This repository reproduces SuperMimic on a dexterous hand, adapting its world-model-based training approach to in-hand cube reorientation with Wuji Hand 2. It is built on mjlab and MuJoCo Warp.

Install uv if it is not already available:

curl -LsSf https://astral.sh/uv/install.sh | sh
source "$HOME/.local/bin/env"

Clone this repository and install the dependencies:

git clone https://github.com/CumulusAlpha/Dex-Supermimic.git
cd Dex-Supermimic
uv sync

uv creates the local Python environment and installs the project dependencies, including the bundled Wuji hand models and meshes. No separate motion dataset or asset download is required.

Run from the repository root:

uv run dex-supermimic-train --num-envs 1024 --device cuda:0

This launches 1024 parallel environments on the first CUDA GPU. Training logs and checkpoints are saved to logs/dex_supermimic/<timestamp>/.

Training logs locally with TensorBoard by default. To use Weights & Biases, append --logger wandb; to disable logging, append --logger none.

Download model_8000.pt from the pretrained checkpoint release and place it at logs/dex_supermimic/pretrained/model_8000.pt. From the repository root:

mkdir -p logs/dex_supermimic/pretrained
curl -fL --retry 3 \
  https://github.com/CumulusAlpha/Dex-Supermimic/releases/download/pretrained-wuji-v1/model_8000.pt \
  -o logs/dex_supermimic/pretrained/model_8000.pt

The pretrained checkpoint is distributed as a Release attachment and is downloaded separately from the source code. The logs/ directory is ignored by Git.

From the repository root, load the pretrained checkpoint in the Viser web viewer:

uv run dex-supermimic-play \
  --checkpoint logs/dex_supermimic/pretrained/model_8000.pt

Playback uses Viser and one environment by default and runs until you press Ctrl+C (--steps 0 is the default). Open the viewer URL printed in the terminal in your browser. To limit playback, pass a positive control-step count such as --steps 2000. Replace the checkpoint path to view your own trained policy.

We thank the SuperMimic authors for the original method.

We thank the authors and contributors of the following repositories:

1. InstinctMJ — the foundation of this repository and its mjlab-based simulation framework.
2. Wuji MJLab — the Wuji hand cube reorientation task and environment implementation.
3. supermimic_vibed — the SuperMimic reproduction.