Detour: Decentralized Multi-Robot Off-Road Navigation via Terrain Diffusion and Curriculum Reinforcement Learning

Haonan Yang, Bolei Chen, Jiaxu Kang, Ping Zhong, Yu Sheng

1. Overview

The multi-robot navigation system has increasingly shown superior performance and broad application prospects in scenarios such as logistics and rescue. Despite the promising performance achieved by existing work in indoor flat terrains, many challenges still need to be overcome in outdoor off-road settings, such as sensing occlusions caused by complex terrains, dangerous tilts, and dynamic collision avoidance. In this paper, a Decentralized multi-robot off-road navigation strategy named Detour is proposed to tackle the above issues by using terrain diffusion and Curriculum Reinforcement Learning (CRL). In particular, the terrain diffusion technique is designed to imagine possible future terrains based on current visual observations, which not only mitigates the occlusion of sensor data by undulating terrains but also enhances the robot's reactivity to potential collisions. By designing an easy-to-hard curriculum for navigation strategy learning, the robot's navigation ability is steadily enhanced to cope with dynamic scenes with complex terrains. In addition, by fully considering the robot's tilt and collision risks, a reward function is crafted for CRL to address the reward sparsity difficulty. Sufficient comparative and ablation studies demonstrate our Detour's superiority. Surprisingly, we experimentally find that terrain diffusion can drastically reduce navigation time steps while improving navigation success and reducing collisions. Experiments with a real robot in outdoor dynamic scenes further validate Detour's feasibility. The experimental code is available at Detour.

This repository contains code for Detour.

2. A demo video for Detour

This is a demo video of the real robot experiments for the paper titled "Detour: Decentralized Multi-Robot Off-Road Navigation via Terrain Diffusion and Curriculum Reinforcement Learning".

3. Installation

Clone

git clone https://github.com/Southyang/Detour.git

Configure the conda environment

conda create -n Detour python=3.8
conda activate Detour

pytorch 1.10.0

Compile

cd Detour
catkin_make

4. Run

Sumilation enviromnent

cd Detour
source ./devel/setup.bash
roslaunch detour train_stage1.launch

Navigation script

cd Detour/src/detour/scripts
mpiexec -n 6 python train_stage1.py

Experimental diagram

5. Acknowledgement

Code acknowledgements:

• Detour/src/Husky is modified from Husky.