Characterizing Human Feedback-Based Control in Naturalistic Driving Interactions via Gaussian Process Regression with Linear Feedback

TL;DR

The paper develops a data-driven framework to model human driver interactions at unsigned intersections using Gaussian Process regression with a mixed linear and nonlinear prior to recover interpretable linear gains. By representing driver behavior as a state-feedback controller and learning the gain matrix G, the authors examine how gains vary across regions (approach, intersection, exit) and populations, and they use SVD to extract dominant maneuver patterns. The results show region-specific gain structures and cross-population differences that have implications for designing socially aware autonomous vehicle controllers, while demonstrating robustness through cross-validation and model comparisons. Overall, the work provides a principled, interpretable method to quantify human driving policies and to distill them into low-dimensional, actionable insights for AV control design.

Abstract

Understanding driver interactions is critical to designing autonomous vehicles to interoperate safely with human-driven cars. We consider the impact of these interactions on the policies drivers employ when navigating unsigned intersections in a driving simulator. The simulator allows the collection of naturalistic decision-making and behavior data in a controlled environment. Using these data, we model the human driver responses as state-based feedback controllers learned via Gaussian Process regression methods. We compute the feedback gain of the controller using a weighted combination of linear and nonlinear priors. We then analyze how the individual gains are reflected in driver behavior. We also assess differences in these controllers across populations of drivers. Our work in data-driven analyses of how drivers determine their policies can facilitate future work in the design of socially responsive autonomy for vehicles.

Figures (9)

• Figure 1: Left: Within the virtual environment, each driver has a realistic driving view that encompasses the dashboard of their own virtual car, the simulated driving environment, and the other vehicle. The green arrow on the driver's dash indicates the maneuver to be taken through the intersection. Right: We consider a single intersection scenario in this analysis, in which both cars are instructed to turn left. Car A is denoted in red and Car B is denoted in blue.
• Figure 2: The state space considered in this analysis is the distance between Car A and Car B (purple arrow) and the distance between the lead car's trajectory and a nominal trajectory at each time step $t$. In this example, we presume that Car A is leading, and so $\Delta x$ is the difference between Car A's actual and nominal trajectories at each time step.
• Figure 3: The SVD of the gain matrix $G$ can be intuitively interpreted as the most important components of the inputs acceleration $a$ and angular velocity $\omega$, with in turn have physical meaning associated with turning behaviors.
• Figure 4: The state space is broken into three regions: the approach region denoted by a solid line for both Car A (red) and Car B (blue), the intersection denoted by the yellow circle with radius 10m, and the exit region denoted by a dotted line for Car A (red) and Car B (blue).
• Figure 5: Relative position acceleration gains, $g_{a}^- \in \mathbb{R}^2$: Approach) Forward-backward relative distance causes deceleration in ISR population and acceleration in NYC population. The ISR population also exhibits a wider spread of gains. Intersection) Forward-backward relative distance causes deceleration for both populations, i.e. cars decelerate less when they are in closer proximity in the intersection. Exit) Forward-backward relative distance causes acceleration for both populations.