AI RESEARCH
Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models
arXiv CS.CV
•
ArXi:2603.06049v1 Announce Type: new We identify a fundamental Narrow Policy limitation undermining the performance of autonomous VLA models, where driving Imitation Learning (IL) tends to collapse exploration and limit the potential of subsequent Reinforcement Learning (RL) stages, which often saturate prematurely due to insufficient feedback diversity. Thereby, we propose Curious-VLA, a framework that alleviates the exploit-explore dilemma through a two-stage design. During IL, we