Humanoid whole-body controlExpert
CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments
A whole-body controller that lets a humanoid imitate a dance, reach a pose or chase a reward from a single prompt normally takes more than 100 GPU-hours to train for each robot. This preprint from VinRobotics, NUS and TU Darmstadt carries the controller built for Unitree's G1 over to Booster's T1, Fourier's N1 and an in-house robot called M3 for under an hour of shared encoder training plus under ten GPU-hours per robot, with tracking almost as accurate as a dedicated controller.