SuperMimic: Training Humanoid Motion Control on One GPU

Can a differentiable world model make large-scale humanoid motion tracking a single-GPU experiment? This episode explains SuperMimic's training architecture, its comparison with SONIC, and what the experiments establish. We cover three questions: how reference motions and simulator transitions provide different kinds of supervision; why differentiating through predicted trajectories adds value beyond faster rollouts; and what a fair compute comparison should measure. The discussion follows simulation-based whole-body control using MuJoCo Warp and G1-retargeted BONES-SEED motions, alongside the authors' physical Unitree G1 demonstrations. We connect the design to SuperTrack, SuperDyno, and short-horizon actor–critic learning, then examine what the approach offers researchers studying motion diversity, policy optimization, and reusable robot dynamics. The central limitation is the unmatched headline comparison: different training setups prevent the reported GPU-hour gap from isolating an algorithmic speedup. For roboticists, the practical question is how to turn shorter training cycles into better experiments. Could learning the body's dynamics change what your next humanoid project can afford to investigate?