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

In three linesPersonaDrive is a retrieval-augmented VLA (vision-language-action) agent pipeline for closed-loop driving simulation, conditioned on retrieved human demonstrations. Trained on CARLA data with aggressive/neutral/conservative instructions, it improves driving score by 4.6% on Bench2Drive and generates style-diverse non-ego agents without per-style retraining.

Summary generated by Claude — human-verified