Isaac Gym Envs for DexHand

This repository provides a framework for training dexterous grasping policies for robotic hands using NVIDIA's Isaac Gym simulator.

Installation

1. Download and install Isaac Gym Preview 4 from NVIDIA's website

2. Verify Isaac Gym installation:

cd isaac-gym/python/examples
python joint_monkey.py

3. Clone and install this repository:

git clone https://github.com/dexrobot/dexrobot_isaac
cd dexrobot_isaac
pip install -e .

Running

Training

python DexHandEnv/train.py task=DexCube num_envs=4096 headless=True

• num_envs: Number of parallel environments (default: 4096)
• headless: Run without visualization for faster training

Testing

To test a trained model:

python DexHandEnv/train.py task=DexCube test=True num_envs=2 checkpoint=runs/DexCube_*/nn/DexHand.pth
python DexHandEnv/train.py task=DexCube test=True num_envs=2 checkpoint=runs/DexCube_06-14-42-29/nn/DexCube.pth

Configuration

The environment and training parameters can be customized through config files:

• Environment config: DexHandEnv/config/task/DexHand.yaml
• Training config: DexHandEnv/config/train/DexHandPPO.yaml

Video Recording

To capture training videos:

python DexHandEnv/train.py task=DexHand capture_video=True capture_video_freq=1500 capture_video_len=100

Multi-GPU Training

For distributed training across multiple GPUs:

torchrun --standalone --nnodes=1 --nproc_per_node=2 DexHandEnv/train.py multi_gpu=True task=DexHand

License

This project is licensed under the Apache License.

Acknowledgements

This work builds upon the Isaac Gym framework developed by NVIDIA.