Everloom/AAWR-DROID

Active Perception AAWR Dataset in GRASP Lab Mock Kitchen

Paper | Project page | Code

Here is our full dataset (13.5GB) in DROID format: https://huggingface.co/datasets/Everloom/AAWR-DROID/

It includes 4 scene as below, each are cleaned with a white list (filter the idle frames)

Sample Usage

To get started with AAWR, follow these steps for a toy simulated task, as described in the code repository:

Setup

1. Install Python dependencies: You can refer to the environment.yaml file, or just manually install the dependencies. There's not too many, mainly torch, gymnasium and their related packages.

2. Install the simulated xarm environment code, and generate demos:

   [email protected]:edwhu/gym-xarm.git
   pip install -e .
   python generate_demos.py # generate demo data.
   
   mkdir gymnasium_xarm_lift_bc
   mv buffer.pkl gymnasium_xarm_lift_bc
   

Train AAWR

Train AAWR on a simple simulated lifting task:

python src/train_il.py agent=asym_awr modality=all task=gymnasium_xarm_lift dataset_dir=data/gymnasium_xarm_lift_bc seed=41 use_wandb=false exp_name=aawr_xarm_lift horizon=1 offline_steps=20000 train_steps=40000 lr=1e-4 grad_clip_norm=10 awr_filter=indicator expectile=0.9 A_scaling=3 save_video=true switch_awr_filter=true

Citation

If you find our project helpful, please cite us:

@inproceedings{
    hu2025realworld,
    title={Real-World Reinforcement Learning of Active Perception Behaviors},
    author={Edward S. Hu and Jie Wang and Xingfang Yuan and Fiona Luo and Muyao Li and Gaspard Lambrechts and Oleh Rybkin and Dinesh Jayaraman},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://openreview.net/forum?id=RkdTtznSAL}
}

Additional Resources

Here is a website on searching behavior under 3rd camera view: https://sites.google.com/view/rwrl-ap/home