Researchers led by Qianzhong Chen, with Mac Schwager and Roberto Martín-Martín among the authors, have built WEVO, a way to improve a humanoid robot's vision-language-action policy without teleoperation, according to a paper on arXiv. The policy learns by reinforcement inside a video world model that predicts what the robot will see after each action.
Simulated play data teaches the world model about contact and changing viewpoints, and 20 real runs per task match it to what the robot actually sees. Across three real robot tasks, training in the world model raised average success from 23.3% to 75.0%, against 48.3% for reinforcement learning on the real robot.



