Real-Time AVPC Scheduling with Collaborative Vehicle-Infrastructure Perception and Adaptive RL

The growing adoption of Autonomous Electric Vehicles (AEVs) and EV-charging infrastructure necessitates efficient real-time Automated Valet Parking and Charging (AVPC) scheduling frameworks that can accommodate both parking and charging demands. Accurate parking resource scheduling relies on precise resource occupancy detection and resource type classification via environmental perception. However, most existing perception models for the parking lot environment are limited by their sole reliance on static infrastructure-side data and lack of optimal parking resource scheduling solutions for the dynamic AVPC scenario. This paper proposes a novel adaptive Reinforcement Learning (RL) framework for AVPC resource scheduling by leveraging collaborative vehicle-infrastructure perception to enhance situational awareness and scheduling decision-making accuracy. We introduce a multi-sensor fusion model that integrates perception data from AEV-mounted sensors (LiDAR, radar, and cameras) with infrastructure-installed sensors (overhead cameras, Wi-Fi, and Bluetooth beacons). The fused perception output is then used by an adaptive Proximal Policy Optimization (PPO) agent with a fusion-aware reward shaping mechanism to guide the global scheduling of AEVs to parking and charging resources. The goal is to jointly optimize parking resource proximity, suitability, and energy cost while improving scheduling success rate. Simulation results using real-world parking lot data demonstrate that the proposed approach significantly outperforms existing baselines, improving convergence and scheduling success rate by at least 18.92% and 10.61%, respectively.