Generate the reference
A diffusion model uses terrain, a target direction, and recent robot states to generate short-horizon motion references from retargeted human demonstrations.
Learning Whole-Body Humanoid Locomotion via Motion Generation and Motion Tracking
Human motion provides the skills; terrain-aware generation and tracking adapt their execution to the robot’s surroundings.
Open the original interactive walkthrough and transcript ↗
This video is embedded from Takahiro Miki’s paper explainer. The original page includes interactive animations, a transcript, and a detailed walkthrough with the paper’s figures.
Reading the results
The quoted 23.0% and 96.2% success rates refer to the 80 cm box with 0° obstacle yaw in simulation, using 500 instances per setting. Hardware footage demonstrates deployment on the Unitree G1; the paper reports those demonstrations qualitatively.
The key ideas
Execute and adapt
An RL-trained tracker follows the references using proprioception and terrain observations. It is fine-tuned in closed loop with the generator frozen, improving execution when generated references are imperfect.
Read the evidence
The fixed-reference comparison is a simulation result in Table III of the revised paper. Hardware experiments demonstrate box climbing, hurdle traversal, stairs, and mixed terrain with perception and computation onboard.