Work

Explain-Then-Act: A Novel LLM-Critic Framework for Safer Reinforcement Learning in Autonomous Driving

Year: 2026
Type: dissertation
Abstract: Autonomous driving systems increasingly rely on deep reinforcement learning for decision making in dynamic traffic environments, yet these models operate as black boxes and lack mechanisms for transparent, safety-oriented reasoning. This thesis proposes an Explain-Then-Act framework that integrates a large language model as a safety critic within a reinforcement learning based driving policy to enable pre-action evaluation and structured self-critique. Before execution, each action generated by a Proximal Policy Optimization agent is assessed using contextual traffic information, including ego state, surrounding vehicles, headway distance, and time-to-collision metrics. The language model classifies actions as safe, risky, or unsafe and provides a natural language explanation, triggering re-planning when necessary. A hybrid reward function incorporates both environmental feedback and safety critique signals to guide policy learning. Evaluation in a simulated highway environment demonstrates improved safety performance and more reliable behavior while maintaining competitive efficiency. The proposed framework advances explainable reinforcement learning by embedding interpretability and safety assessment directly into the decision-making process, contributing to more transparent and accountable autonomous driving systems.
Language: English
Open Access status: green
OpenAlex ID: W7141773084
Cites: 44
Cited by: 1
Related to: 10