Driver distraction gets a new test: a hierarchical model makes lane and speed information compete for sampling, and all eight measures land within one human standard deviation
Synopsis
Deciding whether to look at the lane or the speedometer becomes a problem of allocating limited visual resources under partial observability. Earlier models treated lane keeping and speed control separately, and none formalized sampling decisions alongside coupled lateral and longitudinal control. A hierarchical computational rationality model learns lane keeping, speed maintenance and supervisory attention allocation together; across 30 simulated runs, all eight measure means fell within one human standard deviation, though the model produced no lane deviations.
Interpretation
Visual sampling is recast as a supervisory decision: when lane and speed information can both grow uncertain, the driver must choose which one to update next, not only when to look. Earlier computational models either handled sampling timing in lane keeping alone or combined lateral and longitudinal control without modeling sampling decisions, leaving information allocation under parallel uncertainty unformalized. The model was compared descriptively with simulator data from 30 participants across 30 seeded runs; all eight measure means fell within one human standard deviation and within the observed human minimum-to-maximum range, with a mean absolute standardized distance of 0.48 human standard deviations.
The model has three reinforcement-learning policies: one steers to hold the lane, one accelerates and decelerates to hold speed, and a top-level policy decides each step whether to sample lane information, speed information or stay occluded. The two lower controllers read only internal belief estimates and cannot see true vehicle state unless the matching channel is sampled; during occlusion the beliefs propagate through known dynamics, so estimates drift with occlusion time. Training ran in three stages, lane keeping, then speed maintenance, then attention allocation with the first two frozen; evaluation used a 100–110 km/h target band, 100 km/h initial speed and a sinusoidal undulating road, with episodes up to 60 s.
The attention policy's reward contains only task-failure penalties, a completion reward, a switch cost for initiating a sample and the belief divergence revealed at sampling, with no prescribed sampling pattern. Human data served only as an evaluation target rather than being written into the reward to reproduce, so the observed sampling rhythm emerged from task constraints. Occlusion time averaged 0.97 s for the model versus 0.85 s for humans, occlusions per 10 s were 6.94 versus 7.20, and mean TLC was 9.57 s versus 8.92 s, all within one human standard deviation.
Perspective
The evaluation covers 30 seeded simulation runs and 30 participants (27 with valid eye-tracking) in a self-paced occlusion task on a medium-fidelity simulator, on a 1500 m radius, 3.75 m lane-width circular undulating road with a 100–110 km/h target band and episodes up to 60 s; the authors propose next adding belief noise, perceptual uncertainty or participant-specific steering-skill and cautiousness parameters, plus participant-level parameter fitting.
The authors report that the model produced no lane deviations during evaluation, against 0.29 per 10 s in humans, indicating overly stable lateral control; beliefs are deterministic point estimates without explicit uncertainty; lane and speed are treated as mutually exclusive channels, each glance fixed at 500 ms and startable only from occlusion; the switch cost, divergence weights and timeout thresholds are modeling choices rather than estimates from participants; aggregate agreement does not guarantee reproduction of individual differences.