Safe and Scalable Multi-Agent Optimization for Autonomous Electric Taxi Dispatching via a Two-Stage Reinforcement Learning Framework

The widespread deployment of Autonomous Electric Taxis (AETs) in smart cities introduces critical challenges in large-scale dispatching and charging coordination under energy and operational constraints. Traditional Multi-Agent Reinforcement Learning (MARL) approaches often struggle to ensure both policy feasibility and system scalability in such complex, dynamic environments. In this paper, we propose a safe and scalable two-stage MARL framework for AET dispatching optimization. The proposed method, named Filter-to-Optimization Pipeline (FTOP), decouples constraint handling from reward maximization through a hierarchical architecture. In the first stage, an Action Classification-based Action Filter (ACAF) employs value decomposition to eliminate infeasible actions, enforcing energy and conflict constraints. In the second stage, a Utility-Prioritized Maximization Policy (UPMP) performs a model-based search within the filtered feasible space to optimize system-level utility. Extensive simulations demonstrate that FTOP achieves significant improvements in total fleet revenue, constraint satisfaction, and scalability across varying urban scenarios. These results highlight the potential of decomposed MARL strategies in solving large-scale, safety-critical optimization problems in autonomous transportation systems.