SuperMimic

Project page | Paper

SuperMimic trains a humanoid whole-body motion tracker with one differentiable world model, built on mjlab.

Setup

Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Clone both repos as siblings:

mkdir supermimic_ws && cd supermimic_ws
git clone git@github.com:HybridRobotics/supermimic.git
git clone git@github.com:HybridRobotics/mjlab_research.git
cd supermimic

Usage

Available task IDs:

• SuperMimic-GenCoord-G1 — generalized coordinates representation
• SuperMimic-MaxCoord-G1 — maximal coordinates representation

Train a world model and policy:

CUDA_VISIBLE_DEVICES=<GPU_ID> uv run src/supermimic/scripts/train.py \
  <TASK_ID> --runner.run_name <description>

Play back a trained model:

uv run src/supermimic/scripts/play.py \
  <TASK_ID> --wandb-path <RUN_PATH>

Acknowledgements

SuperMimic contains code inspired by these projects:

• ControlVAE (BSD-3-Clause)
• robotic_world_model (Apache-2.0)
• DreamerV3 (MIT)

License

Copyright 2026 The SuperMimic Authors.

Released under the Apache License 2.0.