PersonaDrive — VLA agents learn human driving styles from demos
arXiv paper introduces PersonaDrive, a retrieval-augmented approach that conditions vision-language-action (VLA) driving agents on human demonstrations labeled by style.
Rather than inferring style through post-hoc labels or LLM reward weights, the system retrieves real examples from a dataset where humans explicitly drove CARLA routes in aggressive, neutral, or other instructed modes.
Moves traffic simulation beyond homogeneous rule-based agents toward realistic behavioral diversity in closed-loop environments.