Dynamic Path Planning for Unmanned Autonomous Vehicles Based on CAS-UNet and Graph Neural Networks

TL;DR

This paper introduces a new model for path planning in dynamic environments using CAS-UNet and GNN, achieving high success rates and efficiency for multi-agent vehicles.

Contribution

The novel CAS-GNN model integrates CAS-UNet and GNN with dynamic edge modules for improved obstacle recognition and adaptability.

Findings

CAS-GNN achieved a 92.8% success rate in dynamic scenarios with six vehicles.

The model reduced the collision rate by 23% compared to the A-GNN model.

Trajectory efficiency improved by 8% with CAS-GNN over traditional methods.

Abstract

This paper proposes a deeply integrated model called CAS-GNN, aiming to solve the collaborative path-planning problem for multi-agent vehicles operating in dynamic environments. Our proposed model integrates CAS-UNet and Graph Neural Network (GNN), and, by introducing a dynamic edge enhancement module and a dynamic edge weight update module, it improves the accuracy of obstacle boundary recognition in complex scenarios and adaptively changes the influence of different edges during the information transmission process. We generate data through online trajectory optimization to enhance the model’s adaptability to dynamic environments. Simulation results show that our proposed CAS-GNN model has good performance in path planning. In a dynamic scenario involving six vehicles, our model achieved a success rate of 92.8%, a collision rate of 0.0836%, and a trajectory efficiency of 64%. Compared with the traditional A-GNN model, our proposed CAS-GNN model improves the planning success rate by 2.7% and the trajectory efficiency by 8%, while reducing the collision rate by 23%.

Figures