Removing human video drops robot OOD success from 60% to 7%: generalization comes from human video

gui_penedo · x · 2026-11-19

A robotics learning study shows a striking ablation: removing only the human video from the dataset drops average OOD success from 60% to 7%. The robot data contains none of the tasks, while the human video covers the needed motion primitives. The authors conclude generalization comes from human video, made transferable by their 3D motion representation—a strong empirical point for scaling robot data with human video.

Original post →

More from Embodied

Embodied channel →