---
license: mit
task_categories:
- robotics
language:
- en
size_categories:
- 100M<n<1B
pretty_name: M
---

# ManipGen-UniDexGrasp
ManipGen-UniDexGrasp is used to train pick and place policies in [ManipGen](https://mihdalal.github.io/manipgen/). We process the [UniDexGrasp](https://pku-epic.github.io/UniDexGrasp/) dataset with the following steps:
1. Augment object size: We add smaller versions of the objects to the dataset.
2. Augment rest poses: The rest object poses in the original dataset are obtained by dropping the object from a height. We add some common rest poses which are less likely to be sampled in the original dataset (e.g., a standing bottle).
3. Sample grasp pose: We sample grasp poses for each object using antipodal sampling. Then we teleport the gripper to grasp the object from the sampled poses, add perturbations, and record valid grasp poses. The ratio of valid grasp pose is also used to filter intractable rest poses (e.g., an upside-down bowl).

### Structure
* `datasetv4.1_posedata.npy`: rest poses for each object
* `meshdatav3_scaled/`: mesh and urdf
* `graspdata/`: pre-sampled grasp poses for Franka arm with UMI gripper
* `trainset3419.txt`: the list of 3419 objects used to train pick and place policies in ManipGen in the format of "[object_code] [object_scale]"