Humanoid whole-body control increasingly relies on tracking large human-motion datasets, but scaling this approach has been costly. Existing pipelines use model-free reinforcement learning, either merging per-motion specialists via distillation or training one policy with cluster-scale computation. SuperMimic instead trains a single differentiable world model of the shared dynamics, fitted from logged transitions of a standard rigid-body simulator, and couples it with a tracking policy in one learning loop.
By backpropagating through short imagined rollouts, the policy receives analytic gradients without the sample-inefficiency of model-free RL. This makes training dramatically cheaper: SuperMimic surpasses a state-of-the-art motion-tracking baseline in 3 hours on a single GPU, against 7 days on 128 GPUs — a more than 1000-fold reduction in GPU-hours. It fails on only 25 of 129,514 benchmark motions, versus 179 for the baseline, and reaches zero failures after 9.8 hours on two GPUs.
The authors also report that world-model prediction error remains stable as motion data grows from 1 to 264 hours, and that GPU-hours needed to reach a target return grow only slowly. They validate the approach on a real Unitree G1 robot across diverse and acrobatic motions. SuperMimic, accepted as an oral at CoRL 2026, positions compute as no longer the bottleneck to scaling humanoid whole-body control, opening the door to systematic scaling-law studies. Since this summary draws on a single paper, all claims reflect that source alone.