An LLM-Explainable DRL Framework for Passenger-Directed Autonomous Driving

Researchers developed a framework combining deep reinforcement learning (DRL) with large language models (LLMs) to make autonomous vehicles safer and more trustworthy by explaining driving decisions to passengers. The system was trained to handle three driving modes—fast, comfort, and stop—while generating safety-focused explanations for its actions, demonstrating effective mode switching and rule compliance.

This research addresses a critical bottleneck in autonomous vehicle adoption: the transparency gap between sophisticated AI decision-making and public trust. By pairing DRL agents trained via Dueling Double Deep Q-Networks with LLM-based explanation modules, the work tackles both technical safety and the psychological dimension of passenger confidence. The framework shows that vehicles can dynamically switch between driving preferences while justifying safety-critical overrides in natural language.

Autonomous vehicles have stalled in public perception despite decades of technical progress, largely because passengers cannot understand why the system makes specific choices. This research fits an emerging trend of explainable AI (XAI) moving from theoretical interest to practical implementation. The dual-layer approach—combining RL's adaptability with LLMs' communication ability—represents a pragmatic engineering solution rather than speculative research.

For the autonomous vehicle industry, this work demonstrates that explainability can be integrated without sacrificing safety or performance. Manufacturers and regulators increasingly recognize that opaque decision-making creates liability and regulatory friction. The framework's effectiveness at handling safety-constraint conflicts (delayed or overridden requests) suggests real-world applicability where passenger preferences must yield to safety requirements.

Developers should watch whether this architecture scales to complex urban environments and diverse passenger preferences. The next critical phase involves testing passenger comprehension and trust improvements from LLM explanations, and whether explanations reduce passenger anxiety during safety-constrained driving. Regulatory bodies may also adopt similar explainability standards as baseline requirements for autonomous vehicle certification.

• →DRL agents successfully learned three distinct driving modes with stable training and safe traffic rule compliance.
• →LLM-generated explanations proved most valuable when safety constraints overrode passenger requests, enhancing transparency during conflicts.
• →The framework balances adaptability, safety, and explainability—three traditionally competing objectives in autonomous driving systems.
• →Integration of reinforcement learning and large language models offers a scalable approach to passenger-vehicle trust building.
• →Results support adopting explainability as a standard component in autonomous vehicle architectures, not an optional feature.