PersonaDrive: Human-Style Retrieval-Augmented VLA Agents for Closed-Loop Driving Simulation

PersonaDrive is a novel pipeline for enhancing closed-loop driving simulations by conditioning a vision-language-action (VLA) driving agent on human-style demonstrations retrieved from a dataset of drivers instructed to adopt various styles (aggressive, neutral, conservative). It employs a three-stage process: triplet mining for style-specific data, training a retrieval head that integrates visual features with control inputs, and fine-tuning a unified VLA backbone to leverage in-context behavioral demonstrations. On the Bench2Drive benchmark, PersonaDrive achieves a 4.6% improvement over SimLingo and consistently outperforms other models across styles, indicating significant advancements in realistic traffic agent behavior for practitioners developing AI-driven simulation environments.