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

Technical Summary: An LLM-Explainable DRL Framework for Passenger-Directed Autonomous Driving

Problem Statement
Despite advancements in autonomous vehicle (AV) perception and control, public trust remains limited due to the "black box" nature of decision-making systems. Passengers often lack transparency regarding why an AV performs specific actions, particularly in critical situations. Existing Explainable AI (XAI) approaches are predominantly developer-oriented, relying on visual representations or technical rule traces that are not accessible to non-expert users. Furthermore, current frameworks rarely integrate real-time passenger voice commands directly into the control loop while simultaneously providing natural language justifications when safety constraints necessitate overriding those commands.

Methodology
The authors propose a hybrid framework operating within a SUMO microscopic traffic simulator that integrates Deep Reinforcement Learning (DRL) for adaptive control with Large Language Models (LLMs) for passenger interaction and explainability. The system architecture comprises four main subsystems:

1. Simulation Environment: A two-way urban road scenario featuring signalized intersections, unsignalized pedestrian crossings, and variable traffic flows. Passenger intent is abstracted into three discrete commands: fast, comfort, and stop.
2. Command Interpretation (LLM-1): Utilizing GPT-4.1-mini, this module transcribes spoken passenger requests (direct or indirect) and reformulates them into a confirmed, unambiguous driving instruction (drive faster, drive more carefully, or stop). It outputs a numeric class (0, 1, or 2) for the control agent.
3. Autonomous Driving Core (DRL): The control policy is modeled as a Markov Decision Process (MDP) using a Dueling Double Deep Q-Network (D3QN).
   • State Space: An 8-dimensional vector including passenger command, ego speed, speed limit, headway distance, relative speed, traffic light distance/state, and pedestrian crossing distance.
   • Action Space: Discrete longitudinal velocity adjustments ranging from -3 to +3 m/s.
   • Reward Functions: Three distinct reward structures are defined for the driving modes:
     • Fast: Prioritizes efficiency and speed limit adherence while maintaining safety (Time-to-Collision > 2s).
     • Comfort: Prioritizes smoothness by penalizing high jerk and favoring conservative headways.
     • Stop: Enforces immediate braking with high penalties for maintaining speed.
   • Safety Mechanism: A virtual leader approach is used to enforce stopping at red lights and pedestrian crossings, replacing rigid distance thresholds with a car-following logic that ensures rule compliance.
4. Explainability Pipeline (LLM-2 & LLM-3):
   • Decision Module (LLM-2): Evaluates the current vehicle action against the passenger request and environment state. It triggers an explanation only if a conflict exists (e.g., decelerating despite a "fast" request due to a red light).
   • Explanation Generator (LLM-3): Generates concise, reassuring, one-to-two-sentence justifications in natural language when a conflict is detected, focusing on the most safety-relevant factor (e.g., "approaching a traffic light").

Key Contributions

1. Passenger-Oriented XAI Pipeline: A novel integration of LLMs with DRL that translates natural language passenger requests into control policies and generates context-aware explanations for non-expert users.
2. Multi-Mode D3QN Agent: A trained agent capable of seamlessly switching between fast, comfort, and stop modes within a single trip, adhering to traffic rules and safety constraints.
3. Conflict-Detection Mechanism: An LLM-based module that dynamically determines when an explanation is necessary, specifically triggering when safety constraints override passenger intent, thereby balancing adaptability with transparency.

Results
The framework was evaluated over 11,000 training episodes and 25 test episodes across varying traffic densities (0.5x to 1.5x baseline flow).

• DRL Performance: Both fast and comfort agents demonstrated stable learning and collision-free operation.
  • The fast agent achieved an average speed of 4.508 m/s with an average Time Headway (TH) of 2.998s and Time-to-Collision (TTC) of 5.603s.
  • The comfort agent achieved a slightly lower average speed (4.396 m/s) but maintained a significantly larger TH (6.373s) and lower jerk (0.205 m/s³), confirming smoother driving behavior.
  • Both agents successfully executed emergency stop commands, overriding their respective driving modes without instability.
• LLM Performance:
  • Command Interpretation (LLM-1): Achieved 96.9% accuracy on a dataset of 33 requests, with perfect precision for the fast class. One error involved misclassifying a stop request as comfort.
  • Decision & Explanation (LLM-2/3): Qualitative evaluation showed the system correctly triggered explanations for safety-driven deviations (e.g., slowing for a red light) and generated appropriate, reassuring justifications. The system effectively avoided redundant explanations for minor, expected adjustments.

Significance and Claims
The paper presents this work as a proof of concept demonstrating the feasibility of integrating DRL-based driving control with LLM-driven passenger interaction. The authors claim that this framework successfully bridges the gap between AV performance and public trust by providing a system that is safe, adaptive to user preferences, and explainable.

The study highlights that the framework is most effective when it must explain why a passenger's request (e.g., "drive faster") is delayed or overridden due to safety constraints. The authors explicitly note limitations, including the reliance on simulation (SUMO), the lack of standardized metrics for real-time explanation evaluation, and the potential for LLM hallucinations. Consequently, they frame the work as a foundational step toward future extensions involving lateral control (lane changing) and real-world deployment, rather than a fully matured solution ready for immediate commercial use.