SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
A UC Berkeley controller that learns inside a differentiable world model beat the return of the published SONIC humanoid motion tracker after 3 hours on one NVIDIA H100, where SONIC reported 128 L40 GPUs for 7 days, about 21,000 GPU-hours, the authors report in a preprint accepted at CoRL 2026. It failed on 25 of 129,514 benchmark motions against 179 for SONIC's released checkpoint, reached zero failures after 9.8 hours on two GPUs, and ran flips and rolls on a real Unitree G1.
Can one learned, differentiable model of humanoid dynamics replace cluster-scale model-free reinforcement learning when training a single controller to imitate a very large library of human motions?
Tracking big motion-capture libraries is now the main route to general humanoid whole-body control, but the cost grows with the library. Current pipelines either train many per-motion specialists and distil them into one policy, losing quality, or train one generalist with model-free reinforcement learning on a GPU cluster.
SONIC, published in Science Robotics in 2026 and built on the same BeyondMimic tracking formulation, trains one PPO policy on 611 hours of motion with 128 NVIDIA L40 GPUs for about 168 hours. It is the closest published comparison and the main baseline.
SuperMimic fits one world model of the robot's dynamics shared across all motions, learning it from logged transitions of a standard rigid-body simulator (MuJoCo), and trains the tracking policy in the same loop. The policy gets analytic gradients by backpropagating through short imagined rollouts of the world model, which are fast and differentiable on a GPU. Training used 264.56 hours of the public BONES-SEED motion set retargeted to the Unitree G1, filtered to 129,514 clips.
Within 3 hours on one H100 the policy surpassed the normalised return of SONIC's released checkpoint, about 56 times less wall-clock time and about 7,200 times fewer GPU-hours by the authors' accounting. On the shared 129,514-motion benchmark it failed on 25 motions after 3 hours on one GPU and on none after 9.8 hours on two GPUs, against 179 failures for SONIC. The world model's prediction error stayed between 0.007 and 0.011 as the training set grew from 1 to 264 hours, and GPU-hours to a target return grew slowly. A single tracker ran a skipping jump, a side roll, a 90-degree flip and an in-place 360-degree flip on a real Unitree G1 without per-motion tuning.
This is a preprint, though accepted at CoRL 2026. The comparison crosses GPU types and data sizes: SONIC used 2.3 times more motion data, and an H100 has up to 5.5 times an L40's peak tensor-core throughput; the authors argue these cannot explain a 1,000-fold gap, but it is not a matched rerun. The world model learns only from simulator data, training covers flat ground only with no terrain or object interaction, and the real-robot test shows four motions rather than a measured success rate. The work was funded in part by Komatsu, Apple and the RAI Institute.
If whole-body control can be retrained in hours on one GPU, compute stops being the gate for humanoid makers that want broad, expressive motion, from Unitree and its G1 customers to start-ups without access to large clusters. It also makes scaling-law studies on motion data affordable, which could shift competition toward who owns the best motion libraries. The near-term limit is that factory and home work needs terrain and object contact, which the authors list as ongoing work; watch for that extension and for real-robot data being folded into the world model.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.