On Agentic Manipulation Policies

Researchers demonstrate that generalist AI models can control robot manipulators more reliably than specialist policies like vision-language-action models. They introduce AMP1, a reference harness that runs across various robot configurations and achieves a high zero-shot success rate on benchmark tasks without demonstrations or task-specific training. Furthermore, adding a persistent memory store allows the agent to learn from experience and significantly lowers failure rates on bimanual tasks. Open article →