metadata

license: mit
task_categories:
  - reinforcement-learning

DMControl 1M

This repository contains the DMControl datasets as used in Learning to Modulate Pre-trained Models in RL:

Trajectories for every task are stored as separate .npz files.
Every .npz file contains 1M transitions.

Download the dataset using the huggingface-cli:

huggingface-cli download ml-jku/dm_control --local-dir=./dm_control --repo-type dataset

For dataloading we refer to our Github repository: https://github.com/ml-jku/L2M

Citation:

@article{schmied2024learning,
  title={Learning to Modulate pre-trained Models in RL},
  author={Schmied, Thomas and Hofmarcher, Markus and Paischer, Fabian and Pascanu, Razvan and Hochreiter, Sepp},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}