LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand

This repository contains the URDF, IsaacGym environment and sim2real deployment code for the paper "LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning" (https://arxiv.org/abs/2309.06440).

Installation

Setup a conda environment (optional)

conda create -n leapsim python=3.8
conda activate leapsim

Install Pytorch using these instructions

Download the Isaac Gym Preview 4 release from the website, then follow the installation instructions in the documentation

cd isaacgym/python
pip install -e .

Clone and install leapsim python packages

git clone https://github.com/leap-hand/LEAP_Hand_Sim
cd LEAP_Hand_Sim
pip install matplotlib gitpython numpy==1.20.3 wandb
pip install -e .

Running a pretrained policy

You can run a pretrained in-hand reorienation policy to check your install. To deploy this policy on the real hand, see the real-world deployment section below.

cd leapsim
python3 train.py wandb_activate=false num_envs=1 headless=false test=true task=LeapHandRot checkpoint=runs/pretrained/nn/LeapHand.pth

Real-world deployment

• Running in the real world requires our LEAP Hand ROS API.
• Follow the instructions in the above link and then run roslaunch example.launch first. The hand should go to the home pose.
• Next, in a separate window run deploy.py using:

cd leapsim
python3 deploy.py wandb_activate=false num_envs=1 headless=false test=true task=LeapHandRot checkpoint=runs/pretrained/nn/LeapHand.pth

• The hand should go to a pre-grasp pose and then rotate a 7.5cm cube by default.

Training your own policy

First, generate a cache of stable grasps for different cube sizes

for cube_scale in 0.9 0.95 1.0 1.05 1.1 
do
	bash scripts/gen_grasp.sh $cube_scale custom_grasp_cache num_envs=1024 
done

This will generate .npy files in the leapsim/cache folder. Next, train a policy using this grasp cache

python3 train.py task=LeapHandRot max_iterations=1000 task.env.grasp_cache_name=custom_grasp_cache

If you wish to not use wandb append wandb_activate=false

for cube_scale in 0.9 0.95 1.0 1.05 1.1 
do
	bash scripts/gen_grasp.sh $cube_scale custom_grasp_cache num_envs=1024 wandb_activate=false
done
python3 train.py task=LeapHandRot max_iterations=1000 task.env.grasp_cache_name=custom_grasp_cache wandb_activate=false

After training, the policy can be visualized by running

python3 train.py wandb_activate=false num_envs=1 headless=false test=true task=LeapHandRot checkpoint=runs/<checkpoint_name>/nn/LeapHand.pth

For training details of this policy refer to the LEAP hand paper section VI-D.

Citing

If you find LEAP hand or this codebase useful in your research, please cite:

@article{
	shaw2023leaphand,
	title={LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning},
	author={Shaw, Kenneth and Agarwal, Ananye and Pathak, Deepak},
	journal={Robotics: Science and Systems (RSS)},
	year={2023}
}

Acknowledgements

Check out the following amazing codebases we build upon

• Hora - https://github.com/HaozhiQi/hora/
• IsaacGymEnvs - https://github.com/NVIDIA-Omniverse/IsaacGymEnvs