On Agentic Manipulation Policies
AMP1 completed 104 of 108 real-world robot trials without demonstrations or task-specific training, the authors report.
The harness lets a generalist AI model command robots directly and adapt as tasks unfold, running unchanged across single-arm, bimanual and mobile setups. On RoboLab-120, it achieved 85.6% zero-shot success, compared with 49.2% for a leading vision-language-action policy and 42.9% for a leading world-action policy. Without built-in grasping, pushing or placing skills, the agent developed strategies including searching for hidden objects, using objects as tools and calibrating cameras. Adding memory across episodes cut failures on RoboDojo bimanual tasks by a further factor of 4.6 compared with zero-shot deployment.
ArXiv · AI/CL/LG's note
The harness lets a generalist AI model command robots directly and adapt as tasks unfold, running unchanged across single-arm, bimanual and mobile setups. On RoboLab-120, it achieved 85.6% zero-shot success, compared with 49.2% for a leading vision-language-action policy and 42.9% for a leading world-action policy. Without built-in grasping, pushing or placing skills, the agent developed strategies including searching for hidden objects, using objects as tools and calibrating cameras. Adding memory across episodes cut failures on RoboDojo bimanual tasks by a further factor of 4.6 compared with zero-shot deployment.
ArXiv · AI/CL/LG's note
score 7