On Agentic Manipulation Policies
arXiv cs.AI · 7 Nov 2026, 04:59 AEST
Recent research indicates that generalist AI models can outperform specialist manipulation policies in robot control. The study introduces agentic manipulation policies (AMPs) that adapt actions in real-time, achieving significant success rates in various benchmarks.
Key facts
- 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).
- 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.
- Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.
Read aloud
Recent research indicates that generalist A I models can outperform specialist manipulation policies in robot control. The study introduces agentic manipulation policies AMPs that adapt actions in real-time, achieving significant success rates in various benchmarks. We show that generalist A I models can control manipulators more reliably than specialist policies such as vision-language-action models VLAs and world-action models WAMs. AMP1 achieves state-of-the-art results on the RoboLab-one hundred twenty and CaP-Gym benchmarks, with no demonstrations or task-specific training. On RoboLab-one hundred twenty, AMP1 reaches an eighty five point six percent success rate zero-shot, compared to forty nine point two percent and forty two point nine percent for leading VLA and WAM policies. The agent succeeds zero-shot in one hundred four of one hundred eight real-world trials spanning twelve tasks and three robot instances, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.
Sources in this cluster
- On Agentic Manipulation Policies
arXiv cs.AI · 7 Nov 2026, 04:59 AEST