PersonaDrive: Human-Style Retrieval-Augmented VLA Agents for Closed-Loop Driving Simulation
Researchers have developed PersonaDrive, a novel pipeline for creating more human-like non-ego traffic agents in closed-loop driving simulations. This system conditions a vision-language-action (VLA) agent on retrieved driving demonstrations from a dataset where humans were instructed to drive in specific styles (aggressive, neutral, conservative). The pipeline efficiently fuses visual features with control signals and fine-tunes a VLA backbone to use these retrieved contexts as behavioral demonstrations, enabling style-diverse agents without per-style retraining. AI
IMPACT Enhances realism in driving simulations by enabling more human-like agent behavior, potentially improving training and testing of autonomous systems.