Presentation
Learning Personalized Human Driving Strategies for Negotiation with Automated Vehicles via Inverse Reinforcement Learning and a Hybrid Markov Decision Process
DescriptionAutomated vehicles (AVs) must safely negotiate with human-driven vehicles in mixed traffic, especially in right-of-way conflict scenarios where human behavior is dynamic and uncertain. This study proposes a data-driven framework for learning and predicting personalized human driving strategies during human–AV negotiation. Using a high-fidelity simulator dataset with 54 participants, 648 trials, and six competitive driving scenarios, we first apply maximum-entropy inverse reinforcement learning to recover scenario-specific reward functions reflecting human tradeoffs among progress, safety, and smoothness. We then use these learned rewards to train a lightweight hybrid Markov decision process policy that combines behavior cloning with reward-based fine-tuning to reproduce representative interaction trajectories. An additional outcome-prediction head estimates whether the ego vehicle will obtain priority from partial trajectory observations. Results show that the learned rewards distinguish expert demonstrations from perturbed rollouts, hybrid training improves trajectory prediction, and reliable outcome prediction can be achieved within several seconds. Demographic-based threshold calibration further improves early prediction robustness without retraining the policy.
Event Type
Lecture
TimeThursday, October 22nd3:20pm - 3:40pm PDT
Location
Similar Presentations