Multi-robot Path Planning Based on Spatio-Temporal Information in Large-scale Unknown Environment

Abstract

Multi-agent path planning is an important re-search direction in robotics due to its wide range of applications such as autonomous driving, automatic storage and so on. Off-the-shelf methods usually only consider the spatial information of the environment where the robot is located. Here we argue that the robot should use both spatial and temporal information to complete action decisions. To this end, We propose a novel multi-robot path planning model mainly composed of three modules: a Convolutional Neural Network (CNN) module, a Graph Neural Network (GNN) module and a Long short-term memory (LSTM) module. Specifically, CNN is responsible for extracting the local spatial features of each robot, and the GNN enables multiple robots to communicate their local spatial features with the neighbours. The LSTM completes the task of fusing temporal information. The combination of temporal and spatial information allows the robot to have a memory function when making behavioral decisions. We validated our approach in some complex environments, and the simulation results show that our approach has a better path planning performance than state of the arts.