About the Project
Walking and kicking are goal-directed: the reward is forward progress or ball contact. Dancing is a different problem. There's no simple scalar reward for "danced well." Instead the target is an entire reference motion, and the policy has to track it closely while staying balanced on real hardware.
I used BeyondMimic to take motion capture dance sequences and retarget them onto the Booster T1's own joint limits and proportions, then trained a tracking policy in IsaacLab to reproduce that motion under the robot's real dynamics instead of just replaying keyframes.
This builds directly on the retargeting and sim-to-real pipeline from the kicking project, applied to a much harder tracking problem. Getting this onto real hardware, not just in simulation, is the next step.