Event

DiffWBC: One-Step Diffusion for Humanoid Whole-Body Motion Tracking

1 reports · 1 sources

Yifei Huang and other authors released a paper proposing DiffWBC, which changes general humanoid whole-body motion tracking from a deterministic policy to a one-step diffusion model: each 50 Hz control step requires only one forward inference, achieving a local tracking error of 20.1 mm on the BONES benchmark, and it has been deployed on a real Unitree G1. The method can be distilled from a reinforcement learning teacher model, or trained from scratch via model-based reinforcement learning; during deployment, a frozen controller is adjusted online using differentiable cost gradients and body-group masks. The authors state that targeted fine-tuning on annotated failure segments can resolve 50% of these failure cases.