Explore the scholarly record.

Abstract

Safe gap negotiation at motorway ramps and lane drops is essential to reliable automated driving in mixed traffic. Existing planning architectures provide limited evidence on how behavioral information and language inputs should influence gap selection while preserving quantitative assessment of traffic responses. This study develops a discrete path selection framework that combines reactive prediction, driver type beliefs, logit quantal response, and conventional vehicle control. Human gap acceptance is calibrated on the highD dataset and the Aerial Dataset for China's Congested Highways and Expressways (AD4CHE). Two large language models (LLMs), DeepSeek and Kimi, supply either constrained type beliefs or direct maneuver commands. In 200 paired merging simulations, the reactive planning configuration increased the minimum time headway proxy from 0.309 to 1.140 s and reduced mean absolute jerk by 68.0%, while all merges were completed. Human models with finite sensitivity improved held-out likelihood on both datasets. Controller variants with inferred driver types increased gap-taking activity, while safety outcomes depended on the scenario. Across both LLMs, belief-based planning completed 94% of mandatory lane changes with a 2% collision rate; direct commands completed 28–32% with 68–72% collisions. These results compare complete controller configurations. Controlled path interventions further linked gap commitments to subsequent traffic outcomes. The findings support integrating behavioral and language information upstream of quantitative path evaluation, providing a structured basis for smoother merging and more reliable completion of required lane changes in mixed traffic.