felixw/itps-act

ITPS Maze2D — Action Chunking Transformer (ACT)

Pre-trained Action Chunking Transformer checkpoint used in Inference-Time Policy Steering through Human Interactions (paper, project page, code).

The model was trained on the D4RL Maze2D dataset and is intended to be loaded with the LeRobot policy classes.

Usage

Clone the inference repo, then load this checkpoint directly from the Hub:

git clone https://github.com/yanweiw/itps.git && cd itps
pip install -e .
python interact_maze2d.py -p act --hf

Or load it programmatically:

from itps.common.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained("felixw/itps-act")
policy.eval()

Citation

@article{wang2024itps,
  title={Inference-Time Policy Steering through Human Interactions},
  author={Wang, Yanwei and others},
  journal={arXiv preprint arXiv:2411.16627},
  year={2024}
}

License

MIT — see LICENSE.