A reference implementation for the Affordance-based Robot Manipulation with Flow Matching.
- Paper page: Affordance-based Robot Manipulation with Flow Matching. https://arxiv.org/abs/2409.01083
- Project page: https://hri-eu.github.io/flow-matching-policy/
- Code: https://github.com/HRI-EU/flow_matching
- Author: Fan Zhang (fan.zhang@honda-ri.de), Michael Gienger
We are in process of integrating flow matching into the Hugging Face 🤗 LeRobot PushT task.
🔬 This repo contains
Training and evaluation examples of using flow matching on PushT and Franka Kitchen benchmarks.
🌷 Getting Started
- Clone this repo and change into it:
git clone git@github.com:HRI-EU/flow-matching-policy.git && cd flow_matching\ - Install the Python dependencies:
python -m venv venv_fm && source venv_fm/bin/activate && pip install --no-cache-dir -r requirements.txt - Enjoy!
🏆 Some Results
Pretrained weights with flow matching: Push-T, Franka Kitchen
| Methods | Push-T1 | Push-T2 | Franka Kitchen |
|---|---|---|---|
| Flow Matching | 0.9035/0.7519 | 0.7363/0.6218 | 0.9960/0.7172 |
sampling range1: [rs.randint(50, 450), rs.randint(50, 450), rs.randint(200, 300), rs.randint(200, 300), rs.randn() * 2 * np.pi - np.pi]
sampling range2: [rs.randint(50, 450), rs.randint(50, 450), rs.randint(100, 400), rs.randint(100, 400), rs.randn() * 2 * np.pi - np.pi]
📝 Acknowledgements
- The model structure implementation is modified from Cheng Chi's diffusion_policy repo. The code is under external/diffusion_policy (MIT license). Some code that we modified is located under external/models.
- We use some functions from Alexander Tong's TorchCFM repo (MIT license). It is installed through pip.
- Please download the PushT demonstration datat from Google Drive (id=1KY1InLurpMvJDRb14L9NlXT_fEsCvVUq&confirm=t) from Cheng Chi's diffusion_policy repo.
- Please download the Franka Kitchen demonstration data from Nur Muhammad Shafiullah's Behavior Transformers repo (MIT license).
This project is licensed under the BSD 3-clause license - see the LICENSE.md file for details