SuperMimic: humanoid motion tracking on 1 GPU in 3 hours

What is new?

UC Berkeley researchers introduced SuperMimic, a whole-body motion tracking controller for humanoid robots (project page). Authors include Pieter Abbeel and Koushil Sreenath, and the paper was accepted to CoRL 2026 (arXiv).

The headline claim: after 3 hours of training on one NVIDIA H100, SuperMimic surpasses the normalized return of SONIC's released checkpoint. SONIC trained for 7 days on 128 NVIDIA L40 GPUs, about 21,000 GPU-hours. By the authors' accounting, that is more than 1,000x fewer GPU-hours and more than 50x less wall-clock time.

How it works

Most current pipelines use model-free reinforcement learning. They either merge per-motion specialists through distillation or brute-force a single policy with cluster-scale compute. SuperMimic instead learns one differentiable world model of the dynamics shared by all motions, fit from logged transitions of a standard simulator (MuJoCo Warp).

The world model and the policy train together in one loop. The policy gets analytic gradients through short imagined rollouts, which makes learning fast.

Results

• On the 129,514-motion benchmark it fails 25 motions, versus 179 for SONIC.
• After 9.8 hours on two GPUs it fails none.
• The world model's prediction error stays stable as training data grows from 1 to 264 hours of motion.
• The policy transfers to a real Unitree G1 robot, including acrobatic motions.

Caveats

The comparison is not like for like. SuperMimic trained on the 264.56-hour public subset of BONES-SEED, SONIC on the full 611 hours. The hardware differs too: H100 versus L40. The authors argue these gaps cannot explain a 1,000x difference, but it is not a matched rerun.

If the result holds, training humanoid controllers comes within reach of smaller labs. The authors say it also makes motion-tracking scaling laws practical to measure.