SuperMimic on mjlab: Model-Based Humanoid Motion Tracking
Differentiable-world-model motion tracking on native MuJoCo-Warp.
Reproduced SuperMimic’s differentiable-world-model approach to humanoid motion tracking on native MuJoCo-Warp through mjlab, achieving 100% full-clip completion across 40 LAFAN1 motions within 5 hours of training on a single RTX 5090. Demonstrated few-shot adaptation of the pretrained policy to highly dynamic MotionCode motions.
Implementation and evaluation
-
Native mjlab implementation: Adapted world-model learning and differentiable policy rollouts to the native MuJoCo-Warp backend, with consistent body-frame observations and goal-conditioned termination modeling.
-
Few-shot policy adaptation: Adapted the pretrained tracking policy to a small set of highly dynamic MotionCode motions through fine-tuning, with mixed-frame-rate data processing and state-preserving checkpoint migration.
-
Full-clip evaluation: Built evaluation tools that measure motion completion and tracking quality separately. Videos show selected policy rollouts alongside the blue reference motion.