Human video emerges as robotics pretraining substrate: effective experience = hours × information per hour
rohanpaul_ai · x · 2026-09-22
Maxinsights argues human egocentric video is now a serious pretraining substrate for robot manipulation, but works best when translated into the robot's own embodiment first. It proposes scaling Physical AI data on two coupled axes: Experience Scale (total hours recorded) and Experience Density (learnable physics per hour — object states, contact, forces, geometry, tool use).
Evidence: Dyna-2, pretrained on 2M+ hours of egocentric video, improves monotonically across four orders of magnitude of human experience and transfers across the embodiment gap. Two recordings of equal duration can differ an order of magnitude in learning value — repetitive tabletop pick-and-place vs. bimanual manipulation with varied contacts.
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12