Heterogeneous Vertiport Selection Optimization for On-Demand Air Taxi Services: A Deep Reinforcement Learning Approach

Urban Air Mobility (UAM) has emerged as a transformative solution to alleviate urban congestion by utilizing low-altitude airspace, thereby reducing pressure on ground transportation networks. To enable truly efficient and seamless door-to-door travel experiences, UAM requires close integration with existing ground transportation infrastructure. However, current research on optimal integrated routing strategies for passengers in air-ground mobility systems remains limited, with a lack of systematic exploration.To address this gap, we first propose a unified optimization model that integrates strategy selection for both air and ground transportation. This model captures the dynamic characteristics of multimodal transport networks and incorporates real-time traffic conditions alongside passenger decision-making behavior. Building on this model, we propose a Unified Air-Ground Mobility Coordination (UAGMC) framework, which leverages deep reinforcement learning (RL) and Vehicle-to-Everything (V2X) communication to optimize vertiport selection and dynamically plan air taxi routes. Experimental results demonstrate that UAGMC achieves a 34% reduction in average travel time compared to conventional proportional allocation methods, enhancing overall travel efficiency and providing novel insights into the integration and optimization of multimodal transportation systems. This work lays a solid foundation for advancing intelligent urban mobility solutions through the coordination of air and ground transportation modes. The related code can be found at https://github.com/Traffic-Alpha/UAGMC.

💡 Research Summary

Urbanization has intensified traffic congestion, environmental degradation, and over‑population in metropolitan areas, prompting the search for novel mobility solutions. While traditional ground‑based traffic optimization focuses on two‑dimensional networks, Urban Air Mobility (UAM) leverages under‑utilized low‑altitude airspace to alleviate pressure on roads. Electric vertical take‑off and landing (eVTOL) aircraft, operating as on‑demand air taxis, promise faster travel and lower emissions, yet their limited passenger capacity and reliance on fixed vertiports introduce new bottlenecks: uneven vertiport utilization and prolonged waiting times.

This paper addresses these challenges by formulating a unified door‑to‑door travel model that simultaneously considers ground autonomous vehicles (CAVs) and aerial eVTOLs. The authors cast the problem as a Markov Decision Process (MDP) where the state captures real‑time V2X data—including road traffic flow, current queue lengths at each vertiport, eVTOL locations, battery levels, and demand forecasts. Actions consist of (i) assigning a CAV route for the first‑mile ground leg, (ii) selecting departure and arrival vertiports, and (iii) planning the aerial flight path and schedule. The reward function penalizes total travel time (ground travel + waiting + flight) while encouraging balanced vertiport usage.

To solve the high‑dimensional, heterogeneous decision problem, the authors develop the Unified Air‑Ground Mobility Coordination (UAGMC) framework based on an Actor‑Critic architecture with a PPO‑style policy update. Two novel neural modules are introduced for state representation:

1. Multi‑Source Contextual Embedding (MSCE) – integrates disparate data streams (traffic sensors, weather forecasts, demand predictions) into a unified latent vector, learning modality‑specific attention weights.
2. Spatio‑Temporal Integration Network (STIN) – combines graph convolutional layers (capturing geographic relationships among vertiports) with transformer‑based self‑attention to model temporal dependencies across the decision horizon.

Both modules feed into the policy and value networks, enabling efficient learning despite the sparse‑reward nature of the problem. The authors also employ a step‑wise shaping reward that provides intermediate feedback when a chosen vertiport reduces expected waiting time, accelerating convergence.

Experiments use a hybrid dataset: real traffic measurements from Shenzhen, synthetic eVTOL demand patterns, and a realistic vertiport layout (8–20 stations). Baselines include a proportional allocation rule, classic queuing‑theoretic dispatch, and existing deep‑RL approaches (DQN, vanilla Actor‑Critic). UAGMC achieves a 34 % reduction in average total travel time relative to the proportional allocation baseline, with especially pronounced gains during peak demand periods where vertiport waiting times drop by over 45 %. Sensitivity analyses reveal robustness to variations in vehicle capacity (2–5 passengers), vertiport count, and V2X communication latency up to 200 ms (performance loss <5 %). Ablation studies show that removing MSCE or STIN degrades performance by 12 % and 18 % respectively, confirming their critical role.

The paper’s contributions are threefold: (1) a unified MDP model that jointly optimizes ground and aerial legs of a multimodal trip, (2) the design of MSCE and STIN to handle heterogeneous, high‑dimensional inputs in a reinforcement‑learning setting, and (3) a comprehensive empirical validation demonstrating substantial efficiency gains over state‑of‑the‑art methods. By releasing the source code (https://github.com/Traffic‑Alpha/UAGMC) and detailed experimental settings, the authors provide a reproducible platform for future research on integrated air‑ground mobility, paving the way toward scalable, intelligent urban transportation ecosystems.