rl_rvo_nav

Introduction

This repository is the source code of the paper "Reinforcement Learned Distributed Multi-Robot Navigation with Reciprocal Velocity Obstacle Shaped Rewards" [RA-Letter]

PDF_ieee | PDF_arxiv | Video_Youtube | Video_Bilibili

Circle 10	Circle 16	Circle 20

Random 10	Random 16	Circle 20

Prerequisites

• Python >= 3.8
• Pytorch >= 1.6.0
• intelligent-robot-simulator == v2.5

git clone -b v2.5 https://github.com/hanruihua/intelligent-robot-simulator.git
cd intelligent-robot-simulator
pip install -e .

Test environment

• Ubuntu 20.04, 18.04
• Windows 10, 11

Installation

git clone https://github.com/hanruihua/rl_rvo_nav.git
cd rl_rvo_nav
./setup.sh

Policy Train

• First stage: circle scenario with 4 robots.

python train_process.py --use_gpu

or

python train_process_s1.py

• Second stage: continue to train in circle scenario with 10 robots.

python train_process.py --robot_number 10 --train_epoch 2000 --load_name YOUR_MODEL_PATH --use_gpu --con_train

or

python train_process_s2.py

Policy Test

You can test the policy trained from the previous steps by following command:

python policy_test.py --robot_number 10 --dis_mode 3 --model_name YOUR_MODEL_NAME --render

Note1: dis_mode, 3, circle scenario; 2 random scenario
Note2: YOUR_MODEL_NAME refer to the path and name of the check point file in the policy_train/model_save folder

Pretrained model

We provide the pre_trained model, you can test this model by following command:

python policy_test_pre_train.py --render

Citation

If you find this code or paper is helpful, you can star this repository and cite our paper:

@article{han2022reinforcement,
  title={Reinforcement Learned Distributed Multi-Robot Navigation With Reciprocal Velocity Obstacle Shaped Rewards},
  author={Han, Ruihua and Chen, Shengduo and Wang, Shuaijun and Zhang, Zeqing and Gao, Rui and Hao, Qi and Pan, Jia},
  journal={IEEE Robotics and Automation Letters},
  volume={7},
  number={3},
  pages={5896--5903},
  year={2022},
  publisher={IEEE}
}

Author

Han Ruihua
Contact: [email protected]