Robots learned manipulation from ordinary human videos, with 82% success on 12 real tasks

A preprint from Southeast University, Astribot and Sun Yat-sen University proposes Action-Intent Transfer (AIT), a way to use ordinary human videos to train robot manipulation models, which normally need costly teleoperated robot data. It extracts an explicit 'action intent' from videos, aligns human and robot versions of it, then turns it into robot actions.

It reports a 91.3% average success rate on the LIBERO simulation benchmark and, on 12 real tasks with an Astribot S1 robot arm, 82.08% against 70.42% for UniVLA, a baseline from earlier work. The model runs at 8 Hz on a single GPU.

Results are the authors' own, from a limited robot dataset plus videos, and the paper is a preprint.