Researchers led by Siyuan Zheng, with Henry X. Liu among the authors, have built M4M, which learns where a mobile robot should move so that a trained manipulation policy can succeed, according to a paper on arXiv. From the robot's stereo cameras, it estimates how likely a task is to succeed from places the robot has not yet visited.
The estimate comes from the manipulation policy's own hidden representations and a limited number of real outcomes. It is built into the training of a navigation policy, and a handoff policy decides when to stop moving and start manipulating. On eight tasks with articulated objects in unseen RoboCasa kitchens, M4M averaged 47.0% success, against 26.5% for the strongest baseline and 19.2% for manipulating without moving first.



