Intelligent Decision-Making for Autonomous Driving in Dynamic and Interactive Environments

The decision-making module is crucial for safe and efficient driving in autonomous vehicles (AVs). However, AVs face significant challenges in coexisting with human road users, making fast and optimal driving decisions, and operating in unknown traffic environments with only incomplete information. Unsignalized intersection and lane-changing scenarios are particularly representative of such challenges, which involves complex dynamic interactions. AVs need to assess and predict the driving preferences of nearby vehicles to optimize and adaptively adjust their own driving policies while considering uncertainties arising from incomplete observations. This dissertation investigates game-theoretic and learning-based methods to address these challenges.

In this dissertation, a novel approach for integrating game-theoretic decision-making with deep reinforcement learning (DRL) is proposed to enable AVs to navigate unsignalized intersections using an onboard sensor. The game model predicts the surrounding vehicles' reactions to the ego-vehicle's movements without relying on coordination or vehicle-to-vehicle communication. The proposed algorithm employs cognitive hierarchy reasoning and a DRL algorithm to achieve a self-play training mode for getting a near-optimal driving policy in a realistic simulator before transferring to the real world.

Second, this dissertation introduces a practical algorithm based on DRL for enhancing lane-changing decision-making, addressing low sample efficiency in DRL, and improving the generalization capability in Imitation Learning (IL). To narrow the gap between simulation and reality (sim-to-real gap), a digital twin platform is developed for simulating LiDAR sensing, model training, and algorithm evaluations. To tackle multi-objective optimization and imbalanced data concerns, a hierarchical decision-making framework is proposed, breaking down the complex decision-making problem into subtasks for improved convergence of driving policies.

Third, a robust adaptive game-theoretic decision-making algorithm by utilizing receding horizon optimization and level-k game theory is presented. To reduce the potential safety risk arising from an inaccurate motion prediction of surrounding vehicles, the proposed approach can estimate driving aggressiveness of surrounding vehicles online. Then, the generated trustworthiness is used to formulate a safe, efficient, and robust adaptive driving policy. Additionally, a switching interaction graph is introduced into the adaptive level-k framework to reduce the computational complexity.

Validation on both a high-fidelity simulator and hardware confirms the feasibility, effectiveness, and real-time performance of the proposed methods. Overall, this dissertation contributes novel approaches to address decision-making challenges in AVs. The integration of game theory and model-based/model-free optimization showcases the potential for improving safety and efficiency of AVs' operation in dynamic traffic environments.