Research on multirobot SLAM path planning based on deep learning

Abstract

Multi-robot cooperative SLAM is a core technology for autonomous exploration in large-scale unknown environments. However, the information isolation between traditional SLAM and path planning modules leads to poor real-time performance and low collaborative efficiency. This paper proposes an end-to-end multi-robot SLAM path planning framework based on deep learning, which deeply integrates map feature extraction, inter-robot interaction modeling, and path planning decision-making. The framework comprises three core modules: a multi-modal fusion perception module based on CNN-Transformer, which extracts environmental features from RGB-D and LiDAR data; an inter-robot interaction module based on Graph Neural Networks (GNN), which models collaborative relationships and obstacle a voidance constraints among robots through a dynamic graph structure; and a distributed path planning decision-making module based on Deep Reinforcement Learning (DRL), which adopts a centralized training and distributed execution framework to achieve multi-robot cooperative motion planning. Experiments were conducted in a comprehensive complex scene built in the Gazebo simulation environment, and the results showed that the proposed framework achieved a success rate of 88.64%, an improvement of 30.4% compared to the baseline method based on RRT*, and demonstrated good scalability