Unitree G1 humanoid performs a traditional Japanese festival dance learned from video

Hesham Shehata of Arithmer's research and development division in Tokyo has described a pipeline that transfers Awa Odori, a traditional Japanese festival dance, from ordinary video to a commercial Unitree G1 EDU humanoid, in a paper posted on arXiv.

It uses human mesh recovery and kinematic retargeting to turn the video into robot motion, trains a tracking policy with reinforcement learning in simulation and runs it on the robot, with training done on a single RTX 5070 graphics card. Over eleven versions tested on the robot, the author found and fixed silent corruption of deployment files, modelled drift in the state estimator during training and edited the reference motion so the robot could perform stably and end in a controlled way.

The final policies completed a 43-second routine three times in a row without a safety rope, gave a 46-second performance with an automatic handoff and danced for 9.7 minutes without a rope. In a live handover on stage, one remote command switched the robot from the maker's walking controller to the dance and back.