Facet-0 predicts and values the contact consequences of its actions. Vision, language, kinematics and causal wrench history condition a joint action–wrench proposal: Cartesian actions are executed; future wrench remains a prediction.

This repository provides the three-stage learning pipeline and policy interfaces:

Use Python 3.11 and the pinned OpenPI checkout below. JAX and PyTorch may also run in separate environments; see the setup and training guide.

git clone https://github.com/PINE-Lab-NTU/FACET.git
git clone https://github.com/Physical-Intelligence/openpi.git
git -C openpi checkout 650c5b0283a49c42784fb5055a0507da2c6d347d
cd FACET
pip install --no-deps -e ../openpi -e ../openpi/packages/openpi-client
pip install -e '.[policy,test]' -e './facet-online-adaptation[test]'

Run the complete CPU example in a new output directory:

export PYTHONPATH="$PWD/../openpi/src:$PWD/../openpi/packages/openpi-client/src"
JAX_PLATFORMS=cpu python examples/small_pipeline.py --output-dir /tmp/facet-small

The example trains small networks on synthetic data and verifies checkpoints, critic learning, token reconstruction, BC → RL → frozen evaluation, and coarse/refinement scheduling. Full-scale training uses external data, weights, judge models and robot controllers.

Training & serving · Model interfaces · Paper mapping · Changes

@article{deng2026facet0,
  title   = {Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation},
  author  = {Deng, Haoyuan and Liu, Haichao and Guo, Wenkai and Ling, Yuan and
             Yang, Zaijia and Xue, Yuanjiang and Sun, Haosheng and Wang, Liangzi and Wang, Ziwei},
  journal = {arXiv preprint arXiv:2609.01596},
  year    = {2026}
}

Apache-2.0. Upstream code, model weights and datasets retain their respective terms.
Research and collaboration: ziwei.wang@ntu.edu.sg.