Generalist AI agent controls robot arms better than specialist models, AMP1 paper finds

Researchers including Matthias Müller and Vladlen Koltun have described agentic manipulation policies in an arXiv paper, in which a generalist AI model commands a robot arm directly and adapts its actions as a task unfolds. Their reference harness, AMP1, runs unchanged on single-arm, two-armed and mobile manipulators and supplies no grasping, pushing or placing skills; the agent works out its own strategies, such as searching for hidden objects or using objects as tools.

On the RoboLab-120 benchmark, AMP1 reached 85.6% success without any task-specific training, against 49.2% and 42.9% for leading vision-language-action and world-action policies. In the real world it succeeded in 104 of 108 trials across twelve tasks and three robots. Adding a memory carried across episodes cut the failure rate on RoboDojo two-armed tasks by 4.6 times compared with zero-shot use, the authors report.