Compliant Object Manipulation for High Precision Prying Task Using Diffusion Policy with Force Modality

[project page] [data] [Paper]

Jeon Ho Kang, Sagar Joshi, Ruopeng Huang, and Satyandra K. Gupta

University of Southern California

Baseline code for diffusion policy was derived from Diffusion Policy

All Real tags are for real robot implementation

However, data_util.py is shared for real and test

Dependencies

Create Conda Environment (Recommended) and run:

$ pip install requirements.txt

Real Robot

For all demonstrations, we used KUKA IIWA 14 Robot

Real Robot Data for Prying Task (Zarr File)

Data collected on Kuka IIWA 14 robot containing robot state, image, force and action will be published here

To collect your own data:

After obtaininig joint state from handguiding or any other methods,

Run

$ python robot_data_collection_joint_state.py

Training Your Own Policy

After loading your own zarr file or ours in real_robot_network.py

$ python train_real.py

You can select or create your own Config file for training configuration

Inference

$ python inference_real.py

Acknowledgement

• Diffusion policy was adapted from Diffusion Policy

📚 Citation

If you find our work useful or interesting, please consider citing:

@article{kang2025robotic,
  author    = {Jeon Ho Kang and Sagar Joshi and Ruopeng Huang and Satyandra K. Gupta},
  title     = {Robotic Compliant Object Prying Using Diffusion Policy Guided by Vision and Force Observations},
  journal   = {IEEE Robotics and Automation Letters},
  year      = {2025},
  note      = {Accepted for publication. Available on arXiv: arXiv:2503.03998},
}