Adaptive Bounded-Rationality Modeling of Early-Stage Takeover in Shared-Control Driving

Paper Title

Adaptive Bounded-Rationality Modeling of Early-Stage Takeover in Shared-Control Driving

Publication Info

• Topic area: Cognitive modeling and adaptive systems for shared-control driving safety.
• Keywords: Bounded rationality, shared-control driving, takeover safety, cognitive modeling, reinforcement learning, particle filtering, physiological alignment, human–automation interaction.

Background and Problem

• Problem / challenge: Current methods for predicting driver behavior during control handover struggle with data scarcity, non-stationary cognitive states, and lack of real-time adaptation, leading to unreliable safety interventions.
• Significance: Early-stage takeover prediction is critical for preventing accidents during semi-automated driving, where human drivers must quickly regain control under time pressure and uncertainty.
• Motivation and related work: Existing behavioral models (e.g., deep learning) achieve high accuracy with ample data but fail in transient post-handover scenarios. Cognitive models improve interpretability but lack direct coupling to executable control actions and real-time adaptability. This paper addresses these gaps by integrating cognitive mechanisms with adaptive prediction.

Solution

• Proposed approach: A cognition-to-control framework that models driver behavior as bounded rationality and adapts latent cognitive parameters in real time using particle filtering.
• Novelty:
  1. Interpretable bounded-rationality model coupling perceptual noise, looming aversion, and action delay to closed-loop control.
  2. Sequential inference method for real-time adaptation of cognitive parameters during takeover.
  3. Empirical validation in a vehicle-in-the-loop study showing improved prediction and physiological alignment.
  4. Demonstration of early collision warnings with extended lead times compared to baselines.
• Procedure and key techniques:
  • Cognitive mechanisms (perceptual uncertainty, looming aversion, action delay) embedded in a Partially Observable Markov Decision Process (POMDP).
  • Reinforcement learning (PPO) with cognition-aware policy modulation.
  • Online particle-filter inference of cognitive parameters from observed trajectories.
  • Evaluation using simulator-based driving scenarios and physiological data (e.g., eye-tracking).

Results

• Concrete findings:
  • Early collision warnings achieved in 89.5% of episodes at ≥ 0.5 s lead time, outperforming Vanilla PPO (41.5%) and Constant-Velocity (22.0%).
  • Lead-time coverage at ≥ 1 s and ≥ 2 s was 80.5% and 58.5%, respectively, compared to 12.0% and 0.0% for Vanilla PPO.
  • Physiological alignment: inferred perceptual noise parameters correlated with gaze entropy and fixation instability; looming-aversion coefficient tracked saccadic bursts and pupil dilation.
• Advantage over baselines:
  • Higher early-warning hit rates and longer lead-time coverage.
  • Improved predictive accuracy during transient takeover phases.
  • Physiological plausibility of inferred cognitive parameters.
• Experiments / evaluation:
  • Vehicle-in-the-loop study with 41 participants performing takeover tasks in a high-fidelity driving simulator.
  • Quantitative evaluation of prediction accuracy, early-warning performance, and physiological alignment.
  • Qualitative analysis of a representative collision episode, illustrating dynamic risk forecasting and cognitive–behavior coupling.
• Limitations and future work:
  • Limited scenario diversity (focused on freeway work-zone avoidance).
  • Sensitivity to particle-filter design choices.
  • Physiological miss rates suggest additional mechanisms may be needed.
  • Future plans include expanding scenarios, testing closed-loop assistance policies, and conducting on-road studies.

Summary

This paper introduces a cognition-to-control framework for predicting driver behavior during early-stage takeover in shared-control driving. By embedding bounded rationality mechanisms (perceptual uncertainty, looming aversion, action delay) into a POMDP and adapting cognitive parameters in real time, the model achieves higher early-warning coverage and longer lead times compared to baselines. Empirical validation demonstrates alignment between inferred cognitive states and physiological metrics, supporting the model’s interpretability and applicability for safety-critical interventions. Future work will expand scenario diversity, test real-world applications, and explore fairness across driver subgroups.