Related Repositories

4.1 MuJoCo Simulation Example

mujoco-sim provides a minimal example for loading and controlling Wuji Hand in the MuJoCo physics simulator.

Run Commands

git clone --recursive https://github.com/wuji-technology/mujoco-sim.git
cd mujoco-sim
pip install -r requirements.txt
python run_sim.py

Expected Result

The script loads the right-hand model (default) and loops through the pre-recorded waving motion trajectory in data/wave.npy. To switch to the left hand, change the side variable in run_sim.py to "left".

4.2 Isaac Lab Simulation Example

isaaclab-sim provides simulation examples for using Wuji Hand in NVIDIA Isaac Lab, supporting GPU-accelerated large-scale parallel simulation, suitable for reinforcement learning training scenarios.

Run Commands

git clone --recurse-submodules https://github.com/wuji-technology/isaaclab-sim.git
cd isaaclab-sim
# Configure the environment according to Isaac Lab official documentation, then execute
python run_sim.py

Expected Result

The script loads the right-hand model (default) and loops through the pre-recorded waving motion trajectory in data/wave.npy. To switch to the left hand, run python run_sim.py --side left.

4.3 Hand Pose Retargeting

wuji-retargeting is a high-precision hand pose retargeting system based on adaptive analytical optimization. It maps human hand poses to dexterous hand joint positions and supports real-time hand tracking with Apple Vision Pro.

Run Commands

git clone --recurse-submodules https://github.com/wuji-technology/wuji-retargeting.git
cd wuji-retargeting
pip install -r requirements.txt
pip install -e .

# Simulation example (using pre-recorded data)
cd example
python teleop_sim.py --play data/avp1.pkl --hand left

# Real robot example (requires Vision Pro connection)
python teleop_real.py --input visionpro --ip <vision-pro-ip> --hand right

Expected Result

In simulation mode, the dexterous hand replays pre-recorded hand pose data. In real-robot mode, it follows the hand movements captured by Vision Pro in real-time. The system uses Huber loss with NLopt SLSQP optimization, adaptively switching between pinch and spread gestures to improve accuracy.