AI product Open source
SuperMimic is open-source software from UC Berkeley's Hybrid Robotics group for humanoid whole-body motion tracking and control. It trains a single differentiable world model that captures shared robot dynamics, then improves a policy by propagating gradients through short imagined rollouts and feeding the results back into a physics simulator, rather than relying on model-free reinforcement learning on large GPU clusters. The repository provides generalized-coordinate and maximal-coordinate G1 task configurations, scripts for training the world model and policy and playing back trained models, and is built on mjlab. It is released under the Apache License 2.0.
1 use taken from transcripts — each links to the moment in the video.
Open-source code from UC Berkeley's Hybrid Robotics group for scaling humanoid whole-body control with one differentiable world model. Instead of model-free RL on huge GPU clusters, it learns a single world model capturing shared robot dynamics and improves the policy via gradients through short imagined rollouts, looping back to a physics simulator. Trainable on a single GPU.
1 in the library.