SWAP: Test-Time Steerable Humanoid Whole-Body Collision Avoidance

Zeji Yi^1,2*, Guanqi He², Xiaofeng Guo², Guannan Qu¹, Pieter Abbeel², Yan Duan², Karen Liu², Guanya Shi^1,2

¹Carnegie Mellon University ²Amazon FAR

^*Work done at Amazon FAR

Videos of the Unitree G1 ducking, sidestepping, and twisting past obstacles will appear here.

Paper: coming soon
Code: coming soon
Videos: coming soon

SWAP is a hierarchical policy that lets a humanoid robot avoid obstacles with its whole body. A diffusion planner proposes many collision-aware motions, a reinforcement learning controller keeps the robot balanced, and simple scoring functions choose among the proposals at test time, so the robot can change its goal without retraining.