A Humanoid Motion Controller Trained in Three Hours on One GPU

Ask a humanoid robot for a backflip and the flip is not the hard part. The hard part is everything around it: the hundred thousand other movements the same controller has to hold at once, so that one machine can walk, kneel, roll and get back on its feet without a separate program for each. The field's answer has been to buy its way through. SONIC is NVIDIA's motion tracker and the strongest published system of its kind. It trained for seven days on 128 graphics processors, or GPUs.

On Oct. 20, 2026, researchers at the University of California, Berkeley posted a paper describing a controller that reaches the same score after three hours on one GPU. Qiayuan Liao led the work, with Pieter Abbeel and Koushil Sreenath sharing the senior-author role, and they call the system SuperMimic. Its claim is about the bill rather than about the robot: three hours of one chip's time against the roughly 21,000 GPU-hours SONIC's own paper reports.

What changes is where the practice happens. The usual pipeline is model-free reinforcement learning: the controller tries a movement in a physics simulator, gets a score, and the learning algorithm works out which way to nudge its weights by averaging over many noisy attempts. The direction it gets is an estimate, and a better estimate means more attempts, which is what the cluster is for.

SuperMimic adds a second network, a world model, fit to the simulator's own logged transitions: given a state and a command, it predicts the next state. Because that stand-in is itself a network, the exact direction to adjust can be read straight out of it. So the training plays out short imagined runs inside the world model and traces the error back through them, taking a precise gradient where the older pipeline could only approximate one. Model and controller learn in a single loop, the model updating every iteration. The authors are explicit that no gradient ever passes through the simulator, which stays an ordinary one.

The benchmark measures coverage, not new ground

The comparison runs on 129,514 motions, a filtered slice of BONES-SEED, a public motion-capture library. Each one plays once from its first frame under identical physics for every controller, and counts as a failure if the run trips SuperMimic's training-termination condition, the rule that cuts off a bad episode during training. At the three-hour mark SuperMimic fails 25 of them. SONIC's released controller fails 179.

The totals hide the shape of the gap, which the authors published in full on their project page. It sits almost entirely in 207 unusually hard motions. There SONIC's checkpoint fails 126, among them every flip in the library and every side roll, against SuperMimic's 11. Across all the other motions the two are close, 53 failures against 14. So the gap is not a broad advantage spread across the library. It is that the cluster-trained controller cannot do the acrobatics at all, and this one largely can.

The test also asks a narrower question than it appears to. Both systems trained on these motions, so the counts measure how completely each has mastered its own library, not how either handles a movement it has never seen. The authors put it as coverage rather than generalization. SONIC's paper leads with the opposite measure, holding motions back and testing on those, and this comparison does not use those splits.

Where the two training runs are not alike

On the evaluation side the comparison is like-for-like: the same motions, the same simulator, one failure rule for both. The training side is not, and the authors set out the differences themselves, in a section of their own page asking whether the comparison with SONIC is fair. SuperMimic trained on 2.3 times less motion data. The two runs used different chip models, and how much of either chip's peak either run reached is unknown. SONIC's compute figure is taken from its paper rather than re-measured, and what SuperMimic was scored against is SONIC's released checkpoint. Volunteering every one of those is to their credit, and the data gap cuts their way.

The metric belongs on that list too. Normalized return, the score the comparison turns on, is SuperMimic's own measure. SONIC reports joint-position error and a tracking success rate, and no return at all. The 0.90 that SuperMimic clears is SuperMimic's measurement of SONIC's released controller under SuperMimic's reward, and the failure rule is SuperMimic's as well, applied to both sides. SuperMimic is reported as above that bar, with no figure for how far above. Running one criterion symmetrically is the defensible way to do a head-to-head, and the authors say plainly that is what they did. But both numbers came from the group proposing the method, and no outside group has re-run the comparison in print.

Zero failures arrive after 9.8 hours on two GPUs, and that result belongs to that operating point alone, not to the three-hour one. Underneath it is the finding the authors care about most: from one hour of motion data up to all 264, the world model's prediction error stays about where it started, and the GPU-hours needed to reach a given score grow slowly. That is what they mean by removing compute as the bottleneck. It is a claim about what a laboratory can now measure, not about humanoid control being solved.

The robot in the clips is real; the numbers off it are not published

The world model is fit to logged transitions of MuJoCo Warp, a standard rigid-body simulator, which makes it a fast stand-in for a simulator and not for the world. A physical robot is where that distinction would show. One policy, trained entirely inside the learned model, runs on a real Unitree G1 across several families of movement, with no tuning for individual motions and no retraining for the hardware. That is more than a clip and more than an intention. It is also not a measurement: no success rate on hardware, no tracking error and no trial count has been published, so what the footage shows is a demonstration.

The paper has been accepted at the 2026 Conference on Robot Learning, so it has been through peer review; the proceedings version is not out, and the public artifact is the arXiv posting. The authors' own record adds that it was accepted for an oral presentation. The motion library behind the benchmark is public and the filtering is described, which means a group with one GPU and an afternoon can now try to take the result apart. On the authors' own framing, that is the point.