NaP-Control Accepted to ECCV 2026: Efficient Physics-based Character Control

rsasaki0109 · x · 2026-08-23

The NaP-Control paper has been accepted to ECCV 2026. The method combines RL with a diffusion prior to predict noise directly, avoiding test-time guidance steps, reducing inference costs, and improving control performance. It can also adapt to uneven terrains using a motion prior trained only on flat ground.

Related event: ETH's NaP-Control Combines RL and Diffusion Priors for Efficient Character Control(2 posts)→

Original post →

More from Embodied

Embodied channel →