Human Preference Modeling Using Visual Motion Prediction Improves Robot Skill Learning from Egocentric Human Video

Researchers developed a new method that allows robots to learn skills from watching human videos 10x faster than previous approaches. The system models human preferences by predicting visual motion between frames, creating a reward function that robots optimize using a modified Soft Actor Critic algorithm. Initialized with just 10 on-robot demonstrations, the approach outperforms prior work across multiple real-world tasks, closing the embodiment gap between human videos and robot execution.