百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
L

LEAP_Hand_Sim

> 编程语言
开源

用于 LEAP 手部 V1 的 Isaac Gym 仓库

200 stars0 点赞1 次浏览
访问官网GitHub

工具介绍

用于 LEAP 手部 V1 的 Isaac Gym 仓库

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

bash
cd isaacgym/python
pip install -e .

Clone and install leapsim python packages

bash
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.

bash
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:
bash
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//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

Issues· 1 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

Python

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月18日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言