HybridRobotics Releases Code for SuperMimic, Full-Body Motion Tracking Using a Differentiable World Model

HybridRobotics has posted code for SuperMimic on GitHub. The system trains humanoids for full-body motion tracking using a single differentiable world model. Built on mjlab, it provides two G1 tasks: one using a generalized-coordinate representation and the other using a maximal-coordinate representation.

SuperMimic provides code for training full-body motion tracking and two types of tasks, using generalized and maximal coordinates. This gives researchers and developers a starting point for exploring implementations with different coordinate representations. However, the repository does not report validation results on physical robots or evaluation metrics, so the availability of the code alone is not enough to assess its practical performance.

The repository’s technical focus is on incorporating a world model into motion-tracking training and handling two coordinate representations: generalized coordinates and maximal coordinates. Separate task IDs distinguish the two representations, allowing users to select different configurations within the same SuperMimic framework. However, the repository does not report results comparing their accuracy or training efficiency, so the tasks’ availability alone does not establish which performs better. On the implementation side, the repository separates the training and playback scripts and provides instructions for specifying a training run name and the run path to play back. This gives users a workflow for running the code, but the source material describes only how to execute it. It does not mention distribution of trained models, the contents of the motion data used for training, evaluation conditions or deployment on physical robots. The released code should therefore not be equated with a ready-to-use tracker or a real-robot control system. The instructions also call for cloning the “mjlab_research” repository alongside mjlab. This suggests that the code is intended to run in combination with related code rather than as a self-contained SuperMimic package. The versions of dependent repositories and the environment requirements will therefore also need to be checked when assessing reproducibility. The Apache License 2.0 provides information about the terms of use, but the repository’s description alone does not clarify the scope of code use or the relationship between the licenses of ControlVAE, robotic_world_model and DreamerV3, which it cites as inspirations. The industry implications should be limited to the fact that the code presents a world-model-based training approach and task design using separate coordinate representations. Assessing its practical value will require information on the motion data used for training, how differences between simulation and real-world operation are handled, and the conditions used to evaluate tracking accuracy and stability. The next points to verify are reproducibility of the released code, training data and evaluation metrics, and whether it has been tested on physical robots. Until those are reported, the approach’s practical utility and advantages over other methods remain undetermined.

The task IDs listed in the repository are “SuperMimic-GenCoord-G1” and “SuperMimic-MaxCoord-G1.” The training script is “src/supermimic/scripts/train.py,” and the script for playing back a trained model is “src/supermimic/scripts/play.py.” For playback, users specify the run path with “--wandb-path.” The instructions show “mjlab_research” being cloned alongside “supermimic.” The repository says its code draws inspiration from ControlVAE, robotic_world_model and DreamerV3. It is released under the Apache License 2.0.