In plain words
Training a humanoid robot controller to follow many different movements can take a lot of simulator runs. SuperMimic learns a model of how the simulated body moves, then uses that model to try out short sequences and improve one controller across many motions.
In the paper’s tests, it reached the tracking score of a published controller using far fewer GPU-hours. The comparison used different hardware and amounts of training data, and the test motions were part of training.
The research
As a humanoid motion library grows, training can become more expensive because model-free methods repeatedly use simulator runs to estimate what actions will do. Can one controller reuse dynamics shared across motions and reduce that training burden?
SuperMimic learns shared humanoid dynamics from simulator data, then uses imagined rollouts and gradients to train one controller across many motions.
SuperMimic alternates between collecting simulator transitions, fitting one differentiable model to them, and improving the controller through short imagined rollouts.
On a shared motion-tracking test of 129,514 filtered BONES-SEED training motions, its normalized return, the full-episode reward sum divided by its maximum possible value, passed SONIC’s released-checkpoint score within three hours on one H100.
The paper reports that this took more than 1,000 times fewer GPU-hours than SONIC’s reported training run, suggesting the approach can reduce compute in this comparison.
The SONIC comparison does not isolate the training method: hardware and training-data volumes differed, and peak GPU use is unknown. The motion test uses training motions, so it measures coverage of those motions, not performance on unseen motions.
Analysis
Deep
- Original paper authors
- Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz
- Submitted to arXiv
- License
- arXiv non-exclusive