Robotics · 2026-11-09

On Agentic Manipulation Policies

Chen Chen, Omid Bagherieh, Naveen Nikamanth Appiah Balaji, Jesse T. Gonzalez, Negin Heravi, Fahad Islam, Xueting Li, Jean-Baptiste Regli, Kejia Ren, Petr \v{S}vec, David J. Yoon, Matthias M\"uller, Vladlen Koltun

arXiv:2611.09892PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors introduce AMP1, a reference harness where a generalist AI model directly commands a robot manipulator in an agentic manipulation policy loop (no demonstrations, no task-specific training). AMP1 reports state-of-the-art zero-shot results on RoboLab-120 and CaP-Gym simulations and strong zero-shot success on real hardware, including 85.6% success on RoboLab-120 (Gemini 4 Argon).

Why it matters

The paper argues that keeping a generalist model in the control loop (direct action, adapting as the task unfolds) yields more general and robust manipulation than specialist VLA/WAM policies. AMP1 is designed to run unchanged across single-arm, bimanual, and mobile manipulation, enabling systematic evaluation and analysis of robustness, failure modes, and the benefit of persistent memory.

Method

  • Define agentic manipulation policies (AMPs) where a generalist AI model commands the manipulator directly, adapting actions as the task unfolds.
  • Provide AMP1 harness with visual observations, code execution, and direct control of manipulator motion, run unchanged across multiple manipulation settings.
  • Evaluate zero-shot performance on RoboLab-120 and CaP-Gym, plus real-world trials, and study effects of recovery and persistent memory.

Limitation

The paper states that AMPs are “not yet practical, as the robot spends much of each episode waiting for the model to reason.”

Abstract (from arXiv)

Recent work on robot manipulation has focused on specialist policies trained on demonstrations. We show that generalist AI models can control manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). We systematically analyze agentic manipulation policies (AMPs), in which a generalist AI model commands the manipulator directly, adapting its actions as the task unfolds. We present AMP1, a reference harness that runs unchanged across single-arm, bimanual, and mobile manipulation. AMP1 achieves state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks, with no demonstrations or task-specific training. On RoboLab-120, AMP1 reaches an 85.6\% success rate zero-shot, compared to 49.2\% and 42.9\% for leading VLA and WAM policies. The agent succeeds zero-shot in 104 of 108 real-world trials spanning twelve tasks and three robot instances, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, the agent can also learn from experience: an augmented harness with a memory representation that is carried across episodes lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment. Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-06 · Indexed by Pub-AI: 2026-11-09 UTC

preprint

On Agentic Manipulation Policies

Chen Chen, Omid Bagherieh, Naveen Nikamanth Appiah Balaji, Jesse T. Gonzalez, Negin Heravi, Fahad Islam, Xueting Li, Jean-Baptiste Regli, Kejia Ren, Petr Švec, David J. Yoon, Matthias Müller, Vladlen Koltun

Preprint posted: 2026-11-06 · Original record ↗

arxiv
2611.09892
Explicit source versions (1)
  • v1 · 2026-11-06
  1. Preprint available

    Source date: 2026-11-06 · First detected: 2026-11-09 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Recent work on robot manipulation has focused on specialist policies trained on demonstrations. We show that generalist AI models can control manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). We systematically analyze agentic manipulation policies (AMPs), in which a generalist AI model commands the manipulator directly, adapting its actions as the task unfolds. We present AMP1, a reference harness that runs unchanged across single-arm, bimanual, and mobile manipulation. AMP1 achieves state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks, with no demonstrations or task-specific training. On RoboLab-120, AMP1 reaches an 85.6\% success rate zero-shot, compared to 49.2\% and 42.9\% for leading VLA and WAM policies. The agent succeeds zero-shot in 104 of 108 real-world trials spanning twelve tasks and three robot instances, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, the agent can also learn from experience: an augmented harness with a memory representation that is carried across episodes lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment. Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.