A hybrid deep reinforcement learning approach for target allocation and routing of multiple nonholonomic vehicles

Abstract

The coordination of multiple nonholonomic vehicles (NVs) for large-scale applications poses a complex dual decision-making problem of target allocation and routing while considering kinematic constraint of NVs. To address this problem, this paper proposes a Hybrid Deep Reinforcement learning approach for Target allocation and Routing of multiple NVs (HyDR-TR). HyDR-TR employs a Transformer-based architecture to solve for target allocation and routing jointly. Experiments on traveling salesman problem library benchmarks (10-130 targets) and randomly generated datasets (25-100 targets) with a fixed number of homogeneous NVs demonstrate that HyDR-TR achieves 10%-30% lower makespan and computes 30-400 times faster compared with recent state-of-the-art methods.