SuperMimic: Berkeley Trains a Humanoid Motion Tracker in 3 Hours on One GPU

A Berkeley team says it can train a humanoid whole-body motion tracker on 129,514 motions in 3 hours on a single GPU, surpassing a state-of-the-art tracker that took 7 days on 128 GPUs. The method, SuperMimic, comes from Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel and Koushil Sreenath, and was accepted as an oral at CoRL 2026. The sources are the arXiv paper, SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model, and the project page.

All numbers below are the authors' own, run by them against a released checkpoint of their comparison system. The paper is a preprint submitted October 20, 2026, though the CoRL acceptance means it has passed peer review.

TL;DR

The problem: scale is expensive

Tracking large human-motion datasets has become the main route to general, expressive humanoid whole-body control: a policy learns to reproduce reference motions such as walking, flips, rolls and dancing, and later a higher-level system can ask for them. Scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation or brute-forcing a single policy with cluster-scale compute. The comparison system here, SONIC, followed the second approach. For another line of whole-body motion work, see our coverage of NVIDIA's MotionBricks on Unitree G1.

The idea: learn the physics once, then imagine

SuperMimic fits one differentiable world model of the shared dynamics of all motions from logged transitions of a standard rigid-body simulator (MuJoCo Warp). The world model and the tracking policy train together in one loop, and backpropagation through time over short imagined rollouts gives the policy analytic gradients, instead of the noisy policy gradients used by methods such as PPO.

The project page answers several questions that matter for understanding it:

• It does not need a differentiable simulator. The simulator is never differentiated; the learned model is, so simulators with non-smooth contacts work as they are.
• The world model is not frozen. It updates every iteration alongside the policy and is held fixed only inside each policy update.
• Rollouts are short and fast. Imagined rollouts are 32 steps, with a critic bootstrapping beyond that. With the critic, a 16-step horizon matches a critic-free 32-step one; without it, training fails at 16 steps or fewer. The page says imagination is more than 10x faster than MuJoCo Warp at practical batch sizes, and first-order gradients reach PPO's best training return with about 64x fewer imagined steps than PPO on world-model rollouts.
• The model scales with data. From 1 to 264 hours of motion, world-model prediction error stays stable (0.007 to 0.011) and GPU-hours to a target return grow slowly.

The results

All systems are scored on the same 129,514 motions, with normalized return from 0 to 1. SONIC's released checkpoint scores 0.90.

The failure table is revealing. SONIC's released checkpoint fails every flip (all 44 motions in the 360, 180 and 90 degree groups) and every side roll (all 22), plus 9 of 10 crossed-leg motions, 28 of 99 lying, resting and standing-up motions, 15 of 24 broad jumps, and all 8 in a dance routine. After 3 hours on one GPU SuperMimic fails 2 side rolls and 1 flip, 8 lying motions and 14 of the rest; after 9.8 hours on two GPUs it fails none.

Is the comparison fair?

The authors raise this themselves. The two runs differ in data and hardware: SONIC used 128 L40 GPUs and 611 hours of motion; SuperMimic used one H100 and 264.56 hours of the public BONES-SEED subset, which is 2.3x less data. An H100 delivers 0.74x the fp32 and 5.5x the tensor-core peak of an L40, they note, and how much of that peak either run used is unknown. Their argument: "2.3x in data and at most 5.5x in peak GPU throughput cannot account for a 1000x gap in GPU-hours."

Readers should still keep three limits in mind:

1. Coverage, not generalization. The benchmark motions are training motions for both methods, so failure counts measure how many training motions a policy can reproduce, not how well it handles unseen ones.
2. Simulation conditions. Each motion plays once from its first frame with foot friction 1.0, no domain randomization or observation noise, and deterministic physics. A motion fails if the rollout triggers SuperMimic's own training termination condition, which is not necessarily how other groups would judge failure.
3. The baseline is one released checkpoint. SONIC's number is from its published training run; a modern re-tuned baseline might narrow the gap.

None of this erases the result, which is a large efficiency gain, but "1000x" is best read as an order-of-magnitude claim from the authors, not a certified measure.

Sim-to-real on the Unitree G1

The authors show a single policy running on a physical Unitree G1 across diverse and acrobatic motions, with no per-motion tuning, "trained in imagination" and transferred zero-shot. These are curated clips on the project page, so they show feasibility rather than a success rate on hardware. The page also hosts an in-browser demo where you can choose a motion and run the policy. In one visualization, switching the policy off makes the robot fall and the world model predicts the fall; in another, making the floor slippery makes the feet slide, and the world model, told the friction, predicts the slide.

A plain-language explanation of the method

Imagine teaching a gymnast 129,514 routines. The model-free approach is to have the gymnast try each routine in a real gym millions of times, rewarding successes and punishing falls, and to nudge technique from those noisy outcomes. SuperMimic instead builds a fast mental model of how bodies move, learned from watching the gym's physics, and lets the gymnast rehearse inside that mental model. Because the mental model is differentiable, the gymnast gets precise feedback about which small change in a muscle command would have improved the landing, not just a verdict after the fall. That precise feedback is the "analytic gradient," and it is why fewer attempts are needed.

Two things keep this from being too good to be true. First, the mental model keeps learning from fresh real simulator data every iteration, so it does not drift far from the physics, and the project page shows the model's predictions staying close to the simulator even for falls and slides. Second, rehearsals are short, 32 steps at a time, with a critic estimating what happens afterward, which limits how far errors in the mental model can compound.

A reasonable skeptic's question is whether a learned model of rigid-body physics can survive contact with a real robot. The authors' answer is the G1 clips and the claim of zero-shot transfer, and their evidence is the demonstration videos plus the fact that the policy was trained against a standard simulator's data, the same route other sim-to-real pipelines use. Readers who care about real-world reliability should look for per-motion hardware success rates in the full paper.

Why it matters

Compute has been the barrier to measuring scaling laws for motion tracking: if one experiment costs 21,000 GPU-hours, few labs can run many. If SuperMimic's efficiency holds up, university labs could explore data scale, architectures and motion datasets on a single GPU, and the authors explicitly say the method makes it "practical to measure motion-tracking scaling laws." It also fits a broader move toward learned world models as training environments, which we explain in what world models are and in coverage of NVIDIA's Cosmos 3 physical AI world model. Data matters too: see the HIW-500 humanoid dataset for LeRobot and Gemini Robotics 2's whole-body work.

What to watch

• Code and checkpoints. We did not find a code repository in the sources we checked, so reproduction depends on a release.
• Generalization. Held-out motions, new robots, and tasks that interact with objects, not only motion tracking.
• Independent hardware results. More than clips: success rates over many trials on real robots.
• Comparisons with other trackers. The same benchmark run against other recent whole-body systems.

Related reading

• NVIDIA MotionBricks on Unitree G1
• What are world models?
• NVIDIA Cosmos 3 world model guide
• HIW-500 humanoid dataset
• Gemini Robotics 2 whole-body intelligence
• Official: SuperMimic paper on arXiv and project page

Details are accurate as of October 22, 2026 and come from the paper and project page. Results are the authors' own.