Constraint-Guided Multi-Task Reinforcement Learning for Autonomous Electric Taxi Dispatch and Energy Management

With the rapid development of autonomous driving and electric vehicle technologies, autonomous electric taxis (AETs) are emerging as a critical pathway toward intelligent and low-carbon urban mobility. However, existing studies often treat order dispatching and energy management as separate tasks, neglecting the dynamic interactions and collaborative optimization between these two subsystems. Additionally, they typically lack explicit modeling of battery constraints and behavioral restrictions, resulting in compromised operational safety and poor policy convergence. To address these issues, this paper proposes a Constraint-guided Multi-task Value Decomposition Network (CM-VDN), a multi-agent reinforcement learning framework designed to achieve coordinated optimization of order dispatching and energy management in AETs. We introduce a constraint-aware multi-objective reward mechanism, explicitly embedding critical operational constraints such as battery boundaries, task mutual exclusions, and penalties for idle actions. Experimental results demonstrate that the proposed CM-VDN significantly outperforms conventional reinforcement learning methods, achieving over 200% improvement in overall operational revenue while effectively eliminating constraint violations. The study provides solid theoretical and practical foundations for deploying autonomous electric taxi fleets in complex, dynamic urban environments.