ETH Zurich Introduces NaP-Control: Merging RL with Diffusion Priors for Fast Character Control

rsasaki0109 · x · 2026-08-23

ETH Zurich presents NaP-Control, a latent noise optimization framework combining reinforcement learning (RL) with a task-agnostic diffusion motion prior. Instead of relying on slow, iterative gradient guidance at test-time, it learns to navigate the input noise space of the diffusion prior to generate task-driven, physically plausible behaviors, achieving efficient and precise whole-body character control.

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

Original post →

More from Embodied

Embodied channel →