Overview
MIMIC-MJX is a framework for neuromechanical emulation of animal behavior using massively parallel imitation learning in MuJoCo-MJX. It enables GPU-accelerated physics simulation of biomechanically realistic virtual animals that learn to reproduce naturalistic behavior from motion capture data.
The project encompasses two core tools:
- track-mjx — GPU-accelerated neuromechanical pose estimation
- stac-mjx — Skeletal tracking and calibration for MuJoCo body models
Key contributions
- Massively parallel imitation learning pipeline that scales across GPU clusters
- Biomechanically realistic mouse body models with musculoskeletal actuation
- Physics-aware constraints that promote naturalistic muscle activity patterns
- Transfer learning from kinematic replay to task-driven behavior
Publication
MIMIC-MJX: Neuromechanical Emulation of Animal Behavior C.Y. Zhang*, Y. Yang*, A. Sirbu, et al. — Under Review at Nature Methods
Code
- talmolab/track-mjx — Main codebase
- talmolab/stac-mjx — Skeletal tracking and calibration