SuperMimic trains humanoid control on 1 GPU in 3 hours
A learned world model stands in for the simulator, and the training bill drops by more than three orders of magnitude.
Qiayuan Liao says his team found a way to train humanoid full-body control that cuts GPU-hours by more than 1000x. Across 129,514 motions, he says SuperMimic surpasses the state of the art using 1 GPU for 3 hours, where the prior result took 128 GPUs and 7 days.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
That is the whole claim, and it is about cost, not about a robot doing something it could not do before. The motions are the same kind of motion tracking work the field already does. What changes is that a single policy covering all 129,514 of them stops being a cluster job.
Why it gets that much faster
Pieter Abbeel gave the two-line version of the trick. First, a world model (a learned stand-in for the physics simulator) runs much faster than the simulator itself. Second, because that model is differentiable, you can backpropagate through time instead of relying on policy gradients, which Abbeel says collects far more signal per step of training.
Kevin Zakka framed it as the two approaches working together rather than one replacing the other. The simulator supplies accurate transitions, and the world model turns those into a cheap, differentiable surrogate that the policy optimizer can hammer on. Zakka says you can play with it in your browser.
What to hold back on
The comparison to the previous state of the art comes from the authors. Nobody outside the team has posted an independent rerun, and the sources do not say how the policy behaves on real hardware versus inside the world model. Two well-known researchers vouching for the method is a strong signal, but it is not a replication.
The interesting part for anyone training robots is the shape of the result. Motion tracking at this scale has been a thing you do because you have a cluster. If one GPU and an afternoon gets you a single policy over a six-figure motion set, the group of people who can iterate on full-body humanoid control gets a lot larger. Whether the resulting policies hold up away from the authors' setup is the question the next few weeks should answer.
Now it's superfast (and cheap, just one GPU) to train a single policy on *all* the motions.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.