30,000 hours of video taught a world model hands, not objects
A rare negative scaling result lands in the middle of the argument about whether video data alone can solve robot manipulation.
Jiahua Dong posted that his team trained a world model (a model that predicts what happens next in a scene) on 30,000 hours of first-person video. It learned exactly where the hands go. What happens to the things those hands touch, much less so.
It learned exactly where the hands go. What happens to the things they touch, much less so.
The part that makes this more than an anecdote is what happened when they turned the dials. Going from 300 hours of video to 30,000 did not narrow the gap. Neither did a model four times larger, which Dong says barely moved it.
Why a negative result lands harder
Most scaling claims in robotics go one direction. More data, more parameters, better numbers, post the chart. This one says the curve flattened on the exact capability that matters for manipulation, which is predicting the consequence of contact rather than the shape of a reach.
That is why it got picked up fast as ammunition. Tonghe Zhang used it to go after the video-scaling camp directly, arguing that people who say scaling video data will solve robotics usually do not understand what a POMDP is, meaning a setting where some of the true state of the world is simply not visible in what the camera sees. His line was that perception is deception, and he credited the post with proving it quantitatively.
Chris Paxton said it matches what he has long argued about video models. His framing was about missing information, and he pushed it further than the model. Even humans, with brains and bodies shaped by millions of years of evolution for manipulation, cannot one shot a new object without touch.
Shown versus claimed
What is in the sources is one team's result and two reactions to it. Dong's thread describes the gap and the two scaling knobs that failed to close it. There is no breakdown here of what the model was, what the evaluation measured, or which dataset the 30,000 hours came from, so nobody reading these posts alone can say how far the finding generalizes to other video-trained world models.
Still, the shape of the claim is the useful bit. Hand trajectories are heavily determined by what you can see, so a video model can learn them. Object outcomes depend on mass, friction, how firmly something is gripped, whether it slipped, and none of that is cleanly readable off pixels. Scaling the thing that is visible does not buy you the thing that is not.
It also sits next to the touch argument that keeps resurfacing in the field, including Paxton's point that humans need contact to handle an unfamiliar object. If you believed video data alone was the path, this is the first number worth arguing with.
Sorry man, perception is deception.
Fundamentally there is a lot of missing information still.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.