diffusion_policy

Diffusion Policy

Install

See here for installation.

Dataset preparation

Collect demonstration data by teleoperation.

Generate a zarr format dataset for learning from teleoperation data:

$ python ../utils/convert_npz_to_zarr.py \
--in_dir ../teleop/teleop_data/<demo_name> --out_dir ./data/<demo_name>.zarr \
--nproc `nproc` --skip 3

Note: If you are using pyenv and encounter the error No module named '_bz2', apply the following solution.
https://stackoverflow.com/a/71457141

Model training

Train a model:

$ python ./bin/TrainDiffusionPolicy.py \
task.dataset.zarr_path=./data/<demo_name>.zarr task.name=<demo_name>

To disable logging by WandB, add the option --enable_wandb=False. You can override the yaml configuration file by adding the following command line arguments, for example: task.shape_meta.obs.joint.shape=\[8\] task.shape_meta.action.shape=\[8\]. The checkpoint files are saved in the log directory.

Note: If you encounter the following error,

ImportError: cannot import name 'cached_download' from 'huggingface_hub'

downgrade huggingface_hub by the following command.

$ pip install huggingface_hub==0.21.4

Policy rollout

Run a trained policy:

$ python ./bin/rollout/RolloutDiffusionPolicyMujocoUR5eCable.py \
--checkpoint ./log/<demo_name>/checkpoints/200.ckpt \
--skip 3 --world_idx 0

Technical Details

For more information on the technical details, please see the following paper:

@INPROCEEDINGS{DiffusionPolicy_RSS23,,
  author = {Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
  title = {Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
  booktitle = {Proceedings of Robotics: Science and Systems},
  year = {2023},
  month = {July},
  doi = {10.15607/RSS.2023.XIX.026}
}