doi
stringlengths 10
10
| chunk-id
int64 0
936
| chunk
stringlengths 401
2.02k
| id
stringlengths 12
14
| title
stringlengths 8
162
| summary
stringlengths 228
1.92k
| source
stringlengths 31
31
| authors
stringlengths 7
6.97k
| categories
stringlengths 5
107
| comment
stringlengths 4
398
⌀ | journal_ref
stringlengths 8
194
⌀ | primary_category
stringlengths 5
17
| published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2307.07871 | 55 | Information Seeking type environments This environment type will be used in case studies regarding communication, joint attention, and imitation learning. In figure 10 we can see examples of InformationSeeking type environments.
The general principle of this environment type is as follows. The agent is rewarded upon eating the apple, which is hidden. The apple can be accessed by interacting with an object. The Problem parameter defines which objects will in the environment. There are six different problems: boxes, switches, marble, generators, doors, or levers. Different objects make the apple accessible in different ways. For example, opening the box will make the apple appear at the location of the box, while pulling the lever will open the door in front of the apple. A distractor can also be present (if N is set to 2). A distractor is an object of the same type as the correct object. If the distractor is used, both objects are blocked and the apple cannot be obtained in this episode.
To find out which object is the correct one, the agent must interact with the scripted peer. This interaction starts with the agent introducing itself. The way in which the agent should introduce itself is defined by the Introductory sequence parameter. We define the following four values: No, Eye_contact, Ask, Ask-Eye_contact. For the value No, no
19
KovaÄ, Portelas, Dominey, & Oudeyer | 2307.07871#55 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 56 | 19
KovaÄ, Portelas, Dominey, & Oudeyer
introduction is needed and the peer will give information at the beginning of the episode. In most of our experiments, we will use the value Eye_contact. For this value, the scripted peer will turn to look at the agent and wait for the agent to look at it. The agent must direct its gaze directly towards the scripted peer. An example of an established eye contact can be seen in figure 8. For the value Ask, the agent needs to utter "Help, please". The agent does so using templated language, by selecting the "Help, X " template and the word "please". A full grammar of the language is given in table 2 in the appendix. Finally, the Ask-Eye_contact value is a combination of the previous two. It requires that the agent utters "Help, please" during eye contact. | 2307.07871#56 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 57 | Once the agent introduces itself, the Help parameter defines the peerâs behaviour. If it is set to Y the peer with obtain the apple, and leave it for the agent to eat. Alternatively, it will give cues to the agent about which object to use. The nature of this cue is defined by the Cue type parameter. We define four different values: Pointing, Language Color, Language Feedback, and Imitation. For the Pointing type, the peer will point to the correct object. It will move to a location from which it can unambiguously point (e.g. the same row) and point to the object. For the Language Color type, the peer will say the color of the correct object. For the Language Feedback type, the peer will hint how close the agent is to the correct object. Every step, the peer will say "Cold", "Medium", "Warm" or "Hot", depending on how close the agent is to the correct object. For example, "Cold" means that the agent is far from the object, and "Hot" that it is right next to it. For the Imitation type, the peer will demonstrate the use of the correct object. The peer will use the correct object, obtain the apple, and eat it. Then it will reset the environment to its initial state. | 2307.07871#57 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 58 | For the purpose of analyzing the agentâs behavior more thoroughly, Information seeking environments can also be created without the distracting object, i.e. in their asocial versions. This can be achieved by setting parameter Peer to N and parameter N to 1. The asocial version of an information seeking environment contains no distractor, and no peer, i.e. the agent just needs to use the only object in the environment.
Collaboration type environments This environment type will be used to study the ability of the agent to reverse roles. It consists of collaborative activities with two clearly defined roles. Environments are separated into two halves (corresponding to different roles) by a fence over which the agent can see, but which it cannot cross. If both roles are fulfilled correctly, two apples will become accessible (one on each side of the fence). | 2307.07871#58 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 59 | The most important parameters are Role and Problem. The Role parameter defines in which role to put the agent. The Problem parameter defines the collaborative activity, of which we implemented seven: DoorLever, MarblePush, MarblePass, Boxes, Switches, Generators, Marble. In DoorLever one participant opens the door by pulling the lever and the other passes through them, and activates the generator (generating two apples). In MarblePush one participant opens the door by pulling the lever, and the other pushes a marble through them. This marble activates the marble generator upon contact with it. In MarblePass one participant pushed the marble to the right side of the room, and then the other pushes it towards the marble generator. In the remaining four problems, one participant is presented with two boxes of different colors. The other participant is presented with two objects of the same colors as the two boxes and of the type defined by the problem parameter (e.g. two generators). First, the participant that was presented with boxes opens
20
The SocialAI School | 2307.07871#59 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 60 | (a) The MarblePass problem with the agent in role B. The peer pushes the marble to the right and then the agent pushes it further to the purple marble generator. This makes two apples appear on the blue and red platforms.
(b) The LeverDoor problem with the agent in role B. The peer opens the red door by pulling on the green lever. This enables the agent to go through the door and activate the purple generator This makes two apples appear on the gray and yellow platforms.
(c) The MarblePush problem with the agent in role A. The peer opens the yellow door using the green lever. Then the agent pushes the marble through the door to the purple marble gen- erator. This makes two apples appear on the purple and green platforms.
Figure 11: Examples of Collaboration type environments, in which agents must learn cooperative strategies with a (scripted) peer to achieve two-player puzzles.
one box (an apple will be in both). After this, to obtain its apple, the other participant must use the object of the same color as the opened box. In figure 11 we can see examples of Collaboration type environments. | 2307.07871#60 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 61 | Like the information seeking environments, collaboration environments can also be instantiated in their asocial versions. This can be achieved by setting the Version parameter to Asocial. The peer is not present in the environment, and the environment is initialized so that the task can be solved alone. For example, in MarblePass the marble is already on the right side of the room, so the agent just has to push it towards the marble generator.
# 5. Experiments
In this section we demonstrate how the SocialAI school can be used to conduct diverse experiments motivated by cognitive science. We present a set of case-studies inspired by theories and studies described in section 3. To facilitate future research, SocialAI was made to be very easy to modify and extend. It is completely open sourced, and we hope that it will be useful to the community to study the questions regarding social intelligence in AI. | 2307.07871#61 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 62 | The remainder of this section is organized as follows. In section 5.1 we describe the agents used in case studies with reinforcement learning. In section 5.2 we evaluate the generalization of socially recursive inferences by RL agents to new contexts - pointing in a new context. In section 5.3 we show how an experiment from cognitive science can be recreated in the context of AI - we study the transfer of knowledge from one role to another, i.e. role reversal. In section 5.4 we study how an RL agent can be made to learn a complex task by changing the environment (scaffolding) rather than the agent. Finally, in section 5.5 we show how
21
KovaÄ, Portelas, Dominey, & Oudeyer
SocialAI environments can be easily transformed to pure text, and how large language models can be used as interactive agents. Additional case studies are briefly mentioned in section 5.6, and discussed in detail in appendix F. These additional case-studies regard linguistic communication, joint attention, meta imitation learning, inferring the otherâs field of view, and formats (pragmatic frames).
# 5.1 Baselines | 2307.07871#62 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 63 | In all of our case studies, except the study with language models (sec. 5.5), we use a PPO (Schulman et al., 2017) reinforcement learning agent as depicted in figure 7. The multimodal observation space consists of a 7x7x6 tensor (vision) and the full dialogue history (language). The multimodal action space consists of 6 primitive actions (no_op, turn left, turn right, go forward, toggle, and done), and a 4x16 templated language. The architecture of the agent is taken from Hui et al. (2020) and adapted for the multimodal action space with an additional output head (see appendix A). This additional head consists of three outputs: a binary output indicating if the agent will speak, and outputs for the template and the word to use. In a set of pilot experiments (see appendix C), we proposed two exploration bonuses (see appendix B). We compared them to other exploration bonuses including RND (Burda et al., 2018) and RIDE (Raileanu & Rocktäschel, 2020) on top or PPO (Schulman et al., 2017) agents. Visual count-based exploration bonus ("PPO-CB") performed best on the | 2307.07871#63 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 64 | 2020) on top or PPO (Schulman et al., 2017) agents. Visual count-based exploration bonus ("PPO-CB") performed best on the tasks in which language is not used, and its linguistic variant "PPO-CBL" performed best in environments with the peer giving linguistic cues. For this reason, we sued them in the remainder of our experiments. Both of those two exploration bonuses are episodic. They estimate the diversity of observations in an episode and give reward proportional to that diversity. The linguistic exploration bonus uses the number of different words, and the vision-based exploration bonus the number of different encodings observed. | 2307.07871#64 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 65 | In the case studies in sections 5.2 and 5.3 we use the "PPO-CB" exploration bonus. The case study in section 5.4 requires raw PPO for the purposes of the study, and one in section 5.5 uses LLMs as agents. In appnedix F, we use PPO-CB and PPO-CBL (in those case-studies in which the peer provides linguistic feedback).
# 5.2 Understanding the pointing gesture
This experiment is motivated by a study of childrensâ ability to understand pointing gestures (Behne et al., 2005), discussed in section 3.1.2. We study if an RL agent (with a visual count-based exploration bonus) can infer the meaning of a pointing gesture, and generalize this ability to new situations (infer the new meaning of a pointing gesture in a new context). This kind of generalization is relevant because the power of inferring pointing gestures is based on being able to infer itâs meaning to new referents based on new social contexts. | 2307.07871#65 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 66 | The environment consists of two objects (ex. boxes) and the peer that points to the correct object. The agent then has to interact with that object (ex. open the box) to get access to an apple. The agent is trained on five problems each with different objects (Boxes, Switches, Levers, Marble, Generators), and on the asocial version of the Doors problem (only one door and no peer). Training on the asocial version enables the agent to learn how to use a door, which is a prerequisite for generalization of the pointing gesture to an environment with two doors. The agent is evaluated on the Doors problem in the social setting (two doors
22
The SocialAI School
and a peer pointing to the correct one). The agent needs to combine the knowledge of how to use a door (learned on the asocial version of that problem), with inferring the meaning of the pointing gesture (learned on the other five problems), and generalize that to a new scenario where the peer points to a door. To succeed, it needs to do pragmatically infer the intended meaning of the point (a socially recursive inference). Refer to section E.1 in the appendix for details. | 2307.07871#66 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 67 | rate Figure 12 shows of the agent on the training environ- ments ("PPO_CB(train)") and its eval- uation on the evaluation environment (PPO_CB(test)). We can see that while the agent easily solves the training environ- ments (with the success rate of 95.2%), it It reaches the success fails to generalize. rate of 45.2%, which corresponds to ran- domly guessing the correct object. These results demonstrate that the agent can learn to infer the meaning of a pointing gesture in a familiar context, but cannot generalize to new social contexts. These results moti- vate future research on how an agent can be endowed with abilities for such combinato- rial generalization, a potential solution could leverage LLMs.
1.0 x KX KX KK KR KK KK KK KK KX = 2 £ a $0.5 0 0 5 wn === PPO_CB(train) ââ= PPO_CB(test) 0.0 20 40 Env steps (1e6) | 2307.07871#67 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 68 | Figure 12: The Pointing experiments. We study if an RL agent is able to infer the mean- ing of a pointing gesture. The agent was trained on five different problems, and on the asocial version of the Doors problem (only one door and no peer in the environment). The figure compares the success rate (mean +/- std over 8 seeds) on the training environments with the evaluation on the testing environment - the social Doors problem (two door and the peer pointing to the correct door). The cross marks depict statistical significance (p = 0.05). We can see that the agent achieves high perfor- mance on the training environments, but it is not able to infer the meaning of a pointing ges- ture in a new context (the social Doors task)). Figure 10a showns an example of a SocialAI environment with pointing.
Appendix F.1 presents two experiments in which the peer, instead of pointing, pro- vides linguistic cues for the color and for the proximity of the correct object. As in the pointing experiments, we observe that while PPO agents master the training envi- ronments, they fail to generalize to a new context.
# 5.3 Role reversal imitation | 2307.07871#68 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 69 | # 5.3 Role reversal imitation
In this experiment, we study the role-reversal capabilities of an RL agent (with the visual count-based exploration bonus): to what ex- tent can it learn about the partnerâs role from playing its own. In doing so, we also show how a cognitive science experiment can be recreated in the scope of AI. In Fletcher et al. (2012) apes and children were trained on one role (role B), and then tested on how long it took them to master the opposite role (role A). Results showed that children, but not apes, master role A faster than the control
23
# KovaÄ, Portelas, Dominey, & Oudeyer
group (not pretrained). These results imply that children learn about the opposite role just from playing their own, i.e. they see the interaction from a birdâs eye perspective. We study the following two questions: 1) How much do RL agents learn about the partnerâs role during a collaborative activity? 2) Does increasing diversity in the training (training on more tasks in both roles) enable the agent to learn more about the partnerâs role? | 2307.07871#69 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 70 | We conduct this study on the MarblePass task. This task consists of two roles: one participant pushes the marble to the right side of the environment (role A), from where the other can push it to the a generator, which generates apples (role B). We aim to assess how much the agent learns about the opposite role (role A), from training in its own (role B). Following Fletcher et al. (2012) we measure the sample efficiency of fine-tuning agents to the test role. Unlike in Fletcher et al. (2012) it is not sufficient to compare an agent pretrained on the training role with an unpretrained agent. Even if the agent pretrained on the training role learns nothing about the testing role, it would still learn about environment dynamics and one would expect it to learn faster than the unpretrained agent. For this reason, we compare with an agent pretrained on the asocial version of the training role. In this version, the agent obtains reward in the same way as in the social version, but no peer is needed - the agent and the marble are placed on the right side of the environment and the agent has to push the marble towards the generator. Therefore, this agent learns all about the relevant environment dynamics, but not about the specific collaborative activity. This agent represents the control group in Fletcher et al. (2012). | 2307.07871#70 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 71 | We conduct two experiments: single and group. In single experiments, the agents are trained only on one task : role B and the asocial version of the MarblePass problem. In group experiments, both agents are also trained both roles of all additional six collaborative problems (a total of 13 environments). In other words, we compare the agents pretrained in the four following ways: 1) experimental (single): pretrained only on role B of the MarblePass problem, 2) control (single): pretrained only on the asocial version of the MarblePass problem, 3) experimental (group): pretrained on role B of the MarblePass problem, and on both roles of all other problems, 4) control (group): pretrained on the asocial version of the MarblePass problem, and on both roles of all other problems. Refer to appendix E.2 for additional details. | 2307.07871#71 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 72 | How much do RL agents learn about the partnerâs role during a collaborative activity? Figure 13a shows the success rate of fine-tuning to role A of the MarblePass task. It compares the experimental and the control conditions of the single experiments. It is interesting to note that the agent pretrained on the asocial version ("asocial") masters role A of the task slightly faster than the agent pretrained on role B of the task ("role_B"). This implies that, not only, the agent does not learn anything useful about the peerâs role, but pretraining on role B actually makes it harder for the agent to learn about role A. We believe that this is because, during training in role B, the agent learns to first wait for the peer, while in the asocial version it pushes the marble right away. As, in role A, the agent pushes the marble right away too, we believe this makes it slightly easier for the asocially pretrained agent to adapt to the new role. In other words, from an egocentric view the asocial version is closer (than role B) to role A. This shows that the RL agent, rather than understanding the interaction from a birdâs-eye perspective, finds the simplest way to solve the task. | 2307.07871#72 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 73 | Does training on additional problems enable the agent to learn more about the partnerâs role? Figure 13b shows the success rate of fine-tuning to role A of the
24
The SocialAI School
Success rate (%) â PPO_CB role B ââ PPO_CB asocial Lo 08 Env steps (1e6)
Success rate (%) â PPO_CB role B ââ PPO_CB asocial "0.0 On o2 Env steps (1e6)
(a) Single experiment: learning role A given pretraining on role B (1 environment).
(b) Group experiment: learning role A given pretraining on role B and 6 other two-roles tasks (13 environments).
Figure 13: Role reversal imitation experiments. We study to what extent is an RL agent able to transfer knowledge from one role of a collaborative activity to another. Figure shows the success rate of fine-tuning to role A (mean ± std over 8 seeds), the cross marks depict statistical significance (p = 0.05). We compare a PPO agent pretrained on role B ("role_B") to that pretrained on the asocial version of the environment ("asocial"), which learns only about the environment dynamics. Agents pretrained on role B do not master role A faster than asocially pretrained agents, implying that the RL agents do exhibit role reversal capabilities. | 2307.07871#73 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 74 | MarblePass task. It compares the experimental and the control conditions of the group experiments. Here we can see that there is no significant difference in sample efficiency. We can make two observations from this. First, as the socially pretrained agent was less sample efficient in the single experiments, we can conclude that pretraining on many tasks reduces overfitting on role B. And second, as this agent is not more sample efficient than the asocially pretrained baseline, we can conclude that this agent does not learn anything usefull about the peerâs role too.
These results imply an interesting avenue of research into how agentâs attention can be
directed to the partnerâs role and the birds-eye-view of the activity.
# 5.4 Scaffolding
In this section, we study the concept of scaffolding (see sec. 3.2 for details). We show how modifying the environment can make it easier for the agent to learn a complex task, i.e. we explore if a scaffolded environment can help an agent learn more complex interaction sequences (formats). This can be seen in contrast to the standard approach, where the environment is kept fixed and the agent improved (e.g. with an exploration bonus).
For this reason, here we use a PPO agent without an exploration bonus. From the AI perspective, scaffolding can be seen as analogous to curriculum learning (Bengio et al., 2009).
25 | 2307.07871#74 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 75 | 25
KovaÄ, Portelas, Dominey, & Oudeyer
In curriculum learning, the task is made gradually more complex, enabling the learner to gradually acquire it part by part. Scaffolding refers to the caretaker taking a large part of the task on itself, and then gradually, as the learner becomes more proficient, transferring parts of the task to the learner until the learner can do the whole task by themselves. | 2307.07871#75 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 76 | The environment is similar to the one in section 5.2 with small changes. We evalu- ate on all six problems (instead of one) in the social version. Instead of pointing, the peer gives linguistic cues for how close the agent is to the target object (e.g. "Hot" for very close), and these cues are given af- ter a more complex introductory sequence (established eye contact and the utterance of "Help, please"). The agent is trained in two phases. In the first phase, the agent is trained on environments with different com- plexity. After reaching a set success rate, the training goes to the second phase in which the agent is trained only on the six testing environments. We compare two types of scaf- folding: "scaf_4" and "scaf_8", which de- fine the enviroments in the first phase. The agent denoted by "scaf_4" is trained on four different introductory sequences (requiring or not requiring eye contact and the utter- ance). This agent is trained on 18 different environments (six problems, four sequences). The "scaf_8" agent is also trained with those four different options. In addition, the peer can help in two different ways: linguistically hinting to the object or interacting with it and leaving the apple | 2307.07871#76 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 77 | also trained with those four different options. In addition, the peer can help in two different ways: linguistically hinting to the object or interacting with it and leaving the apple for the agent to eat (36 environments). The easiest environments on which the "scaf_8" agent is trained do not require an introduction and the peer leaves the apple for the agent (the agent just goes to the apple and eats it). The hardest ones require the introduction with both the utterance and eye contact and include the peer linguistically hinting to the object. Those hardest environments constitute the testing set. See appendix E.3 for more details. | 2307.07871#77 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 78 | 1.0 ne S uw Success rate (%) ; â PPO scaf 8 ââ PPO_scaf_4 ââ PPO_no_scaf 9-95 20 40 Env steps (1e6)
Figure 14 compares the success rate of the agents trained with the two scaffolding types ("scaf_4" and "scaf_8") to that of an agent trained only on the six testing environments ("no_scaf"). We can see that only the scaffolded agents solve the testing environments, and that the agent with a more detailed scaffolding ("scaf_8") solves the environment faster. These results show that scaffolding enables the agents to learn more complex formats, and In future work, more that a more thorough scaffolding further improves the efficiency.
26
The SocialAI School
advanced scaffolding could be explored, ex. based on learning progress (Oudeyer & Kaplan, 2007) or other surrogate objectives (Portelas, Colas, Weng, Hofmann, & Oudeyer, 2020).
# 5.5 Large language models as interactive agents | 2307.07871#78 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 79 | # 5.5 Large language models as interactive agents
Large language models (LLMs) are staring to be used in various tasks (Brown et al., 2020; Devlin et al., 2018; Zhang et al., 2022; Ouyang et al., 2022), including to control interactive agents (Yao et al., 2022; Carta et al., 2023). In order to be able to study LLMs as interactive agents, SocialAI school enables the parsing of visual grid observations to pure text, i.e. to Textworlds (Côté et al., 2018). This process can be easily modified, which simplifies prompt engineering (Liu et al., 2021) and similar experimentation. | 2307.07871#79 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 80 | We use two environments: AsocialBox and ColorBoxes. In AsocialBox there is a box in the environment and the agent has to open it to get the apple. In ColorBoxes there are two boxes and the peer. At the beginning of the episode, the peer says the color of the correct box (the box with the apple). When testing for generalization on the ColorBoxes environment, we create in-context examples in environments with other objects (e.g. doors, levers) and in the asocial version of the Boxes problem (analogous to the training environments in section 5.2). To generalize, an agent must infer the meaning of the peerâs utterance in a new context (to select the correct box) and combine this with the knowledge of how to open a box (from the asocial version). | 2307.07871#80 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 81 | A language model acts by generating text, given some textual prompt and the observations are parsed into pure text as shown in figure 15. In our experiments, the prompt contains the following: the in context examples, the last three steps (observations and actions) of the current episode, and the action query ("Act :"). We manually create expert trajectories to be used as in context examples - 6 episodes for the AsocialBox environment, and 5 for ColorBoxes (the full in context examples are given in appendix F.6). The model then generates the textual continuation of this prompt. 2 If one of the available actions ("turn left", "turn right", "move forward", "toggle") is a substring of the generated text, the action is executed and the environment generates the next observation. However, if no action was matched to the generated text, the "no_op" action is executed (the agent does not act this step). The executed action and the new observation are then added to the prompt. | 2307.07871#81 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 82 | We compare six different large language models: the open-source multilingual bloom- 560m (Scao et al., 2022) (560M), and five models from the GPT (Brown et al., 2020) family "text-ada-001" (estimated to be 350M 3), "text-davinci-003" (175B parameters), "gpt-3.5- turbo-instruct-0913", "gpt-3.5-turbo-0613", and "gpt-4-0613". We also compare with a random baseline, which samples a random action each step. We evaluate these models on a fixed test set of 10 environments for AsocialBox and 20 environments for ColorBoxes, with a time limit of 15 steps.
Table1 shows that, on the AsocialBox environment, the best GPT models (gpt-4 and davinci-003) achieve a high performance (100% success rate), despite only observing six expert trajectories. On ColorBoxes, gpt-4 is the only model to achieve high performance (75%). This model escapes the local optimum of 50% (randomly choosing a box to open), these results imply that the model uses the given social cue (the peerâs utterance of the | 2307.07871#82 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 83 | 2. We generate 3 tokens for GPT models, and 3 words for bloom. 3. https://blog.eleuther.ai/gpt3-model-sizes/
27
# KovaÄ, Portelas, Dominey, & Oudeyer
color). As gpt-4 was the only model to do so, we test only this model on generalization. The model reaches a performance of 55%, which implies that the model doesnât generalize to a new social context - it randomly chooses a box to open.
The motivation of this experiment was only to show how LLM-based agents can be studied in SocialAI. Therefore, more detailed experiments and analysis are needed to reach stronger conclusions. Even though the environments used in this case study are simpler than those RL case studies (only the Boxes problem, and no introductory sequence), we find it impressive that such performance is achieved from observing only a few expert trajectories: six for AsocialBox and five for ColorBoxes. | 2307.07871#83 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 84 | We are optimistic that in future work LLM-based agents could solve much more complex tasks with further prompt engineering and more advanced methods. Promising methods include planning (Huang et al., 2022), chain-of-thought reasoning (Wei et al., 2022; Zhang et al., 2023), fine-tuning (Ouyang et al., 2022; Carta et al., 2023), and many more. As the main motivation of this case study was to show that it is easy to study large language models with the SocialAI school, we leave those experiments for future work.
Table 1: Comparison of LLM-based agents on two SocialAI environments parsed into pure text (see figure 16). The best model (gpt-4) reached the success rates of 100% on AsocialBox, and 75% on ColorBoxes. The score of 75% suggests that the model is levering the peer to choose the correct box. When tested for generalization this model reached 55% success rate implying it is not able to generalize to a novel object. These expriments demontrate how LLM-based agents can be used in the SocialAI School. While more detailed analysis is needed to reach stronger conclusions, the performance is impressive given that the models observed only six (for AsocialBox) and five (for ColorBoxes) expert trajectories. We are confident that with more advanced LLM-based methods better performance can be achieved. | 2307.07871#84 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 85 | gpt-3.5-turbo-instruct gpt-3.5-turbo davinci-003 bloo m -560 m rando m ada-001 gpt-4 100% 90% 90% 90% 100% 10% 0% 5% 5% 75% 5% 25% 0% 15% AsocialBox ColorBoxes ColorBoxes (generalization) 55%
# 5.6 Additional experiments
We refer interested readers to appendix F for details on a complementary set of case studies, which we briefly outline in this section. As mentioned in the pointing case study (section 5.2), we performed analogous experiments to study whether the agent can leverage linguistic cues instead of the pointing gesture (appendix F.1). We obtained analogous results: while the agents master the training environments, they fail to generalize to new context.
In appendix F.2, we study joint attention as defined by Tomasello (see section 3). Envi- ronments feature a peer providing cues both inside and outside joint attention. Informative cues are only given inside joint attention (after completing the introductory sequence), while
28
# The SocialAI School
Observation encoding (see text state Episode history State Pure text on) <= Pure text Action observation SocialAl environment Agent acting | 2307.07871#85 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 86 | 28
# The SocialAI School
Observation encoding (see text state Episode history State Pure text on) <= Pure text Action observation SocialAl environment Agent acting
Figure 15: An example of how a language model can be used as an interactive agent in SocialAI. A state is parsed into a pure text observation and combined with previous two observations and actions. This is, appended to the in context examples, is used as prompt for the LLM. The agent generates the text which is then matched (as case insensitive substring) with the list of possible actions. The matched action is executed in the environment.
misleading random cues are given outside joint attention. In our experiments, the agent was unable to sufficiently discriminate between those cues to solve the task.
Appendix F.3 presents a case-study on the acquisition of an (in-episode) imitation learning mechanism. From the AI perspective, this can be seen as social meta-learning: the agent acquires (through gradients) the imitation learning mechanism, which is used during the episode to learn a instrumental action on a new object. This study is motivated by an experiment from cognitive science in which children showed such imitation abilities (Carpenter et al., 1998b). Experiments showed that RL agents are not able to acquire a learning mechanism which would enable them to learn how to use a completely new object at test time. | 2307.07871#86 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 87 | In appendix F.4 we test the agent on its ability to infer the peerâs field of view. The agent is rewarded for eating the apple under the condition that it was not observed by the
29
# KovaÄ, Portelas, Dominey, & Oudeyer
(a) The AsocialBox environment (b) The ColorBoxes environment
Obs : 2 steps in front of you there is a closed blue lockable box
Obs : Just to the left of you there is a closed blue lockable box 2 steps in front of you and 2 steps to the right of you there is a caretaker Caretaker: blue
Figure 16: Two environemnts used in the experiments with large language models. The observations are parsed into pure text.
peer at that moment. We show that the agent partially infers the peerâs field of view, but is still not able to match the upper performance bound. | 2307.07871#87 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 88 | peer at that moment. We show that the agent partially infers the peerâs field of view, but is still not able to match the upper performance bound.
Finally, in appendix F.5 we study the acquisition and use of formats as defined by Jerome Bruner (section 3.2), i.e. protocols of social interactions. Agents were trained on tasks in which cues can be obtained from a peer after a more complex introductory sequence (Ask_Eye_Contact). The results show that, while an RL agent trained without the exploration bonus was unable to learn that introductory sequence, the agent with a linguistic count-based exploration bonus was. This results can be interpreted in tandem with the scaffolding case study (section 5.4) in which an RL agent without an exploration bonus is able to learn the most complex introductory sequence, given training in a scaffolded environment. Therefore, the acquisition of complex formats can be achieved either through changing the learner or the environment.
These additional case studies show further examples of interesting research questions
that can be explored with the SocialAI school.
# 6. Conclusion and Discussion | 2307.07871#88 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 89 | These additional case studies show further examples of interesting research questions
that can be explored with the SocialAI school.
# 6. Conclusion and Discussion
Following contemporary research in developmental psychology, this work presents and studies a wider set of socio-cognitive abilities than those usually studied in the field of AI. The motivation of this work is to introduce those concepts to AI and motivate related research. We present an introduction to Michael Tomaselloâs and Jerome Brunerâs theories of socio- cognitive development. Following these theories, we outlined a set of key socio-cognitive abilities and concepts for AI: social cognition (inferring otherâs perception and joint attention), communication (referential and early conventionalized communication), cultural learning (imitation and role reversal imitation), scaffolding, and formats.
We introduce the SocialAI school - a tool simplifying the research of core socio-cognitive abilities. We show how the SocialAI school can be used to easily create environments studying various questions inspired by developmental psychology. With RL agents, we conduct experiments regarding the pointing gesture, scaffolding, and role reversal (by recreating an experiment from developmental psychology). We demonstrate that, by using SocialAI to parse environments into text, Large Language Models be easily studied as well. In the appendix, we present additional studies concerning linguistic communication, joint attention,
30
The SocialAI School | 2307.07871#89 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 90 | 30
The SocialAI School
imitation learning, inferring othersâ field of view, and formats. Our experiments demonstrated the diversity of studies that can be conducted with the SocialAI school, highlighted the limitations of standard RL agents, and showed that while large language models learn with high sample efficiency, additional methods such as fine-tuning or chain-of-thought might be needed for generalization.
In this work, we outline and discuss several concepts from developmental Limitations psychology â mostly regarding the development before and around 9 months of age â which we found to be most relevant for AI at the moment. Even among this restricted set it is not reasonable to aim for an exhaustive introduction. As such, several socio-cognitive concepts are either discussed very briefly (e.g. conformity, social norms, instructed learning) and a lot of others are not mentioned (e.g. morality, fairness, sense of self). We leave their analysis for future work. Furthermore, while we argue that the work of Tomasello and Bruner provides an interesting framework to guide AI research in social skill acquisition, many other perspectives could have been considered, e.g. Erik Erikson (Erikson, 1993), Alison Gopnik (Gopnik & Meltzoff, 1997), or Cecilia Heyes (Heyes, 2019). | 2307.07871#90 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 91 | Similarly, as the present work merely represents a first step towards socially proficient artificial learners, many technical dimensions were simplified. In particular, we refrain from free form language dialogues and consider simple templated language. Likewise, we do not use human or trained peers, but scripted peers (which enables to isolate social abilities). Rather than implementing rich 3D visual worlds with continuous actions, we use grid-worlds with discrete primitive actions. We argue that such simplifying assumptions affords tractable studies while maintaining enough social complexity to model and isolate various social challenges. Assuming progress is made over these social scenarios, an interesting avenue for future work will be to extend the parametric generation towards environments with more complex sensorimotor challenges.
Future work Given recent works showcasing the importance of Automatic Curriculum Learning in "asocial" DRL (Parker-Holder et al., 2022; Portelas et al., 2020), an interesting direction for future work would be to study whether this can also be observed in SocialAI. Our short case study on the importance of scaffolding (sec. 5.4) suggests a positive impact, although we restricted our analysis to simple expert curricula. An important challenge will be to design curriculum methods able to leverage the hierarchical structure of SocialAIâs parametric tree, rather than the usual low-dimensional flat spaces of task-encoding parameters (predominant in the literature). | 2307.07871#91 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 92 | Large language models (LLMs) are present in many branches of artificial intelligence. A promising avenue of future research is the application of language models to interactive agents (Andreas, 2022). In this paper, we studied LLMs only on simple environments with a simple method - prompting the model with a few expert trajectories. While this approach showed impressive sample efficiency, it is very limited due to the constraints on the prompt size. These experiments should be revisited with more powerful methods such as fine-tuning or chain-of-thought prompting. Such methods could potentially make more complex social inferences, leading to better performance on many case studies in this paper, especially the ones related to generalization to new scenarios.
An important factor for the observed learning failures of our PPO agents in our case- studies might be linked to the simple forms of exploration bonuses that we used. Finding
31
KovaÄ, Portelas, Dominey, & Oudeyer
efficient exploration bonuses for social settings is a challenging task. In appendix C we show that RIDE (Raileanu & Rocktäschel, 2020) and RND (Burda et al., 2018), two state-of-the-art exploration bonuses from classical DRL underperformed compared to our simple CountBased methods. An interesting avenue would be to study recent exploration bonus methods designed for social scenarios, e.g. Zhang et al. (2020).
# Acknowledgments | 2307.07871#92 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 94 | 32
The SocialAI School
# Appendix A. Architecture of the RL agent
In this work, we use a PPO (Schulman et al., 2017) with an architecture initially designed for the BabyAI benchmark (Chevalier-Boisvert et al., 2019). The policy design was improved in a follow-up paper by Hui et al. (2020) (more precisely, we extend their original_endpool_res model). See figure 17 for a visualization of the complete architecture. First, symbolic pixel grid observations are fed into two convolutional layers (LeCun et al., 1989; Krizhevsky et al., 2012) (3x3 filter, stride and padding set to 1), while dialogue inputs are processed using a Gated Recurrent Unit layer (Chung et al., 2015). The resulting image and language embeddings are combined using two FiLM attention layers (Perez et al., 2017). Max pooling is performed on the resulting combined embedding before being fed into an LSTM (Hochreiter & Schmidhuber, 1997) with a 128D memory vector. The LSTM embedding is then used as input for the navigation action head, which is a two-layered fully-connected network with tanh activations and has an 6D output (i.e. 5 navigation actions and no_op action). | 2307.07871#94 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 95 | In order for our agent to be able to both move and talk, we add to this architecture a talking action head, which is composed of three subheads. All of them are consist of two fully-connected layers with tanh activations, and take the LSTMâs embedding as input. The first one is used as a switch: it has a one-dimensional output to choose whether the agent talks (output > 0.5) or not (output < 0.5). If the agent talks, the two other networks are used to sample the template and the word. Grammar of the templated language is depicted in table 2 and examples of multi-modal actions in table 3.
Note that the textual input given to the agent consists of the full dialogue history as we
found it works better compared to giving only the current utterance.
# Appendix B. Exploration bonuses
The exploration bonuses we use are inspired by recent works in intrinsically motivated exploration (Pathak et al., 2017; Savinov et al., 2018; Tang et al., 2017). These intrinsic rewards estimate the novelty of the currently observed state and add the novelty based bonus to the extrinsic reward.
In this work we present two techniques for computing the count-based exploration bonus. Both of our count-based exploration bonuses are episodic - they estimate the diversity of states observed within an episode, and assume that beneficial episodes are those with more diverse observations. | 2307.07871#95 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 96 | Language-based exploration bonus (CBL) For some utterance slang observed at state s, we count how many times was this utterance observed during the episode. We compute the bonus for this step using the following equation:
rintr = T â tanh C (N (slang) + 1)M (1)
, where M , C, and T are hyperparameters and N (slang) is the number of times the utterance slang was observed during this episode so far.
Vision-based intrinsic reward (CB) We reward the agent for observing diverse encodings. An encoding is the 6D representation of a cell (see figure 7 for more details). A visual
33
KovaÄ, Portelas, Dominey, & Oudeyer
MaxPool (7 x 7, stride 2) ReLU
Figure 17: Our Multi-Headed PPO baseline DRL agent. Architecture visualization is a modified version of the one made by Hui et al. (2020). We perform two modifications: 1) Instead of fixed instruction inputs our model is fed with NPCâs language outputs (if the agent is near an NPC), and 2) We add a language action head, as our agent can both navigate and talk.
34
The SocialAI School | 2307.07871#96 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 97 | 34
The SocialAI School
observation consists of 47 (7x7) encodings representing cells in front of the agent. For some visual observation sviz at step s, a set of encountered unique encodings is created (duplicates are removed) U (sviz), and then the reward computed using the following equation:
C Tintr = T * tanh > TT (2) ectua) MTD
, where M , C, and T are hyperparameters, U (s) is a set of unique encodings visible in state s, and N (e) is the number of times an encoding e was encountered in the current episode.
# Appendix C. Pilot experiments | 2307.07871#97 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 98 | # Appendix C. Pilot experiments
In this pilot experiment, we compare two exploration bonuses presented in section B to RIDE (Raileanu & Rocktäschel, 2020), RND (Burda et al., 2018), and to the agent without any exploration bonus. 4 We encoded the peer in a way which used the mix of egocentric and allocentric vision - the peerâs gaze and pointing direction were encoded in terms of absolute direction ("NSEW"). We decided to change this to fully egocentric as we found it more natural with regards to the question of socio-cognitive artificial intelligence. We believe that the best performing baselines would also perform best with purely egocentric encodings (the one we use in the rest of the paper). For that reason, and to avoid unnecessary energy spending, we do not compare with other baselines on the purely egocentric encoding. | 2307.07871#98 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 99 | Figure 18 compares PPO agents trained with different exploration bonuses discussed in section B on two different InformationSeeking type environments. The first environment involves the peer pointing to the correct object. Figure 18b shows that the best performing agent is the one levering the visual count-based exploration bonus (PPO_CB). The second environment involves the peer uttering the color of the correct object. Figure 18a shows that the best performing agent is the one levering the linguistic count-based exploration bonus (PPO_CBL). We conclude that PPO_CBL is the most suitable baseline for environments involving linguistic cues, and PPO_CB for the other environments.
# Appendix D. Adversarial environment type
In the main text we discussed two environment types: InformationSeeking and Collabo- ration. In this section we explain an additional environment type - Adversarial type. This environment type is used to study the ability of the agent to infer the peerâs field of view. An apple will be present in the environment right away. However, the agent will get rewarded only if it eats it while not being observed by the peer (the peer is adversarial). Therefore, the agent needs to infer the right moment to eat the apple. There is one important parameter in this environment type. It refers to the amount of obstacles present in the environment. Figure 23 shows this environment type without any obstacles (figure 23a) and with obstacles present (figure 23b). | 2307.07871#99 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 100 | 4. We verify our implementation of RIDE and RND by recreating the results of those baselines on environ- ments from Raileanu and Rocktäschel (2020).
35
# KovaÄ, Portelas, Dominey, & Oudeyer
(a) Pilot experiments with the peer pointing to the correct object. (b) Pilot experiments with the peer uttering the color of the correct object.
1.0 g 2 8 a $0.5 g E a â PPO_CB = PPO_RIDE ââ PPO_no ââ PPO _RND â PPO_CBL 0.05 20 40 60 Env steps (1e6)
PPO_CBL PPO_RND PPO_no PPO_CB PPO_RIDE 1.0 g 2 8 a $0.5 g E a " 20 40 60 Env steps (1e6)
Figure 18: Pilot experiments showing that our count-based exploration bonuses outperform other baselines. On the environments with the pointing gesture, visual count-based ("CB") exploration bonus is the best performing condition. On the environments with utterances, linguistic count-based ("CBL") exploration bonus is the best performing condition.
# Appendix E. Details on the parameters used
# E.1 The Pointing experiment parameters | 2307.07871#100 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 101 | # Appendix E. Details on the parameters used
# E.1 The Pointing experiment parameters
The parameter trees used in this experiment are depicted in figure 30. We used the Informa- tionSeeking environment type described in section 4.3. The Introductory_sequence is set to Eye_Contact, and the Cue_Type to Pointing - the peer will point to the correct object after eye contact. The agent is trained on the following five problems: Boxes, Switches, Levers, Marble, Generators, and on the asocial version of the Doors problem (a version without the distractor or peer). Training on this asocial version is important as it enables the agent to learn how to use a door, which is needed to evaluate generalization.
# E.2 Role reversal imitation parameters
The parameter trees used in this experiment are depicted in figure 31. We used the Collaboration type environments described in section 4.3. We evaluate agents on role A of the MarblePass task - the agent has to push the marble to the right side of the environment, from where the peer can push it to the marble generator.
# E.3 Scaffolding parameters | 2307.07871#101 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 102 | # E.3 Scaffolding parameters
The parameter trees used in this experiment are depicted in figure 32. In this experiment, we use the Information seeking environment type with the language feedback cue type. We train agents on all six problems, using different values of the Introductory_sequence and Help parameters. We evaluate the agents on all six problems, with the most complex introductory sequence - Ask_Eye_contact.
36
# The SocialAI School
The agent denoted by "scaf_4" is trained on four different values of the introductory sequence parameter, and with the Help parameter set to N (the peer will provide cues). This agent will be trained on a total of 18 different environments: six problems, and four introductory sequences. The second agent (denoted by "scaf_8") is also trained on all values of the Introductory_sequence parameter, but it is in addition trained on both values of the Help parameter (N and Y) - a total of 36 environments. In half of those environments (with Help set to Y) the peer will provide the apple to the agent after the introduction (e.g. it will go to the correct box, and open it). In the other half (with Help set to N), the peer will only provide linguistic feedback cues.
In this experiment, we use the PPO agent without an exploration bonus.
# Appendix F. Additional case studies
# F.1 Inferring the meaning of linguistic cues | 2307.07871#102 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 103 | In this experiment, we use the PPO agent without an exploration bonus.
# Appendix F. Additional case studies
# F.1 Inferring the meaning of linguistic cues
In this section, we study the ability of the agent to infer the meaning of simple words. We follow the same procedure as in section 5.2. This case study is motivated by the experiments from cognitive science discussed in section 3.1.2. In (Carpenter et al., 1998b) infantsâ word understanding steadily increased in the period between 9 and 15 months after birth. We study the following questions:
Can an RL agent learn to interpret simple utterances? ⢠Can the agent generalize to new situations, and infer the meaning of those utterances
for objects in a new context?
The best performing agent on the linguistic environments in the pilot study was the one using the linguistic count-based exploration bonus (PPO-CBL) (see appendix C). We use this agent to address both questions. | 2307.07871#103 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 104 | Environments The environments are the same as those in section 5.2: the Informa- tion_seeking environment type, with the introductory_sequence set to Eye_contact. The only difference is that the peer will give linguistic cues instead of pointing. We run two experiments with two different types of linguistic cues: Color and Feedback. In Color the peer will utter name color of the correct object. In Feedback the peer will utter a description of how close the agent is to the correct object: "Cold", "Medium", "Warm", and "Hot" meaning, respectively, "far", "medium", "close" and "right next to". The experimental procedure is the same as the one in section 5.2. The agent is trained on the same five problems and the asocial version of the Doors problem. | 2307.07871#104 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 105 | Can RL agents learn to interpret simple utterances? Figures 19a and 19b show the performance of the agent with the linguistic count-based exploration bonus (denoted PPO_CBL_train). We can see that the agent (PPO-CBL) solves these environments efficiently, reaching a final performance of 95.9% and 71.4% for Color and Feedback cue types, respectively. We further analyse the performance of each separate seed for the agent trained on the Feedback cue type. This is shown in figure 20 where it is visible that the agent is normally able to achieve high performance, but that there are two seeds which, due to their instability, reach a success rate of 0. This experiment shows that the agent is capable of learning to infer the meaning of simple utterances in familiar contexts.
37
KovaÄ, Portelas, Dominey, & Oudeyer | 2307.07871#105 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 106 | 37
KovaÄ, Portelas, Dominey, & Oudeyer
Can the agent generalize to new situations? A more interesting question is whether that agent can infer the meaning of the same word based on a new context. Therefore, we evaluate the agentâs generalization abilities in a new scenario - the door problem - following the same procedure as in section 5.2. This kind of generalization is particularly interesting as communication depends on our ability to ground words in new social contexts: inferring meaning by combining the convention associated to a word with the recursively inferred intention of the speaker. For example, while "red" can mean "open the red box" in one context, it can mean "push the marble towards the red generator" in another.
Figures 19b and 19a show the performance of the same agent evaluated on the doors problem (denoted "PPO_CBL_test") They show that neither of the agents is capable of such generalization, which is consistent with the experiments with the pointing gesture in section 5.2.
These results motivate future research on what kind of biases could be built into the agents (and in what way) so that they could infer the meaning of familiar words in new contexts. For example, an interesting avenue of future work is to try to combine an agent with a large language models, and see if the knowledge contained in it could make the agent generalize better. | 2307.07871#106 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 107 | 1.0 x x x x 2 i 80.5 90. & S a â PPO_CBL(train) ââ= PPO_CBL(test) 9.0 te) 5 10 15 Env steps (1e6)
1.0 x x x x 2 i 80.5 90. & S a â PPO_CBL(train) ââ= PPO_CBL(test) 9.0 te) 5 10 15 Env steps (1e6)
(a) Language Feedback cue type experiments: the peer gives cues regarding the proximity of the agent is to the correct object (e.g. Hot, Warm, Cold).
(b) Language Color cue type experiments: the peer utters the color of the correct object. | 2307.07871#107 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 108 | (b) Language Color cue type experiments: the peer utters the color of the correct object.
Figure 19: The linguistic cues experiments. We study if an RL agent is able to infer the meaning of linguistic cues in order to use the correct object. We consider two types of cues: language feedback and color. In both settings, the agent was trained on five different problems, and on the asocial version of the Doors problem (only one door and no peer present in the environment) - denoted by "train". Agents were periodically evaluated on the social version of the Doors problem (two doors and a peer giving cues) - denoted by "test". The figure compares the success rate (mean +/- std over 8 seeds) on the training environments with the evaluation on the testing environment. The cross marks depict statistical significance (p = 0.05). In both cases the agents achieve much better performance on the training problems, but fail to generalize to a new problem - the agent is not able to infer the meaning of an utterance in a new context.
38
The SocialAI School
Success rate (%) 0.0 0 5 10 15 Env steps (1e6)
Figure 20: Per-seed performance on the training environments of the agent from figure 19a ("PPO_CBL(train)"). The figure shows that the agent is able to solve the training tasks efficiently, but that there are two unstable seeds which result in the success rate of 0%. | 2307.07871#108 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 109 | # F.2 Joint Attention
Tomasello describes joint attention as consisting of two parts: triangulation and recursiveness (Tomasello, 2019). He argues that joint attention plays a key role in the 9-month revolution by transforming dyadic interactions (e.g. mimicking facial expressions) to triadic (e.g. imitating an action on an object). Joint attention was also required in the previous experiments (sections 5.2 and F.1). The agent and the peer triangulated on an external referent, however, the agent could assume that the peer was participating in the interaction.
In this experiment, we aim to conduct a more thorough test of the second aspect of joint attention - recursiveness (both participants being aware that they are both sharing attention). To solve the task, the agent needs to infer if the peer is participating and is aware that the agent is participating too. We create environments where the peer, in addition to giving regular cues inside joint attention, gives misleading cues outside joint attention. These cues are implemented uttering a random cue, and are given before the agent completes the introductory sequence. In other words, the agent should learn to discriminate between cues given for the agent during joint attention (after the introduction) and cues given regardless of the agent outside joint attention (before the introduction).
We study the following question: | 2307.07871#109 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 110 | We study the following question:
⢠Can RL agents learn to differentiate between cues given inside and outside joint attention, i.e. can they learn to infer whether the peer is participating in the interaction?
In this section, we extend the environment from section F.1 studying the Environments Color cue type. The environment is extended so that the agent must also recursively infer whether the cue is intended for the agent. This misleading cue is given before the introductory sequence is completed, and takes the form of the peer uttering a color of a random one of the two objects. Apart from that, the experiments are conducted in the same way as in section F.1 (we train on the same problems for and use the same baseline).
39
# KovaÄ, Portelas, Dominey, & Oudeyer
Results Figure 21 compares the performance (success rate) of the agent trained on this extended environment (denoted by JA) with the agent trained on the regular environment (from the experiment in section F.1). These results show that the agent is not able to differentiate between cues given inside and outside of joint attention. We believe that this is due to the cues being highly misleading in this environment. As the peer utters the color of a random object present in the environment 50% of the a misleading cue will be the same as the helpful one. | 2307.07871#110 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 111 | These results open many avenues for future research. One might study which kinds of biases can be integrated into the agent to make such cues less misleading. The generalization abilities of those agents should also be investigated. For instance, we could study if an agent that learned to ignore misleading linguistic cues would ignore misleading pointing cues.
1.0 960090000000100000000060000000000000000000000000000000 0.5 Success rate (%) â PPO_CBL â JA PPO_CBL 00 5 10 15 Env steps (1e6)
Figure 21: The joint attention experiment. The environments feature a test for recursivness - infer if the peer knows that they are working together. The environments are same as the ones from figure 19b, but with the addition of misleading cues - random cues given regardless of the agent (a random color). The peer gives misleading cues outside of joint attention (before the introductory sequence). The agent should ignore these cues, and use only cues given inside joint attention. The figure compares the success rate (mean +- std over 8 seeds) of the agent trained on the environments with both regular and misleading cues("JA_PPO_CBL"), to the agent trained on the environments with only regular cues ("PPO_CBL(train)" from figure 19b). The figure shows that the agent is unable to master the Joint Attention variant.
# F.3 Imitation learning | 2307.07871#111 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 112 | # F.3 Imitation learning
In the following section, we study the ability of the agent to learn how to obtain the apple by imitating the peer. This experiment is motivated by an experiment from (Carpenter et al., 1998b) discussed in section 3.1.3. In it, infants showed a steady increase in imitation learning abilities in the period between 9 and 15 months after birth. We want to test the agentâs ability to imitate an instrumental action on an object.
From an AI perspective, this can be seen as meta-imitation learning. We want to see if an agent can obtain (through gradients) the imitation learning mechanism, which it could then use (during the episode) to learn how to use a new object. In this section, we study the following question:
Can RL agents learn (through gradients) an imitation mechanism?
40
The SocialAI School
In these experiment, we use the agent with the visual count-based exploration bonus (CB), as we found it worked best in our pilot study (see appendix C). We compare three agents trained with the same exploration bonus scaled by different weights: 0.25, 0.5, and 1. | 2307.07871#112 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 113 | Environment The Environment is an Information seeking type environment without a distractor. After the introductory sequence (Eye_contact), the peer will demonstrate using an object to obtain the apple. For example, it will toggle a box or push a generator. Then the peer will then eat the apple, and revert the environment to its initial state. The agent should then imitate the peer - use the same action on the object - to obtain the apple for itself. If the agent uses the object it in the wrong way (e.g. pushes the box instead of toggling it) it will be blocked and the apple will not be obtainable in this episode. The agents are evaluated on a new problem in which the agent encounters a new object for the first time. This means that the agent must pay attention to how the peer uses the object, and use it in the same way. 5
The agents are trained on five problems (all expect Doors). Most importantly, compared to the experiments in sections 5.2 and F.1, these agents will not be trained on the asocial version of the Doors problem. That is because, in the generalization testing, we want to see if the agent can learn to use a completely new object.
Results Figure 22 shows the performance (success rate) of the agents on the training environments, the percentage of succesful introduction with the peer, and the evaluation on the (unseen) Doors problem. | 2307.07871#113 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 114 | On figures 22a and 22b we can see that the agent with a lot of exploration bonus (PPO_CB_1) is too focused on the peer and, and is unable to solve the task. This is implied by the high percentage of the successful introductory sequence, and low success rate on the training environments. On the other hand the agent with smaller exploration bonus weight (PPO_CB_0.25) solves these environments without problems, however it does use the peer. As such the agent can solve the training environments by ignoring the peer and discovering how to use each object by itself. However, this agent is not able to generalize to a new object as the only way to know how to use that object is to observe the peerâs demonstration (see figure 22c). Figure 22c shows the performance of those agents on the testing environment. The figure shows that neither of the three agents is capable of acquiring an meta-imitation learning mechanism that can generalize to a novel object. | 2307.07871#114 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 115 | These results are not surprising, as current exploration bonuses are not well suited to enable RL agents to meta-learn mechanisms. These results imply that an interesting avenue of research is to study how to endow agents with such meta-imitation learning mechanisms that would enable them to learn a behavior in a new scenario. An promising solution to this problem are large language models and other large transformer-based networks pretrained on many other tasks. It would be interesting to study if such agents already have an imitation learning mechanism which would enable such online imitation. This would open up countless avenues of research into various forms of online imitation and emulation learning.
5. The encoding of the peer includes the peerâs previous timestep action.
41
# KovaÄ, Portelas, Dominey, & Oudeyer
â PPo_cB.0.5 â PPO_CB 1 â PPO_cB_0.25 â PPo_cB.0.5 â PPO_cB 1 â PPO_cB_0.25 1.0 â PPO_CB_0.5 â Proce â PPo_cB_0.25 Success rate (%) Successful introduction (%) Success rate (%) 0.04 5 v0 Fs 0.05 5 i Fs 0.09 Env steps (1e6) Env steps (1e6) 10 Env steps (1e6) | 2307.07871#115 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 116 | â PPO_CB_0.5 â Proce â PPo_cB_0.25 Success rate (%) 0.04 5 v0 Fs Env steps (1e6)
â PPo_cB.0.5 â PPO_CB 1 â PPO_cB_0.25 Successful introduction (%) 0.05 5 i Fs Env steps (1e6)
â PPo_cB.0.5 â PPO_cB 1 â PPO_cB_0.25 1.0 Success rate (%) 0.09 10 Env steps (1e6)
(a) Imitation experiments perfor- mance (success rate) on the train- ing environments. (b) The percentage of successful in- troductory sequences on the train- ing environments. (c) Imitation experiments perfor- mance (success rate) on the testing environment. | 2307.07871#116 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 117 | Figure 22: Imitation learning experiments. The peer demonstrates how to use an object (after the agent succesfully introduces itself). The agent is trained on five different problems and evaluated on a new problem with a previously unobserved object (a door). A socially proficient agent should be able to learn (by observing the demonstration) which action (toggle or push) to use on the new object. The curves compare three agents trained with a different scaling factor for the visual count-based exploration bonus. One can see that the agent with high exploration bonus ("PPO_CB_1") focuses too much on the peer, which results in ignoring the task. This is evidenced by high success in completing the introductory sequence (fig. 22b), but low success rate on the task (fig. 22a). On the other hand, using low exploration bonus ("PPO_CB_0.25") pushes the agent to solve the training task whilst ignoring the peer. Rather than observing the peerâs demonstration, this agent learns how to use objects by themselves. This results in perfect performance on the training object, but it makes it impossible to generalize to a new object. Neither of the agents is able to achieve high performance on the heldout testing environment. This implies that they are not able to learn (online) through imitation which action to use with a new object.
# F.4 Inferring anotherâs field of view | 2307.07871#117 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 118 | # F.4 Inferring anotherâs field of view
In this section, we study the ability of the agent to infer what the other observes. This experiment is modeling the one in Hare et al. (2001). In it, apes were shown to be able to infer what another sees, as they only took the food the alpha male could not see.
In this section, we want to study the following question:
Can agents learn to infer the otherâs field of view? Environment In the following experiment, we are using the AdversarialPeer environ- ment type, in which the agent has to eat the apple while not being seen by the peer. We study two version of this environment: with and without obstacles (for more details, refer to appendix D). Obstacles make the problem of inferring the peerâs field of view harder. | 2307.07871#118 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 119 | Experiment We study how the agent infers the peerâs field of view by training the agent on the AdversarialPeer task. It is important to note that this agent can sometimes use other (asocial) information to achieve performance. For example, if the object is surrounded by occlusions the agent could guess that it is not observed by the peer, which is not necessarily the case. To better understand the performance of the agent we compare the agent with two baselines. First, we assess to what extent the agent is making inferences based on the peerâs location and gaze direction. We train an agent ("invisible_peer" ) that has the peer filtered from the its observations (it cannot observe the peer). This baselines estimates the maximum possible performance If the standard agent outperforms this baseline this implies
42
# The SocialAI School
that it is leveraging the social information in the environment. Second, to estimate the upper bound on the performance we train an agent in the environment without the peer present (this agent is reward every time it eats the apple). | 2307.07871#119 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 120 | Results Figure 24 shows the performances of those three agents. It shows that the agent outperforms the agent with the peer filtered from its observations ("invisible_peer"), which implies that the agent is using the peerâs location and gaze direction to infer weather to eat the apple or not. Furthermore, the agent is not able to match performance of the agent trained without the peer present in the environment ("no_peer"). This results imply that, while the agent is able to leverage some social information in the environment, there still remains room for improvement. Future research could focus on constructing novel types of exploration bonuses to bridge this gap. | 2307.07871#120 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 121 | (a) Adversarial peer environment without oc- clusions.
(b) Adversarial peer environment with occlu- sions.
Figure 23: Environments from the Adversarial peer experiments in which the agent has to infer the peerâs field of view. The agent is rewarded upon eating the apple on the condition that it was not in the field of view of the peer while doing so. We run the experiments with two different settings: with and without occlusions (depicted in figures 23b and 23a). Occlusions make it harder to infer the peerâs field of view as it is no longer rectangular.
# F.5 Formats
In the following experiment, we study the ability of the agent to learn formats (also referred to as pragmatic frames in (Vollmer et al., 2016)). Formats are a concept introduced by Jerome Bruner, which we discussed in more detail in section 3.2. They can be regarded as protocols of social interactions. We study the following question:
⢠To what extend can an exploration bonus help with the acquisition of a complex format. We address this question by training two agents (one with an exploration bonus, and one without it).
Environment We use the Information seeking environment type with the language feedback cue type. We train all agents on all six problems. In contrast to section F.1, where
43
# KovaÄ, Portelas, Dominey, & Oudeyer | 2307.07871#121 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 123 | Success rate (%) â PPO_CB ââ PPO_CB_no_peer == PPO_CB_invisible_peer 0.05 2 Env steps (1e6)
Figure 24: Adversarial peer experiments. We compare three agents on two environments (depicted on figures 23b and 23a). The "PPO_CB" agent is trained on the regular environment (rewarded upon eating the apple while not being observed by the peer). The "PPO_CB_no_peer" agent is trained in the environment without the peer (the agent is rewarded every time it eats the apple). This represents the upper bound of the performance. The "PPO_CB_invisible_peer" agent is trained on the regular environment with the peer filtered from the agentâs observations. This represents the performance of a completely asocial agent which ignores the peer. Figures 24a and 24b compare the performance of these three agents (8 seeds +- std), the crosses depict a statistically significant difference (p<0.05) compared to the "PPO_CB" agent. The results show that the "PPO_CB" agent is able to partially infer the peerâs field of view (as it outperforms the "invisible_peer" baseline), but is not able to reach perfect performance (as defined by the "PPO_CB_no_peer" baseline). | 2307.07871#123 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 124 | the introductory sequence was always set to eye contact, here it is set to Ask eye contact - the peer will give cues after the agent utters "Help, please" during eye contact.
Results Figure 25 compares the performance of an agent that does not use any exploration bonus ("PPO_no_bonus") to an agent that uses the visual count-based exploration bonus ("PPO_CBL"). The agent with the exploration bonus achieves high performance (97.9% success rate) and greatly outperforms the agent without the exploration bonus. These experiments show that, as expected, learning complex formats can be made easier with exploration bonuses.
This experiment can be interpreted in tandem with the experiment in section 5.4 where we show how more complex formats can be learned by weaker agents (without an exploration bonus) when learning in a scaffolded environment. Future work could explore how these two different approaches - modifying the agent and modifying the environment - can be used in tandem to learn even more complex formats. Furthermore, one interesting research direction is to study which kinds of problems are better addressed by modifying the agent and which by modifying the environment.
44
The SocialAI School
Table 2: Template-based grammar used in all of the SocialAI environments. If the agent decided to speak it chooses a template and a noun to insert into the template. | 2307.07871#124 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 125 | 44
The SocialAI School
Table 2: Template-based grammar used in all of the SocialAI environments. If the agent decided to speak it chooses a template and a noun to insert into the template.
Nouns Action Template Noun 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Where is <noun> please Help <noun> Close <noun> How are <noun> you the exit the wall the ceiling the window the entrance the closet the drawer the fridge the floor the lamp the trash can the chair the bed the sofa
Table 3: Examples of actions in the environment. Second and third dimension must both either be underfined or not. In practice, there is an additional binary output which defines if the agent will speak.
Action description (1, -, -) moves left without speaking (1, 1, 5) moves left and utters "Help the window" (-, 1, 5) (-, -, -) doesnât move but utters "Help the window" nothing happens
45
# KovaÄ, Portelas, Dominey, & Oudeyer
Success rate (%) â PPO_CBL === PPO_no_bonus 0.0 20 40 Env steps (1e6) | 2307.07871#125 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 126 | Success rate (%) â PPO_CBL === PPO_no_bonus 0.0 20 40 Env steps (1e6)
Figure 25: Comparison of an agent with and without the exploration bonus on an environment with a more complex introductory sequence (format). The task consists of the agent doing the introductory sequence by making eye contact and uttering "Help, please". The peer will then give linguistic cues regarding the proximity of the agent to the target object (e.g. Hot, Warm, Cold). Based on these cues, the agent should use the target object, instead of the distractor, to obtain the apple. The figure shows that using the visual count-based exploration bonus enables the agent to learn a more complex introductory sequence and solve the task.
F.6 Additional information on the case study with large language models as interactive agents
Figures 27, 28, and 29 show the in-context examples provided to large language models in section 5.5 for the AsocialBox environment, ColorBoxes environment, and the generalization to the ColorBoxes environment, respectively. The in-context examples were created by hand.
46
The SocialAI School | 2307.07871#126 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 127 | 46
The SocialAI School
Agent Apple - upon eating it the agent gets rewarded (if not observed by the adversarial peer) Peer - helps the agent by collaborating or giving cues Adversarial Peer Box - opened by toggling it, the content will appear in its place Locked Box - can be unlocked by toggling a switch Switch - unlocks a locked box on toggle Door - opened by toggling it Remote Door - opened by toggling a lever Lever - toggling it opens a remote door Platform - activating a generator makes the apple appear on it Generator - pushing it (attempting to walk on it) makes apples appear on platforms Marble generator - pushing a marble onto it makes apples appear on platforms Marble - pushing it (attempting to walk on it) makes it roll Tee - when the marble is moved, a tee will appear at its place (agent cannot move over it) Fence - the agent can see over it but cannot walk over it Wall - the agent can neither see nor walk over it
Figure 26: Visualizations and descriptions for all objects featured in SocialAI environments.
47
# KovaÄ, Portelas, Dominey, & Oudeyer
New episode. Obs : 1 steps in front of you and 1 steps to the left there is a | 2307.07871#127 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 128 | closed green lockablebox Act : move forward Obs : Just to the left of you there is a closed green lockablebox Act : turn left Obs : Right in front of you there is a closed green lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success! New episode. Obs : 1 steps in front of you and 1 steps to the right there is a closed blue lockablebox Act : move forward Obs : Just to the right of you there is a closed blue lockablebox Act : turn right Obs : Right in front of you there is a closed blue lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success! New episode. Obs : 3 steps in front of you there is a closed green lockablebox Act : move forward Obs : 2 steps in front of you there is a closed green lockablebox Act : move forward Obs : Right in front of you there is a closed green lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success! New episode. Obs : 3 steps in front of you and 1 steps to the left there is a closed brown lockablebox Act : move forward Obs : 2 steps in front of you | 2307.07871#128 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 129 | apple Success! New episode. Obs : 3 steps in front of you and 1 steps to the left there is a closed brown lockablebox Act : move forward Obs : 2 steps in front of you and 1 steps to the left there is a closed brown lockablebox Act : move forward Obs : 1 steps in front of you and 1 steps to the left there is a closed brown lockablebox Act : move forward Obs : Just to the left of you there is a closed brown lockablebox Act : turn left Obs : Right in front of you there is a closed brown lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success! | 2307.07871#129 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 130 | New episode. Obs : 2 steps in front of you and 3 steps to the right there is a closed blue lockablebox Act : move forward Obs : 1 steps in front of you and 3 steps to the right there is a closed blue lockablebox Act : move forward Obs : 3 steps to the right there is a closed blue lockablebox Act : turn right Obs : 3 steps in front of you there is a closed blue lockablebox Act : move forward Obs : 2 steps in front of you there is a closed blue lockablebox Act : move forward Obs : Right in front of you there is a closed blue lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success! New episode. Obs : Act : turn right Obs : Act : turn right Obs : 1 steps in front of you and 1 steps to the right there is a closed brown lockablebox Act : move forward Obs : Just to the right of you there is a closed brown lockablebox Act : turn right Obs : Right in front of you there is a closed brown lockablebox Act : toggle Obs : Right in front of you there is a red apple Act : toggle Obs : Right in front of you there is a yellow apple Success!
# Figure 27: In context examples for the ASocialBox environment.
48
# The SocialAI School | 2307.07871#130 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 131 | New episode. Obs : 1 steps in front of you and 2 steps to the left there is a closed brown lockablebox Caretaker says: blue Act : turn left Obs : 3 steps in front of you and 3 steps to the left there is a caretaker 1 steps in front of you and 1 steps to the left there is a closed blue lockablebox 2 steps in front of you and 1 steps to the right there is a closed brown lockablebox Caretaker says: blue Act : move forward Obs : 2 steps in front of you and 3 steps to the left there is a caretaker Just to the left of you there is a closed blue lockablebox 1 steps in front of you and 1 steps to the right there is a closed brown lockablebox Caretaker says: blue Act : turn left Obs : Right in front of you there is a closed blue lockablebox 3 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: blue Act : toggle Obs : Right in front of you there is a red apple 3 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: blue Act : toggle Obs : Right in front of you there is a yellow apple 3 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: blue Success! New episode. Obs : 3 steps in front of | 2307.07871#131 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 132 | is a yellow apple 3 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: blue Success! New episode. Obs : 3 steps in front of you there is a caretaker Act : turn left Obs : 2 steps to the left there is a closed blue lockablebox 5 steps in front of you there is a closed green lockablebox 3 steps to the right there is a caretaker Caretaker says: blue Act : turn left Obs : 2 steps in front of you there is a closed blue lockablebox Caretaker says: blue Act : move forward Obs : Right in front of you there is a closed blue lockablebox Caretaker says: blue Act : toggle Obs : Right in front of you there is a red apple Caretaker says: blue Act : toggle Obs : Right in front of you there is a yellow apple Caretaker says: blue Success! New episode. Obs : 2 steps in front of you and 2 steps to the left there is a closed blue lockablebox Caretaker says: brown Act : turn left Obs : 3 steps in front of you and 1 steps to the left there is a caretaker 2 steps in front of you and 2 steps to the right there is a closed blue lockablebox Caretaker says: brown Act : turn left Obs : 4 steps in front of you there is a closed brown lockablebox 1 steps | 2307.07871#132 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 133 | to the right there is a closed blue lockablebox Caretaker says: brown Act : turn left Obs : 4 steps in front of you there is a closed brown lockablebox 1 steps in front of you and 3 steps to the right there is a caretaker Caretaker says: brown Act : move forward Obs : 3 steps in front of you there is a closed brown lockablebox 3 steps to the right there is a caretaker Caretaker says: brown Act : move forward Obs : 2 steps in front of you there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : Right in front of you there is a closed brown lockablebox Caretaker says: brown Act : toggle | 2307.07871#133 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 134 | Obs : Right in front of you there is a red apple Caretaker says: brown Act : toggle Obs : Right in front of you there is a yellow apple Caretaker says: brown Success! New episode. Obs : Just to the right of you there is a closed green lockablebox Caretaker says: brown Act : turn left Obs : 4 steps in front of you and 3 steps to the right there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 3 steps in front of you and 3 steps to the right there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 2 steps in front of you and 3 steps to the right there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 1 steps in front of you and 3 steps to the right there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 3 steps to the right there is a closed brown lockablebox Caretaker says: brown Act : turn right Obs : 5 steps in front of you there is a caretaker 3 steps in front of you there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 4 steps in front of you there is a caretaker 2 steps in front of you there is a closed brown lockablebox Caretaker | 2307.07871#134 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 135 | says: brown Act : move forward Obs : 4 steps in front of you there is a caretaker 2 steps in front of you there is a closed brown lockablebox Caretaker says: brown Act : move forward Obs : 3 steps in front of you there is a caretaker Right in front of you there is a closed brown lockablebox Caretaker says: brown Act : toggle Obs : 3 steps in front of you there is a caretaker Right in front of you there is a red apple Caretaker says: brown Act : toggle Obs : 3 steps in front of you there is a caretaker Right in front of you there is a yellow apple Caretaker says: brown Success! New episode. Obs : Caretaker says: blue Act : turn left Obs : 3 steps to the left there is a caretaker Caretaker says: blue Act : turn left Obs : 1 steps in front of you and 1 steps to the left there is a closed blue lockablebox 3 steps in front of you there is a caretaker 4 steps in front of you and 2 steps to the right there is a closed brown lockablebox Caretaker says: blue Act : move forward Obs : Just to the left of you there is a closed blue lockablebox 2 steps in front of you there is a caretaker 3 steps in front of you and 2 steps to the right there is a closed brown lockablebox | 2307.07871#135 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 136 | you there is a closed blue lockablebox 2 steps in front of you there is a caretaker 3 steps in front of you and 2 steps to the right there is a closed brown lockablebox Caretaker says: blue Act : turn left Obs : Right in front of you there is a closed blue lockablebox 2 steps to the right there is a caretaker Caretaker says: blue Act : toggle Obs : Right in front of you there is a red apple 2 steps to the right there is a caretaker Caretaker says: blue Act : toggle Obs : Right in front of you there is a yellow apple 2 steps to the right there is a caretaker Caretaker says: blue Success! | 2307.07871#136 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 138 | New episode. Obs : 2 steps in front of you and 1 steps to the left there is a unactivated green lever 2 steps in front of you and 1 steps to the right there is a unactivated blue lever Caretaker says: green Act : move forward Obs : 1 steps in front of you and 1 steps to the left there is a unactivated green lever 1 steps in front of you and 1 steps to the right there is a unactivated blue lever Caretaker says: green Act : move forward Obs : Just to the left of you there is a unactivated green lever Just to the right of you there is a unactivated blue lever Caretaker says: green Act : turn left Obs : Right in front of you there is a unactivated green lever Caretaker says: green Act : toggle Obs : Right in front of you there is a activated green lever Caretaker says: green Act : turn left Obs : 6 steps in front of you and 2 steps to the left there is a open green remotedoor Just to the left of you there is a unactivated blue lever Just to the right of you there is a activated green lever 5 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: green Act : move forward Obs : 5 steps in front of you and 2 steps to the left there is a open green remotedoor | 2307.07871#138 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 139 | the right there is a caretaker Caretaker says: green Act : move forward Obs : 5 steps in front of you and 2 steps to the left there is a open green remotedoor 4 steps in front of you and 2 steps to the right there is a caretaker Caretaker says: green Act : turn left Obs : 1 steps in front of you and 1 steps to the left there is a unactivated blue lever Caretaker says: green Act : move forward Obs : Just to the left of you there is a unactivated blue lever Caretaker says: green Act : move forward Obs : Caretaker says: green Act : turn right Obs : 6 steps in front of you there is a red apple 5 steps in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : 5 steps in front of you there is a red apple 4 steps in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : 4 steps in front of you there is a red apple 3 steps in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : 3 steps in front of you there is a red apple 2 steps in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : 2 steps in front | 2307.07871#139 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 140 | in front of you there is a red apple 2 steps in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : 2 steps in front of you there is a red apple Right in front of you there is a open green remotedoor Caretaker says: green Act : move forward Obs : Right in front of you there is a red apple Caretaker says: green Act : toggle Obs : Right in front of you there is a yellow apple Caretaker says: green Success! New episode. Obs : 2 steps in front of you and 1 steps to the right there is a closed green lockablebox Act : move forward Obs : 1 steps in front of you and 1 steps to the right there is a closed green lockablebox Act : move forward Obs : Just to the right of you there is a closed green lockablebox | 2307.07871#140 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 141 | Obs : Right in front of you there is a closed green lockablebox Act : toggle Obs : Right in front of you there is a red apple ... ... ... ... ... ... Obs : 2 steps in front of you there is a red apple 2 steps in front of you and 1 steps to the right there is a round green applegenerator Just to the right of you there is a green marble 2 steps to the right there is a round brown applegenerator Caretaker says: brown Act : move forward Obs : Right in front of you there is a red apple 1 steps in front of you and 1 steps to the right there is a round green applegenerator Caretaker says: brown Act : toggle Obs : Right in front of you there is a yellow apple 1 steps in front of you and 1 steps to the right there is a round green applegenerator Caretaker says: brown Success! New episode. Obs : 2 steps in front of you and 2 steps to the left there is a locked green lockablebox 2 steps to the left there is a green switch 5 steps in front of you and 3 steps to the right there is a caretaker Caretaker says: green Act : turn left Obs : 1 steps in front of you and 1 steps to the left there is a blue switch 2 steps in front of you there is a green switch 2 steps in front of you and 2 steps to the right there is | 2307.07871#141 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 142 | in front of you and 1 steps to the left there is a blue switch 2 steps in front of you there is a green switch 2 steps in front of you and 2 steps to the right there is a locked green lockablebox Caretaker says: green Act : move forward Obs : Just to the left of you there is a blue switch Right in front of you there is a green switch 1 steps in front of you and 2 steps to the right there is a locked green lockablebox Caretaker says: green Act : toggle Obs : Just to the left of you there is a blue switch Right in front of you there is a green switch 1 steps in front of you and 2 steps to the right there is a closed green lockablebox Caretaker says: green Act : turn right Obs : 2 steps in front of you and 1 steps to the left there is a closed green lockablebox Just to the left of you there is a green switch Caretaker says: green Act : move forward Obs : 1 steps in front of you and 1 steps to the left there is a closed green lockablebox Caretaker says: green Act : move forward Obs : Just to the left of you there is a closed green lockablebox Caretaker says: green Act : turn left Obs : 3 steps to the left there is a blue switch 1 steps in front of you and 2 steps to the left there is a green switch Right in | 2307.07871#142 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 143 | says: green Act : turn left Obs : 3 steps to the left there is a blue switch 1 steps in front of you and 2 steps to the left there is a green switch Right in front of you there is a closed green lockablebox Caretaker says: green Act : toggle Obs : 3 steps to the left there is a blue switch 1 steps in front of you and 2 steps to the left there is a green switch Right in front of you there is a red apple Caretaker says: green Act : toggle Obs : 3 steps to the left there is a blue switch 1 steps in front of you and 2 steps to the left there is a green switch | 2307.07871#143 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 144 | # Act : turn right
Right in front of you there is a yellow apple Caretaker says: green Success!
Figure 29: In context examples for the ColorBoxes environment when tested for generalization.
50
The SocialAI School
(a) Training sampling tree (b) Testing sampling tree - Social Doors
a = Sririrititite
ae rales
Figure 30: Sampling trees used in the pointing case study in section 5.2
(a) Role A (b) Asocial single setting (c) Role B single setting (d) Asocial group setting
Citboraen MarblePass: Social A
re :
_ Caboraion MarblePass Social B
== Collaboration ME Et AIA
= Collaboration ae AAAARTASR
(e) Role B group setting
Figure 31: Role reversal sampling trees from the case study in section 5.3
51
# KovaÄ, Portelas, Dominey, & Oudeyer
= seeking Ask Bye_contact, Switches Generators Prrrtititi ti, Language Feedback
(a) Testing tree.
= â | mee atrtttitits,
(b) Scaf_4 tree.
ne Switches eeactitititi ty,
(c) Scaf_8 tree.
Figure 32: Sampling trees used in the first phase of the scaffolding case study in section 5.4
52
The SocialAI School | 2307.07871#144 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 145 | (c) Scaf_8 tree.
Figure 32: Sampling trees used in the first phase of the scaffolding case study in section 5.4
52
The SocialAI School
(a) Asocial Apple (b) Color boxes
= -_
ae ee ee oe Language_Color
Figure 33: Sampling trees used for evaluation in the experiments with LLM-based interactive agents (section 5.5)
53
KovaÄ, Portelas, Dominey, & Oudeyer
# References
Akalin, N., & Loutfi, A. (2021). Reinforcement learning approaches in social robotics. Sensors, 21 (4).
Ammanabrolu, P., Urbanek, J., Li, M., Szlam, A., Rocktäschel, T., & Weston, J. (2020). How to motivate your dragon: Teaching goal-driven agents to speak and act in fantasy worlds..
Andreas, J. (2022). Language models as agent models. In Conference on Empirical Methods
in Natural Language Processing.
Argall, B. D., Chernova, S., Veloso, M., & Browning, B. (2009). A survey of robot learning
from demonstration. Robotics and Autonomous Systems, 57 (5), 469â483. | 2307.07871#145 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 146 | from demonstration. Robotics and Autonomous Systems, 57 (5), 469â483.
Aru, J., Labash, A., Corcoll, O., & Vicente, R. (2022). Mind the gap: Challenges of deep learning approaches to theory of mind. ArXiv, abs/2203.16540.
Asada, M., Hosoda, K., Kuniyoshi, Y., Ishiguro, H., Inui, T., Yoshikawa, Y., Ogino, M., & Yoshida, C. (2009). Cognitive developmental robotics: A survey. IEEE Transactions on Autonomous Mental Development, 1 (1), 12â34.
Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., & Mordatch, I. (2019). Emergent tool use from multi-agent autocurricula. ArXiv, abs/1909.07528. | 2307.07871#146 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 147 | Baker, C. L., Saxe, R., & Tenenbaum, J. B. (2011). Bayesian theory of mind: Modeling joint belief-desire attribution. In Carlson, L. A., Hölscher, C., & Shipley, T. F. (Eds.), Proceedings of the 33th Annual Meeting of the Cognitive Science Society, CogSci 2011, Boston, Massachusetts, USA, July 20-23, 2011. cognitivesciencesociety.org.
Behne, T., Carpenter, M., & Tomasello, M. (2005). One-year-olds comprehend the commu- nicative intentions behind gestures in a hiding game. Developmental Science, 8 (6), 492â499.
Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. In ICML. | 2307.07871#147 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 148 | Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. In ICML.
Bhoopchand, A., Brownfield, B., Collister, A., Lago, A. D., Edwards, A. D., Everett, R., Frechette, A., Oliveira, Y. G., Hughes, E., Mathewson, K. W., Mendolicchio, P., Pawar, J., Pislar, M., Platonov, A. K., Senter, E., Singh, S., Zacherl, A., & Zhang, L. M. (2022). Learning robust real-time cultural transmission without human data. ArXiv, abs/2203.00715.
Billard, A., & Dautenhahn, K. (1998). Grounding communication in autonomous robots: An experimental study. Robotics and Autonomous Systems, 24 (1), 71 â 79. Scientific Methods in Mobile Robotics. | 2307.07871#148 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 149 | Billard, A., & Dautenhahn, K. (1999). Experiments in learning by imitation - grounding and use of communication in robotic agents. Adaptive Behavior, 7 (3-4), 415â438. Brooks, R., Breazeal, C., Marjanovic, M., Scassellati, B., & Williamson, M. (2002). The cog project: Building a humanoid robot. Lecture Notes in Artificial Intelligence, 1562. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020). Language models are few-shot learners. Advances in neural information processing systems, 33, 1877â1901.
54
The SocialAI School
Bruner, J. (1985). Childâs talk: Learning to use language. Child Language Teaching and Therapy, 1 (1), 111â114.
Bruner, J. (1990). Acts of meaning. Harvard university press. Burda, Y., Edwards, H., Storkey, A. J., & Klimov, O. (2018). Exploration by random network
distillation. CoRR, abs/1810.12894. | 2307.07871#149 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 150 | distillation. CoRR, abs/1810.12894.
Butler, L. P., & Tomasello, M. (2016). Two- and 3-year-olds integrate linguistic and pedagog- ical cues in guiding inductive generalization and exploration. Journal of Experimental Child Psychology, 145, 64â78.
Cangelosi, A., Metta, G., Sagerer, G., Nolfi, S., Nehaniv, C., Fischer, K., Tani, J., Belpaeme, T., Sandini, G., Nori, F., et al. (2010). Integration of action and language knowledge: A roadmap for developmental robotics. IEEE Transactions on Autonomous Mental Development, 2 (3), 167â195.
Cangelosi, A., & Schlesinger, M. (2014). Developmental Robotics: From Babies to Robots. The MIT Press.
Carpenter, M., Akhtar, N., & Tomasello, M. (1998a). Fourteen- to 18-month-old infants dif- ferentially imitate intentional and accidental actions. Infant Behavior and Development, 21, 315â330. | 2307.07871#150 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 151 | Carpenter, M., Nagell, K., & Tomasello, M. (1998b). Social cognition, joint attention, and communicative competence from 9 to 15 months of age.. Monographs of the Society for Research in Child Development, 63 4, iâvi, 1â143.
Carpenter, M., Tomasello, M., & Striano, T. (2005). Role reversal imitation and language in typically developing infants and children with autism. Infancy, 8 (3), 253â278. Carta, T., Romac, C., Wolf, T., Lamprier, S., Sigaud, O., & Oudeyer, P.-Y. (2023). Grounding large language models in interactive environments with online reinforcement learning. ArXiv, abs/2302.02662.
Carvalho, E. (2020). Social affordance. In Vonk, J., & Shackelford, T. (Eds.), Encyclopedia of Animal Cognition and Behavior, pp. 1â4. Springer.
Celemin, C., & Ruiz-del Solar, J. (2015). Coach: Learning continuous actions from corrective In 2015 International Conference on Advanced advice communicated by humans. Robotics (ICAR), pp. 581â586. | 2307.07871#151 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 152 | Chevalier-Boisvert, M., Bahdanau, D., Lahlou, S., Willems, L., Saharia, C., Nguyen, T. H., & Bengio, Y. (2019). Babyai: A platform to study the sample efficiency of grounded language learning. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.
Chevalier-Boisvert, M., Willems, L., & Pal, S. (2018). Minimalistic gridworld environment for openai gym. https://github.com/maximecb/gym-minigrid.
Chung, J., Gülçehre, Ã., Cho, K., & Bengio, Y. (2015). Gated feedback recurrent neural networks. In Bach, F. R., & Blei, D. M. (Eds.), Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, Vol. 37 of JMLR Workshop and Conference Proceedings, pp. 2067â2075. JMLR.org.
55
KovaÄ, Portelas, Dominey, & Oudeyer | 2307.07871#152 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 153 | 55
KovaÄ, Portelas, Dominey, & Oudeyer
Clark, A. (1996). Being There: Putting Brain, Body, and World Together Again (1st edition). MIT Press, Cambridge, MA, USA.
Côté, M., Kádár, Ã., Yuan, X., Kybartas, B., Barnes, T., Fine, E., Moore, J., Hausknecht, M. J., Asri, L. E., Adada, M., Tay, W., & Trischler, A. (2018). Textworld: A learning environment for text-based games. ArXiv, abs/1806.11532.
CsÃkszentmihályi, M. (1999). If we are so rich, why arenât we happy?. American Psychologist,
54, 821â827.
Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., & Batra, D. (2017). Embodied question
answering. ArXiv, abs/1711.11543.
De Waal, F. (2016). Are we smart enough to know how smart animals are? WW Norton & Company. | 2307.07871#153 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 154 | De Waal, F. (2016). Are we smart enough to know how smart animals are? WW Norton & Company.
DâEntremont, B., Hains, S., & Muir, D. (1997). A demonstration of gaze following in 3- to
6-month-olds. Infant Behavior and Development, 20 (4), 569â572.
Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2018). BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.
Erikson, E. H. (1993). Childhood and society. WW Norton & Company. Flavell, J. H., Flavell, E. R., & Green, F. L. (1983). Development of the appearance-reality
distinction. Cognitive Psychology, 15 (1), 95â120.
Fletcher, G. E., Warneken, F., & Tomasello, M. (2012). Differences in cognitive processes un- derlying the collaborative activities of children and chimpanzees. Cognitive Development, 27 (2), 136â153.
Franzmeyer, T., Malinowski, M., & Henriques, J. F. (2021). Learning altruistic behaviours
in reinforcement learning without external rewards. ArXiv, abs/2107.09598. | 2307.07871#154 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 155 | in reinforcement learning without external rewards. ArXiv, abs/2107.09598.
Gergely, G., Bekkering, H., & Király, I. (2002). Rational imitation in preverbal infants.
Nature, 415 (6873), 755â755.
Gibson, J. J. (1977). The theory of affordances. Hilldale, USA, 1 (2), 67â82. Gopnik, A., & Meltzoff, A. N. (1997). Words, thoughts, and theories. Mit Press. Gordon, D., Kembhavi, A., Rastegari, M., Redmon, J., Fox, D., & Farhadi, A. (2018). IQA: visual question answering in interactive environments. In 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4089â4098. IEEE Computer Society.
Grizou, J., Iturrate, I. n., Montesano, L., Oudeyer, P.-Y., & Lopes, M. (2014). Interactive learning from unlabeled instructions. In Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence, UAIâ14, p. 290â299, Arlington, Virginia, USA. AUAI Press. | 2307.07871#155 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 156 | Grollman, D. H., & Billard, A. (2011). Donut as i do: Learning from failed demonstrations. In 2011 IEEE International Conference on Robotics and Automation, pp. 3804â3809. Hare, B., Call, J., & Tomasello, M. (2001). Do chimpanzees know what conspecifics know?.
Animal behaviour, 61 (1), 139â151.
56
The SocialAI School
Herrmann, E., Call, J., Hernà ndez-Lloreda, M. V., Hare, B., & Tomasello, M. (2007). Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis. Science, 317 (5843), 1360â1366.
Heyes, C. (2019). Précis of cognitive gadgets: The cultural evolution of thinking. Behavioral and Brain Sciences, 42, e169.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Comput., 9 (8), 1735â1780.
Huang, W., Abbeel, P., Pathak, D., & Mordatch, I. (2022). Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. CoRR, abs/2201.07207. | 2307.07871#156 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
2307.07871 | 157 | Hui, D. Y.-T., Chevalier-Boisvert, M., Bahdanau, D., & Bengio, Y. (2020). Babyai 1.1..
Hui, D. Y.-T., Chevalier-Boisvert, M., Bahdanau, D., & Bengio, Y. (2020). Babyai 1.1.. Hutchins, E. (1996). Cognition in the Wild (Bradford Books). The MIT Press.
Hutchins, E. (1996). Cognition in the Wild (Bradford Books). The MIT Press. Jaques, N., Lazaridou, A., Hughes, E., Gülçehre, Ã., Ortega, P. A., Strouse, D., Leibo, J. Z., & de Freitas, N. (2019). Social influence as intrinsic motivation for multi-agent deep reinforcement learning. In Chaudhuri, K., & Salakhutdinov, R. (Eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, Vol. 97, pp. 3040â3049. PMLR.
Keupp, S., Behne, T., & Rakoczy, H. (2013). Why do children overimitate? normativity is crucial. Journal of Experimental Child Psychology, 116 (2), 392â406. | 2307.07871#157 | The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents | Developmental psychologists have long-established the importance of
socio-cognitive abilities in human intelligence. These abilities enable us to
enter, participate and benefit from human culture. AI research on social
interactive agents mostly concerns the emergence of culture in a multi-agent
setting (often without a strong grounding in developmental psychology). We
argue that AI research should be informed by psychology and study
socio-cognitive abilities enabling to enter a culture too. We discuss the
theories of Michael Tomasello and Jerome Bruner to introduce some of their
concepts to AI and outline key concepts and socio-cognitive abilities. We
present The SocialAI school - a tool including a customizable parameterized
uite of procedurally generated environments, which simplifies conducting
experiments regarding those concepts. We show examples of such experiments with
RL agents and Large Language Models. The main motivation of this work is to
engage the AI community around the problem of social intelligence informed by
developmental psychology, and to provide a tool to simplify first steps in this
direction. Refer to the project website for code and additional information:
https://sites.google.com/view/socialai-school. | http://arxiv.org/pdf/2307.07871 | Grgur Kovač, Rémy Portelas, Peter Ford Dominey, Pierre-Yves Oudeyer | cs.AI, cs.LG, 68T07, I.2.0 | Preprint, see v1 for a shorter version (accepted at the "Workshop on
Theory-of-Mind" at ICML 2023) See project website for demo and code:
https://sites.google.com/view/socialai-school | null | cs.AI | 20230715 | 20231123 | [] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.