A Hybrid Dynamic Model for Predicting Human Cognition and Reliance during Automated Driving

TL;DR

The paper introduces a hybrid dynamic framework that jointly models continuous cognitive states (trust, perceived risk, workload) and discrete reliance in automated driving using a simple piecewise-affine structure. It demonstrates that this approach can be personalized from a single trajectory and remains interpretable through threshold-based triggering, enabling tailored interventions such as take-over requests. Empirical results from 16 participants show meaningful, participant-specific fits, with trust and risk as primary drivers of reliance and some capacity for predicting unseen data. The work advances cognition-aware automation by offering a tractable model suitable for online adaptation and design insights for human-centric vehicle automation.

Abstract

We propose a simple (12 parameter) hybrid dynamic model that simultaneously captures the continuous-valued dynamics of three human cognitive states-trust, perceived risk, and mental workload-as well as discrete transitions in reliance on the automation. The discrete-time dynamic evolution of each cognitive state is modeled using a first-order affine difference equation. Reliance is defined as a single discrete-valued state, whose evolution at each time step depends on the cognitive states satisfying certain threshold conditions. Using data collected from 16 participants, we estimate participant-specific model parameters based on their reliance on the automation and intermittently self-reported cognitive states during a continuous drive in a vehicle simulator. The model can be estimated using a single user's trajectory data (e.g. 8 minutes of driving), making it suitable for online parameter adaptation methods. Our results show that the model fits the observed trajectories well for several participants, with their reliance behavior primarily influenced by trust, perceived risk, or both. Importantly, the model is interpretable, such that the variations in model parameters across participants provide insights into differences in the time scales over which cognitive states evolve, and how these states are influenced by task complexity. Implications on the design of human-centric vehicle automation design are discussed.

Figures (12)

• Figure 1: Binary signal representing task complexity during the main drive.
• Figure 2: A photo of the driving simulator used in this study.
• Figure 3: Participant's point-of-view when approaching a construction zone.
• Figure 4: Flowchart describing the experimental procedure. Pink boxes depict steps involving subjective data collection, while green boxes involve use of the driving simulator.
• Figure 5: State trajectory for Participant 01. The blue dashed lines show the simulated trajectory generated by the identified model, while the black dashed lines indicate the estimated active thresholds. The blue asterisks denote the cognitive states self-reported by the participant. Fit: RMSE$_T=0.0177$, RMSE$_R=0.0269$, RMSE$_W=0.0644$, Acc.(%)=91.21.