SuperMimic trains a humanoid controller in under ten hours, then takes it to a real robot

UC Berkeley researchers say SuperMimic trained a humanoid robot control system in 9.8 hours on two H100 GPUs, then deployed it on a real Unitree G1. The result pairs a strikingly short training run with a system that, according to the project, completed every motion in its benchmark.

Berkeley BAIR Watch analysis

What happened

SuperMimic uses a single differentiable world model to train whole-body control, working in an imagined environment rather than differentiating the simulator directly. The researchers say it matched or surpassed existing motion-tracking methods such as SONIC, and achieved a 100% success rate across the benchmark motions.

The team’s SuperMimic project page describes both the training result and deployment on the Unitree G1. These are the researchers’ reported results; the supplied account does not give further benchmark detail for judging how broadly they generalise.

Why it matters

Humanoid robots are often introduced with impressive hardware and a rather less impressive answer to “how long did it take to teach the thing?” A reported training time of 9.8 hours on two H100s makes the method worth attention, especially alongside a real-robot deployment rather than a result confined to simulation.

If the approach holds up beyond the reported benchmark, it could make it faster to develop controllers for complex, whole-body movement. That is useful progress in robotics, though a clean sweep of benchmark motions is not the same as a robot coping with every untidy thing a real environment can throw at it.

Our read

This is a meaningful robotics result: the combination of short reported training time, strong benchmark performance and real hardware testing gives SuperMimic more substance than a simulation-only promise. The next question is not whether the benchmark is impressive, but how well the method transfers to new tasks and less rehearsed conditions.

What to watch

• Whether the researchers publish further detail on the benchmark motions and comparisons.
• How reliably SuperMimic handles tasks outside the reported benchmark.
• Whether the training approach transfers to other humanoid platforms or movement tasks.

Discussion spark: For judging humanoid control research, which deserves more weight: fast training, benchmark success, or reliable performance on unfamiliar tasks?

Independent WittyWires tracker for public updates about Berkeley BAIR. Not affiliated with or endorsed by Berkeley BAIR; this is not an official account.