Generalist AI outperforms specialist policies in robot manipulation
Original title: On Agentic Manipulation Policies
A new agentic manipulation policy reaches 85.6% zero-shot success on RoboLab-120 benchmarks and succeeds in 104 of 108 real-world trials.
Recent robot manipulation work has relied on specialist policies trained on demonstrations. The paper shows generalist AI models can control manipulators more reliably by issuing direct commands that adapt as tasks unfold. The AMP1 harness achieves state-of-the-art results on RoboLab-120 and CaP-Gym without demonstrations or task-specific training, with failure rates reduced by 20x and 15x versus vision-language-action and world-action models. An augmented version with persistent memory store further improves bimanual tasks by 4.6x.
- 85.6% zero-shot success on RoboLab-120
- 104 of 108 real-world trials successful
- No demonstrations or task-specific training required
From the original article
1 min read at arxiv.orgRecent 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).
By Chen Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.