Data-Driven Routing for Autonomous Trucks: Learning from Human Behavior with Context Awareness and Privacy Protection

Abstract

Efficient and resilient freight transportation is a cornerstone of modern economies. In Canada, where long distances, dispersed populations, and reliance on cross-border trade amplify the importance of trucking, heavy-duty vehicles carry most domestic and international shipments. However, trucking operations increasingly face challenges related to congestion, infrastructure limitations, and disruptive extreme weather events. These factors directly affect delivery reliability, operating costs, and emissions. As the industry moves toward automation, the need for adaptive and data-driven route recommendation for autonomous truck fleets has become urgent. Unlike human drivers who draw on experiential knowledge to navigate disruptions, autonomous systems must infer driver expertise, predict traffic under adverse conditions, and coordinate decisions across fleets, while respecting privacy constraints. This thesis focuses specifically on these data-driven and behavior-learning challenges, not on hardware or sensor-level issues. This work presents a unified framework for smart route planning for autonomous heavy-duty trucks and pursues three technical objectives: (i) accurate map matching under sparse GPS sampling, (ii) truck-specific traffic prediction under extreme weather, and (iii) privacy-preserving, context-aware route planning informed by human driving behavior. The first objective introduces the Multi-Intention Deep Inverse Reinforcement Learning (MIDIRL) framework for reconstructing truck trajectories from low-frequency GPS data. MIDIRL models diverse driver preferences by learning reward functions from high-frequency trajectories and identifying intention groups using expectation–maximization clustering. A Q-learning search then selects the most plausible driving path. Compared with classical Hidden Markov Model (HMM) and deep-learning baselines, MIDIRL achieves substantially higher segment-level and length-based accuracy on Calgary and Edmonton datasets. The second objective develops the Multi-Task Context-Based GRU Graph Convolutional Network (MT-C2G) for predicting truck traffic under extreme weather. MT-C2G integrates Graph Convolutional Networks (GCNs) for spatial structure, Gated Recurrent Units (GRUs) for temporal dynamics, and attention mechanisms for environmental context. To address the scarcity of rare weather events, Synthetic Minority Oversampling (SMOTE) is applied during training. Experiments on Alberta data show that MT-C2G reduces prediction errors compared with state-of-the-art spatio-temporal baselines, improving robustness required for proactive routing. The third objective proposes MetaFAIRL-Routing, a two-stage framework for learning human-like routing behavior while preserving data privacy. First, Meta-Federated Adversarial Inverse Reinforcement Learning enables multiple fleet operators to jointly learn a shared reward model without exchanging raw trajectory data. Second, a Multi-Agent Deep Q-Network (MA-DQN) uses this learned reward to coordinate decentralized routing decisions across trucks under dynamic conditions. Evaluations demonstrate 5–12% reductions in travel time and 15–35% reductions in congestion exposure, achieving near-centralized performance despite decentralized and privacy-restricted data. Collectively, this thesis advances behavior-aware and climate-resilient routing for autonomous truck fleets by addressing sparse trajectory data, extreme weather prediction, and privacy-sensitive fleet coordination. The proposed methods integrate human expertise, environmental awareness, and adaptive multi-agent decision-making, delivering practical benefits for logistics operators and offering policymakers tools for designing climate-resilient freight networks.