taehoon-kai/supermimic_vibed

How taehoon-kai/supermimic_vibed works

SuperMimic is a humanoid motion-tracking system trained with a learned differentiable world model. Its main workflow loads retargeted G1 reference motions, trains a policy through batched simulation and model-based updates, evaluates saved checkpoints across the clips, then plots scores or renders recorded trajectories. The graph also includes the documented robot-asset conversion path. Relationships are based on the shown excerpts, SOURCE INDEX references, and README; unsampled implementation details are not inferred.

Main entry points: train.py Training Orchestrator · env.py Tracking Environment · motion.py Motion Library · evaluate.py Checkpoint Evaluation

Components 21

Motion and Robot Assets

URDF Conversion tools/urdf_to_mjcf.py
USD Conversion tools/mjcf_to_usd.py
G1 Articulation supermimic/robot.py
Differentiable FK supermimic/kinematics.py
Motion Library supermimic/motion.py

Tracking Simulation

Rotation Math supermimic/rotations.py
Tracking MDP supermimic/mdp.py
Tracking Environment supermimic/env.py
Reset Sampling supermimic/reset_sampling.py

Model-Based Training

Training Orchestrator supermimic/train.py
Replay Buffer supermimic/buffer.py
World Model and Policy supermimic/networks.py

Evaluation and Results

Checkpoint Evaluation supermimic/evaluate.py
Score Plotting tools/plot.py
Trajectory Rendering tools/render.py

Other components

Experimenter
Retargeted Motions
G1 URDF
Isaac Lab / Newton
Evaluation Logs
Recorded Trajectory

Connections 33

Experimenter → Training Orchestrator: starts training README.md
Retargeted Motions → Motion Library: loads clips README.md
G1 URDF → URDF Conversion: compiles README.md
URDF Conversion → USD Conversion: converts model README.md
USD Conversion → G1 Articulation: produces USD asset README.md
G1 Articulation → Differentiable FK: provides model supermimic/kinematics.py
Motion Library → Differentiable FK: computes reference FK supermimic/motion.py
Motion Library → Rotation Math: uses rotation math supermimic/motion.py
Differentiable FK → Rotation Math: uses rotation math supermimic/kinematics.py
Tracking MDP → Differentiable FK: uses FK supermimic/mdp.py
Tracking MDP → Motion Library: reads reference states supermimic/mdp.py
Tracking MDP → Rotation Math: uses rotation math supermimic/mdp.py
Tracking Environment → Isaac Lab / Newton: steps physics README.md
Tracking Environment → G1 Articulation: configures articulation supermimic/env.py
Tracking Environment → Tracking MDP: labels transitions supermimic/env.py
Tracking Environment → Motion Library: resets from motions supermimic/env.py
Tracking Environment → Reset Sampling: updates reset scores supermimic/env.py
Training Orchestrator → Differentiable FK: initializes FK supermimic/train.py
Training Orchestrator → Motion Library: loads library supermimic/train.py
Training Orchestrator → Tracking MDP: builds MDP supermimic/train.py
Training Orchestrator → Tracking Environment: collects rollouts supermimic/train.py
Training Orchestrator → Replay Buffer: stores transitions supermimic/train.py
Training Orchestrator → World Model and Policy: fits and updates supermimic/train.py
Checkpoint Evaluation → Tracking Environment: runs evaluation supermimic/evaluate.py
Checkpoint Evaluation → Motion Library: evaluates all clips supermimic/evaluate.py
Checkpoint Evaluation → World Model and Policy: loads policy supermimic/evaluate.py
Checkpoint Evaluation → Evaluation Logs: writes scores README.md
Checkpoint Evaluation → Recorded Trajectory: records rollout supermimic/evaluate.py
Evaluation Logs → Score Plotting: plots scores README.md
Recorded Trajectory → Trajectory Rendering: renders video README.md
Experimenter → Checkpoint Evaluation: evaluates checkpoint README.md
Experimenter → Score Plotting: requests plot README.md
Experimenter → Trajectory Rendering: requests render README.md

Drawn by GitDiagram on November 10, 2026 from the repository's file tree, README and source files. Read as Markdown ·