Abstract
This study presents a data-driven framework for quantifying within-driver variability during complex driving tasks. By leveraging Kernel Distribution Embedding and the Maximum Mean Discrepancy (MMD) metric, we characterize individual learning profiles in a simulated obstacle-avoidance scenario. Our results demonstrate that driving trajectories can be mapped into a high-dimensional functional space to identify patterns of convergence, consistency, or disengagement, providing a foundation for more personalized human-centric autonomous systems.
Introduction
Autonomous and driver-assistance systems (ADAS) typically target the "average" human driver. However, human behavior is inherently inconsistent, exhibiting significant variability not only between individuals but also within a single driver across repeated trials. Traditional control-theoretic models often fail to capture this "naturalistic" decision-making because they rely on rigid assumptions like the Markov property or Gaussian noise. To develop truly human-centric autonomy, systems must distinguish between a driver becoming more proficient, remaining erratic, or losing focus.
Methodology
We propose a nonparametric machine learning approach to quantify driving variability:
- Kernel Distribution Embedding: Driving trajectories (comprising position, velocity, and heading) are treated as realizations of a stochastic process and projected into a Reproducing Kernel Hilbert Space (RKHS). This allows a complex trajectory to be represented as a single "embedding" point.
- Maximum Mean Discrepancy (MMD): We utilize MMD as a distance metric to calculate the separation between probability distributions within the RKHS.
- Convergence Benchmarking: For each participant, a "final distribution" is defined using the last four trajectories of the experiment. The MMD is then calculated between each initial trajectory and this benchmark to track convergence toward a consistent driving style.
Experimental Setup
A pilot study was conducted with six senior engineering students using an advanced driving simulator. Participants drove at 45 mph on a nighttime rural highway and encountered a hidden obstacle (a tire) behind a blind S-curve. To induce a learning curve and increase task difficulty, participants were required to steer using only their non-dominant hand. Data collected included 12 trajectories per participant and eye-tracking fixation durations.
Results and Discussion
The MMD metric effectively visualized "learning" as a decrease in distance to the final benchmark over time.
The study identified three distinct behavioral profiles:
- Learners (Participants 1, 3, 4): Showed a steady decrease in MMD, indicating successful adaptation and increased consistency.
- Consistent/Ceiling Group: Exhibited flat MMD curves, suggesting they either reached peak performance immediately or lacked motivation for further improvement.
- Disengaged (Participant 2): Showed a "U-shaped" curve. Post-experiment observation revealed this participant switched to two-handed driving against instructions before losing focus.
The Eye-Tracking Paradox
Correlating MMD with eye-tracking data revealed divergent strategies. Some "Learners" reduced road fixation as they became more efficient, while others increased fixation to maintain focus. This suggests that ADAS monitoring eye-tracking alone may misinterpret a driver's state without considering individual strategy differences.
Conclusion
This research demonstrates that MMD is a sensitive metric for capturing the interplay between physical action and cognitive attention. While limited by a small sample size, the framework provides a path toward real-time, customized interventions. Future systems could provide steering assistance to "Learners" showing high variability while withholding it once the driver's MMD stabilizes, thereby improving user acceptance of safety technologies.
Summary updated