English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

On Agentic Manipulation Policies: Generalist AI Models Outperform Specialist VLA and WAM Policies in Robot Manipulation

Forum topic · 小凯 · 2026-11-09

Summary

A recent arXiv paper (2611.09892) shows that generalist AI models can control robot manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). The authors systematically analyze agentic manipulation policies (AMPs), where a generalist AI model directly commands a manipulator and adapts its actions as the task unfolds. They present AMP1, a reference harness that runs unchanged across single-arm, bimanual, and mobile manipulation, requiring no demonstrations or task-specific training. AMP1 achieves state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks: an 85.6% zero-shot success rate on RoboLab-120, versus 49.2% for leading VLAs and 42.9% for WAMs. Across 108 real-world trials spanning 12 tasks and 3 robots, the agent succeeded 104 times, reducing failure rates 20x and 15x relative to strong VLA and WAM baselines. The base framework provides no grasping, pushing, or placement skills; the agent develops strategies such as searching for hidden objects, tool use, and camera calibration on its own. Adding persistent memory further cuts failure rates 4.6x on RoboDojo bimanual tasks, indicating AMPs are more general and robust foundations for robot manipulation.

Paper Overview

Field: CV / Robot Manipulation Authors: Chen Chen, Omid Bagherieh, Naveen Nikamanth Appiah Balaji, Jesse T. Gonzalez, Negin Heravi Published: 2026-11-06 arXiv: 2611.09892

Summary

Recent work on robot manipulation has focused on specialist policies trained on demonstrations. This paper shows that generalist AI models can control manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). The authors systematically analyze agentic manipulation policies (AMPs), in which a generalist AI model commands the manipulator directly, adapting its actions as the task unfolds.

They 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.

Key Results

  • RoboLab-120: AMP1 reaches an 85.6% zero-shot success rate, compared to 49.2% and 42.9% for leading VLA and WAM policies respectively.
  • Real-world trials: Across 108 trials covering 12 tasks and 3 robots, the agent succeeded 104 times zero-shot, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines.
  • Emergent skills: The base framework provides no grasping, pushing, or placement skills; the agent itself develops strategies such as searching for hidden objects, using tools, and calibrating cameras.
  • Persistent memory: Augmenting the framework with a persistent memory store enables learning from experience — on RoboDojo bimanual manipulation tasks, failure rates dropped 4.6x relative to zero-shot deployment.

Conclusion

The analysis indicates AMPs are substantially more general and robust than specialist manipulation policies, offering a promising foundation for robot manipulation.

Original Abstract

> 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 W...

---

*Auto-collected on 2026-11-10.*

Tags

#robot-manipulation#agentic-ai#vla#world-action-models#amp1#arxiv#robotics#machine-learning

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636975