paper web signal

OneSteer: Frozen humanoid controller clears 90% under 1.25m bar

TL;DR

  • OneSteer, a one-step diffusion whole-body controller, lifts simulated clean passage under a 1.25 m bar from 2% to 90% without weight updates.
  • On a physical Unitree G1, the controller completed all 27 nominal tracking trials across nine diverse whole-body motion sequences without falling.
  • A prediction-variation scoring method ranks likely tracking failures more than 10x faster than simulation-based scoring with five rollouts per clip.

A simulated humanoid walked under a 1.25 m bar 2% of the time. After the same controller, weights unchanged, was handed a cost function describing the barrier, it cleared the bar in 90% of runs.

That is the headline figure in OneSteer, a paper describing a one-step diffusion whole-body controller for humanoid robots. The method's pitch is that the controller stays frozen at deployment time, and new objectives arrive as differentiable task costs that optimize the generator's input noise instead of the policy weights.

'To accommodate new objectives, differentiable task costs guide optimization of the generator's input noise with frozen weights, while body-group reference masks relax selected tracking requirements,' the authors write.

On a physical Unitree G1, they report 27 completed nominal tracking trials across nine diverse whole-body motion sequences without a fall. With a steering cost built from a signed distance field derived from onboard LiDAR, the robot passed under a 1.1 m low bar in four of five attempts and stepped over a 12 cm hurdle in all three attempts.

The paper also claims its prediction-variation scoring ranks likely tracking failures 'more than 10x faster than simulation-based scoring with five rollouts per clip.' The abstract does not report baseline steering numbers from competing whole-body controllers, nor what happened in the one 1.1 m attempt that did not succeed.