legged-robots-manipulation

Introduction

legged-robots-manipulation is a loco-manipulation repository for (wheel-)legged robots. The code is built on legged_gym. We provide two pre-trained models for airbot and b2w.

Project Page:wheel-legged-loco-manipulation

The current repository contains airbot and b2w. The current repository is a partial implementation of the paper. The whole code will be released once the paper is accepted.

airbot

airbot is a loco-manipulation task implemented using PPO. It is the baseline for comparison in the Paper.

b2w

b2w is a locomotion demo for the Unitree b2w robot.

Installation

install isaac_gym,rsl_rl,legged_gym

1. You are supposed to install the isaac_gym,rsl_rl and legged_gym.Please install the above dependencies (Isaac Gym Preview 4, rsl_rl, legged_gym) as described in legged_gym.

Copy required files to legged_gym.

1. Copy the contents of ~/legged-robots-manipulation/legged_gym/envs/init.py into legged_gym.

   # Add those contents

import os
from legged_gym.envs.airbot.airbot_config import AirbotRoughCfg, AirbotRoughCfgPPO

from legged_gym.envs.b2w.b2w_config import B2wRoughCfg, B2wRoughCfgPPO


from legged_gym.utils.task_registry import task_registry

from .airbot.airbot_robot import Airbot
from .b2w.b2w_robot import B2w


from .airbot.airbot_config import AirbotRoughCfg, AirbotRoughCfgPPO
from .b2w.b2w_config import B2wRoughCfg, B2wRoughCfgPPO

task_registry.register( "b2w", B2w, B2wRoughCfg(), B2wRoughCfgPPO() )
task_registry.register( "airbot", Airbot, AirbotRoughCfg(), AirbotRoughCfgPPO() )

2. Copy envs files into legged_gym/envs corresponding.

   # Add those files
cp path/to/legged-robots-manipulation/legged_gym/envs path/to/legged_gym/legged_gym/envs

Train

python tran.py --task=b2w 

# or

python train.py --task=airbot  

Citing

If you use this work, please cite this paper:

      @misc{wang2024armconstrained,
        title={Arm-Constrained Curriculum Learning for Loco-Manipulation of the Wheel-Legged Robot}, 
        author={Zifan Wang, Yufei Jia, Lu Shi, Haoyu Wang, Haizhou Zhao, Xueyang Li, Jinni Zhou, Jun Ma and Guyue Zhou},
        year={2024},
        eprint={2403.16535},
        archivePrefix={arXiv},
        primaryClass={cs.RO}
  }