ForceCapture

Official implementation for ForceCapture in the paper ForceMimic: Force-Centric Imitation Learning with Force-Motion Capture System for Contact-Rich Manipulation, accepted by ICRA 2025.

For more information, please visit our project website.

---

Hardware

Download the fixed-tool version assets from Onshape and print them as needed. The force-torque sensor used here is an inner product by Flexiv, but you can easily replace it with a commonly-used alternative, like ATI. The SLAM camera used here is Intel RealSense T265, and the observation camera used here is Intel RealSense L515, while they are discontinued. And the angle encoder used here is by pdcd.

Installation

git clone git@github.com:ForceMimic/forcecapture.git --recurse-submodules
cd forcecapture
# `main` branch for fixed-tool version, `git checkout gripper` to `gripper` branch for gripper version

conda create -n focap python=3.10
conda activate focap

cd r3kit
pip install -e .
cd ..

pip install -r requirements.txt

Pre-collection

python tare_pyft.py --config configs/tare_pyft.txt
python test_tare_pyft.py --config configs/tare_pyft.txt

Collection

python collect_data.py --config configs/collect.txt

NOTE: During collection, use htop to watch memory usage, and use sudo sync and sudo sysctl -w vm.drop_caches=3 to free memory.

Post-collection

python visualize_data.py --config configs/visualize.txt # also be used as annotation
python create_hdf5.py --config configs/hdf5.txt
python visualize_hdf5.py --config configs/visualize_hdf5.txt
python merge_hdf5.py --config configs/merge.txt
python visualize_merge.py --config configs/visualize_merge.txt

Example Data

You can download our processed dataset from Google Drive.

---

Acknowledgement

Our design and implementation are inspired by DexCap and UMI. Kudos to the authors for their amazing contributions.

Citation

If you find our work useful, please consider citing:

@inproceedings{liu2025forcemimic,
  author={Liu, Wenhai and Wang, Junbo and Wang, Yiming and Wang, Weiming and Lu, Cewu},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={ForceMimic: Force-Centric Imitation Learning with Force-Motion Capture System for Contact-Rich Manipulation}, 
  year={2025}
}

License

This repository is released under the MIT license.