PersonaDrive lets autonomous driving agents adopt human-like styles

PersonaDrive tackles a key limitation in closed-loop driving simulators: the lack of behavioral diversity among traffic agents. Current methods either use rule-based managers or single-mode learned models, with style variations added via post-hoc labels or LLM-inferred rewards that are proxies rather than true human demonstrations. The pipeline includes three stages: offline triplet mining over per-style human driving data using a combined image-text similarity score; training a lightweight retrieval head that fuses frozen visual features with a small control encoder; and fine-tuning a single VLA backbone to treat retrieved context points as in-context behavioral demonstrations during waypoint prediction. At inference, style is selected simply by swapping which per-style database the retrieval head queries.

On the Bench2Drive benchmark, PersonaDrive (without style conditioning) achieves a driving score 4.6% higher than SimLingo and 2.5% higher than HiP-AD. Under style conditioning, it attains the highest driving score in every style within a roughly 2% band, with its weakest style still surpassing the strongest baseline (DMW) by 5.4%. Average speed increases by 18% and acceleration by 25% from the conservative to the aggressive instruction. This work enables more realistic, human-like traffic simulations by allowing agents to mimic distinct driving personas—critical for training robust autonomous systems.