On Agentic Manipulation Policies

Abstract

AMP1 is an agentic manipulation policy in which a frontier model controls a robot by writing Python code through a harness for sensing, perception, and motion. The model decides what to observe, can correct a scene model, and checks action outcomes before acting again. The project reports results across RoboLab-120, LIBERO-PRO, robosuite, BEHAVIOR-1K, and RoboDojo, including 86% on RoboLab-120, 95.2% on LIBERO-PRO, 100% on robosuite, 96.0% on BEHAVIOR-1K, and 84.1% on RoboDojo.

Resources

Citation

@article{2026amp,
  title   = {On Agentic Manipulation Policies},
  author  = {Chen, Chen and Bagherieh, Omid and Appiah Balaji, Naveen Nikamanth and
             Gonzalez, Jesse T. and Heravi, Negin and Islam, Fahad and Li, Xueting and
             Regli, Jean-Baptiste and Ren, Kejia and {{S}}vec, Petr and Yoon, David J. and
             M{"u}ller, Matthias and Koltun, Vladlen},
  journal = {arXiv preprint},
  year    = {2026}
}