UAGMC: Urban Air-Ground Mobility Coordination

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

📌 Overview

UAGMC proposes an intelligent decision-making framework that leverages multimodal observational data and reinforcement learning to optimize vertiport selection in on-demand air-taxi services. The framework addresses complex, dynamic, and heterogeneous mobility environments by jointly considering passenger demands, ground traffic conditions, vertiport states, and aerial vehicle capabilities.

🚀 Key Features

• Multi-Source State Representation: Integrates OD demands, ground speed, vertiport status, and eVTOL capabilities.
• Temporal Feature Extraction: Employs LSTM to capture temporal dependencies in heterogeneous state sequences.
• PPO-based Policy Optimization: Uses Proximal Policy Optimization for robust and stable learning.
• Incremental Reward Design: Tackles sparse reward problem with time-step-wise reward shaping for better training efficiency.

🧠 Framework Architecture

The framework consists of three major modules:

1. State Encoding Block: Encodes heterogeneous observational data into latent features.
2. Feature Extractor Block: Extracts temporal dependencies from sequential encoded states.
3. Policy Network: Generates actions (vertiport assignments) using PPO.

🚀 Usage

🏋️‍♂️ Training Strategies

This project provides three training strategies, each using different state representations. Below are the descriptions and usage instructions for each training script:

Command to run:

python rl.py