Human-like driving behaviour emerges from a risk-based driver model
DOI: 10.1038/s41467-020-18353-4
URL: https://doi.org/10.1038/s41467-020-18353-4
archive: archived pipeline: cataloged verified
Summary
Proposes the Driver's Risk Field (DRF), a two-dimensional Gaussian-like field representing the driver's belief about the probability of being at a position in the next preview interval. Multiplied by a cost map of objects in the scene, the DRF yields a scalar perceived-risk metric. A controller that keeps perceived risk below a threshold reproduces human-like driving across seven scenarios (curve radii, lane widths, obstacle avoidance, roadside furniture, car-following, overtaking, oncoming traffic). The DRF formulation expands with steering angle and speed, capturing speed-dependent and curvature-dependent lateral behaviour.
Key finding
A single risk-threshold model based on the Driver's Risk Field reproduces human-like driving behaviour across seven qualitatively different scenarios without scenario-specific switching, supporting the view that satisficing risk minimisation under perceptual/motor noise is a unifying principle of driving behaviour with applications to automated vehicles.
Methodology
Computational driver-modelling study. Authors formulate a kinematic vehicle model coupled to a parametric DRF whose width and length scale with steering angle and speed; a predictive controller minimises perceived risk subject to a threshold. Model predictions are compared qualitatively and quantitatively to driver-behaviour patterns reported in prior simulator and on-road literature (curve negotiation, lane width, obstacle avoidance, roadside furniture, car-following, overtaking, oncoming traffic). For one scenario, predictions are compared to a fixed-base driving-simulator session in which one volunteer drove a track 10 times normally and 10 times in a sporty style.
Sample size: Modelling study; for the in-house simulator validation, N=1 (one 25-year-old male driver, 10 normal trials + 10 sporty trials). Other scenario validations rely on previously published datasets cited from the literature.