WARP-RM is a self-supervised reward model that learns dense relative-progress signals from demonstrations and uses them to filter and reweight action chunk loss during behavior cloning for long-horizon robot manipulation.
The code used in this project to deploy high-frequency visuomotor policies on the bimanual I2RT YAM robot is hosted on repo uynitsuj/robots_realtime
- Coming Soon! Initial code release - Targeting end of July
- Coming Soon! ABC Simulation Reproducible Results - Targeting end of July
BibTeX coming soon. See the project page in the meantime.
This project is released under the MIT License.