Small Talk, Big Impact? LLM-based Conversational Agents to Mitigate Passive Fatigue in Conditional Automated Driving

Paper Title

Small Talk, Big Impact? LLM-based Conversational Agents to Mitigate Passive Fatigue in Conditional Automated Driving

Publication Info

• Topic area: Human–Machine Interaction in automated driving contexts.
• Keywords: Conversational agents, passive fatigue, automated vehicles, SAE Level 3, human–AI teaming, vigilance, safety-critical systems, Karolinska Sleepiness Scale, user archetypes, adaptive design.

Background and Problem

• Problem / challenge: Passive fatigue during monotonous SAE Level 3 automated driving undermines driver readiness and situational awareness, posing safety risks. Existing countermeasures lack adaptability to individual driver needs and often rely on intrusive or simulator-based approaches.
• Significance: Addressing passive fatigue is critical for ensuring safe transitions between automated and manual driving, particularly in conditional automation scenarios where vigilance is essential.
• Motivation and related work: Prior studies have explored gamification, non-driving-related tasks, and Wizard-of-Oz setups to mitigate fatigue, but these approaches often fail to adapt to diverse user profiles or real-world contexts. The potential of real-time, LLM-based conversational agents remains largely unexplored in authentic driving environments.

Solution

• Proposed approach: A real-time conversational agent (CA) named Zoe, powered by GPT-4, designed to combat passive fatigue through brief, context-aware dialogues during monotonous automated driving.
• Novelty:
  1. Deployment of a fully functional LLM-based CA in a genuine SAE Level 3 automated vehicle prototype during real-world driving.
  2. Identification of user archetypes (Safety-First, Entertainment-Seeking, Social-Connection Oriented) and their alignment with established CA design frameworks.
  3. Integration of environment-related prompts to enhance driver alertness while minimizing cognitive overload.
• Procedure and key techniques:
  • Zoe engaged drivers with short, safety-oriented dialogues tied to the driving environment (e.g., wildlife observations).
  • Interaction protocols included recovery strategies and safety reminders to prioritize road attention.
  • A test-track study with 40 participants assessed the CA’s impact on alertness using the Karolinska Sleepiness Scale (KSS), in-car video analysis, and post-drive interviews.

Results

• Concrete findings:
  • KSS scores showed significant improvement in alertness post-interaction with the CA (p = .028), reversing the trend of increasing fatigue observed in the control group.
  • Observed behaviors included reduced signs of drowsiness and boredom, increased scanning of surroundings, and positive emotional expressions during CA interaction.
• Advantage over baselines:
  • CA interaction mitigated passive fatigue more effectively than the control condition, where fatigue symptoms persisted.
  • Participants described the CA as less intrusive than traditional safety alerts and more engaging than non-driving-related tasks.
• Experiments / evaluation:
  • A between-subjects study with 40 participants (25 CA, 15 control) conducted on a closed test track using a Renault Zoe SAE Level 3 prototype.
  • Metrics included KSS ratings, thematic analysis of interviews, and behavioral observations from in-car video recordings.
• Limitations and future work:
  • Single-session study limits insights into long-term adoption and habituation.
  • Prototype AV and CA experienced occasional latency issues.
  • Future research should explore longitudinal effects, cross-cultural applicability, and adaptive systems that dynamically adjust to user preferences.

Summary

This study demonstrates the effectiveness of LLM-based conversational agents in mitigating passive fatigue during SAE Level 3 automated driving. The agent, Zoe, successfully improved driver alertness through brief, context-aware dialogues, as evidenced by KSS ratings and behavioral observations. User archetypes (Safety-First, Entertainment-Seeking, Social-Connection Oriented) were identified, highlighting the need for adaptive CA designs tailored to diverse driver needs. These findings advance the understanding of conversational interfaces as proactive safety tools in automated vehicles, offering design heuristics for future development and broader applications in safety-critical domains.