Human videos can improve robot policies after alignment

Robot interaction data is expensive, but human videos contain far more varied hand-object activity. It remains unclear whether those videos can teach a robot policy anything useful when they do not record robot commands and human movement cannot be directly executed by a robot.

It learns continuous action-like signals from human-video frame changes, then aligns them with robot actions so more human video can support policy pre-training.

The decisive research move is to learn a continuous latent action from each human-video frame transition, then use it for initial policy training before replacing it with executable target-robot action chunks during embodiment alignment and task training.

In the controlled scaling comparison, only the amount of cleaned human video used in Stage 1 changed. Across three AgiBot G1 tasks, average task success rose from 63.3% at 1K hours to 85.0% at 5K hours, with robot data and evaluation conditions fixed.

This supports the authors' view that the latent supervision makes additional human video useful under this protocol.

The method assumes that visual changes learned from human videos can become useful robot supervision after embodiment alignment, but that alignment is not guaranteed.

The real-robot evidence covers seven tasks on one platform with 20 trials per condition, so it does not establish broad deployment robustness.