Masked visual actions help video-action models generalize across robots and even orangutans

CSProfKGD · x · 2026-07-25

A robotics thread says video-action models are now the state of the art for generalist robotics, following RhodaAI’s Direct Video-Action Model.

It highlights a new approach using masked visual actions to make video-action models more data-efficient and less embodiment-specific. The quoted work says it trained a video world model on just 15 hours of single-arm robot video, generalized zero-shot to unseen embodiments — even orangutans — and could synthesize robot motion from desired object motion.

Related event: Zero-Shot Robot World Model Trained with 15 Hours of Video(3 posts)→

Original post →

More from Embodied

Embodied channel →