A new algorithm enables the training of humanoid robots across 129,514 motions using a fraction of the compute required by current industry standards. The SuperMimic system trains a generalist whole-body tracker in 3 hours with 1 GPU, surpassing a state-of-the-art benchmark that required 128 GPUs running for 7 days. This represents a reduction in total GPU-hours of more than 1000x for the development of full-body robot control.
The system utilizes a ‘world model + APG’ paradigm, which Pieter Abbeel called a game changer for the field. This approach increases efficiency because the world model runs faster than traditional simulations and backpropagation through time captures more signal than policy gradients. The result allows for the creation of a single policy across all movements in a dataset using significantly less hardware and time.
