Optimizing the Dynamic Drone-Assisted Pickup and Delivery Problem with Deep Reinforcement Learning

Timothy Mulumba

Keywords: Vehicle routing, reinforcement learning, deep neural networks, discrete optimization

Abstract:

We investigate the dynamic drone-assisted pickup and delivery problem (DAPDP), which concerns real-time, on-demand routing decisions in scenarios where new orders arrive stochastically throughout the day. By leveraging a fleet of trucks each equipped with a drone, operators can split tasks between ground vehicles and aerial vehicles, aiming to minimize total travel costs while respecting constraints on time windows, capacity, and drone flight endurance. We propose a deep reinforcement learning (DRL) approach based on Q-learning, augmented by a neural network function approximator, to decide dynamically which newly arrived orders to dispatch and how to integrate drone sorties effectively. Our experiments on a large, real-world-inspired dataset demonstrate substantial performance gains over greedy, random, and lazy dispatch baselines, yielding 10.6%, 22.6%, and 37.2% savings, respectively, in total travel cost.
Additionally, comparison against a clairvoyant oracle solution shows that our approach is near optimal.

Submission Number: 29