amathislab/musclemimic

What it solves

MuscleMimic addresses the challenge of scaling musculoskeletal motor learning. It provides a benchmark for training AI agents to imitate complex human-like motions (locomotion and manipulation) using biomechanically accurate models that use muscles rather than simple joints, which is typically computationally expensive and difficult to scale.

How it works

The project uses a JAX-based framework with the MuJoCo Warp backend to enable GPU-parallel training across thousands of environments simultaneously. It employs Hill-type muscle models with physiological activation dynamics to simulate realistic muscle behavior. To train these models, it uses DeepMimic-style rewards and retargets motion-capture data (such as AMASS/SMPL and C3D marker data) into trajectories that the muscle-actuated models can follow.

Who it’s for

It is designed for researchers and developers working in embodied AI, biomechanics, and robotics who want to simulate and train full-body musculoskeletal movements at scale.

Highlights

• Muscle-actuated dynamics: Uses physiologically accurate Hill-type muscle models.
• Accelerated training: Leverages JAX JIT compilation and MuJoCo Warp for massive GPU parallelism.
• Generalist imitation policies: Implements validation metrics and rewards for diverse motion datasets.
• Retargeting tools: Includes support for GMR-Fit, AMASS/SMPL motions, and C3D marker data.