A new controller let Neil Nie's ablated model fold a whole pile of boxes
The policy did not change. The low level controller did, and the box count went up.
Neil Nie says a single change to the low level controller was enough to push his team's robot past its previous best at folding boxes.
In a post on X, @neil_nie_ wrote that Allen, the team's controls engineer, implemented a new controller that day. With that change alone, an ablated version of their model, meaning a stripped down version rather than the full one, folded and stacked an entire pile of boxes back to back. Nie called it more boxes than the team had ever managed in a single trial.
More boxes than we've ever managed to fold in a single trial.
What the number actually is
It is a relative claim and nothing more. Nie gives no absolute count of boxes, no success rate, and no comparison to a baseline anyone outside the team can check. More than before is the whole measurement. He also does not say which company or robot this is, what the controller does differently, or what was ablated out of the model.
The interesting part is where the gain came from. The model was the weaker version, not the stronger one. The thing that moved was the controller, the layer underneath the policy that turns intended motion into joint commands. That is a familiar argument in robot learning right now, that a chunk of what looks like a policy problem is really a controls problem, and that better hardware control quietly raises the ceiling on everything stacked above it. It runs alongside the ongoing fight over RL versus inverse kinematics, where researchers have been arguing about how much classical control a learned system should replace.
Not everyone reads the trend the same way. @scav, posting as Mattie Fairchild, offered a hot take that this is a signal frontier models are more likely to abstract away hardware controllers, not lean on them.
My hot take is this is actually a signal frontier models are *more* likely to abstract away hardware controllers
No video, paper or logs accompany Nie's post, so nobody outside the team can verify the result yet.
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.