Web-video scaling finally proves itself on a real robot
The strongest evidence yet that web-video pre-training transfers to robotics comes from an uncomfortable place: 200+ hours of real-robot testing, not a proxy benchmark. Bigger Direct Video-Action models performed better on a complex industrial manipulation task, while more pre-training compute helped most when robot demonstrations were scarce. Held-out web-video prediction also tracked downstream robot performance, giving teams a practical scaling signal.