HybridRobotics/supermimic

What it is

SuperMimic is a research implementation of SuperTrack, a supervised-learning approach to motion tracking for physically simulated characters. It trains a world model together with a policy so that a simulated humanoid mimics recorded motion data. The code is built on a research fork of mjlab (mjlab_research) and targets the G1 humanoid, with two task variants: generalized coordinates (SuperMimic-GenCoord-G1) and maximal coordinates (SuperMimic-MaxCoord-G1) (README).

It is aimed at robotics and graphics researchers who already work with RL and physics-based character animation, not at end users who want a ready-made animation tool.

Setup

You need a CUDA-capable GPU and Git access to both repositories. The README uses uv to manage the Python environment; install it with the provided script if it is not already available:

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

Then clone SuperMimic and its mjlab fork as sibling directories:

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

The clone commands use SSH URLs, so a configured SSH key for GitHub is required (README).

Training and playback

Pick a GPU and a task ID, then run the training script with a name for the run:

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

To replay a trained model, point the play script at the Weights & Biases run path:

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

(<GPU_ID>, <TASK_ID>, <RUN_PATH> and <description> are placeholders.) Playback takes a W&B run path, so experiment logging through Weights & Biases is part of the workflow (README).

Constraints

• Runtime assumes CUDA (CUDA_VISIBLE_DEVICES is set in the training command); CPU-only setups are not described.
• The dependency is the mjlab_research fork, not upstream mjlab; the code expects the two repositories to sit side by side.
• Motion data, checkpoint and W&B configuration details are not covered in the README; the setup and usage sections are the only documented entry points.
• Released under Apache 2.0; the code draws on ControlVAE (BSD-3-Clause), robotic_world_model (Apache-2.0) and DreamerV3 (MIT), whose terms also apply to the borrowed parts.

HotRepo's take

SuperMimic appeared in HotRepo's new-project list on 2026-11-11 with 107 stars gained over a roughly 13.5-hour window from 2026-11-10T10:46:56Z to 2026-11-11T00:15:38Z, alongside 7 forks. There is no GitHub daily-rank record and no HN discussion for this entry. The numbers describe a single early snapshot of a research repository published the day before, and there is not yet a same-metric series long enough to say anything about how sustained that interest is.