Berkeley's SuperMimic trains humanoid control on 1 GPU in 3 hours, cutting GPU-hours by 1000x
alex_verem · x · 2026-11-15
A UC Berkeley team published SuperMimic (arXiv:2610.28480, CoRL 2026), which trains humanoid whole-body control with a single differentiable world model, cutting compute by over 1000x:
- Cost: 3 hours on one H100 surpasses the return of the previous SOTA baseline (128 L40 GPUs for 7 days)
- Fewer failures: only 25 failures across 129,514 benchmark motions vs 179 for the baseline; zero failures after 9.8 hours on two GPUs
- Method: instead of model-free RL with distillation or cluster-scale brute force, it fits one differentiable world model of the shared dynamics from rigid-body simulator logs, training jointly with the tracking policy; analytic gradients come from backprop-through-time over short imagined rollouts
- Real robot: the policy transfers to a physical Unitree G1, doing side rolls and 360° flips
- Limits: flat ground only, though world-model prediction error stays stable from 1 to 264 hours of motion, making motion scaling laws practical to measure
Takeaway: training humanoid whole-body control no longer requires a GPU cluster.
Related event: UC Berkeley's SuperMimic Trains Humanoid Control on a Single GPU in 3 Hours(2 posts)→
More from Embodied
- RaiSim vs Newton benchmark: CPU robot sim beats GPU by up to 95x at practical batch sizes — ChongZzZhang · 2026-11-15
- Researcher teaches a Unitree G1 humanoid to ride a mechanical bull after CoRL — RemiCadene · 2026-11-15
- PrimeSim: world model trained on 30k hours of robot demos runs near real-time on one A100 — jiqizhixin · 2026-11-15
- Viral claim: Claude runs a robot arm on a real factory task with zero training — but the model name doesn't exist — IgorCarron · 2026-11-15
- NEO-VLA ditches visual encoders for robot actions, hitting 90.9% on LIBERO frozen — 机器之心 · 2026-11-15
- Zipline's Keller Cliffton on 3M drone deliveries and what breaks at 10x scale — a16z Podcast · 2026-11-15