Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Chen, Canyu, Yang, Yuguang, Tan, Zhewen, Wang, Yizhi, Zhan, Ruiyi, Liu, Haiyan, Mao, Xuanyao, Bao, Jason, Tang, Xinyue, Yang, Linlin, Sun, Bingchuan, Wang, Yan, Zhang, Baochang
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917318060146688
author Chen, Canyu
Yang, Yuguang
Tan, Zhewen
Wang, Yizhi
Zhan, Ruiyi
Liu, Haiyan
Mao, Xuanyao
Bao, Jason
Tang, Xinyue
Yang, Linlin
Sun, Bingchuan
Wang, Yan
Zhang, Baochang
author_facet Chen, Canyu
Yang, Yuguang
Tan, Zhewen
Wang, Yizhi
Zhan, Ruiyi
Liu, Haiyan
Mao, Xuanyao
Bao, Jason
Tang, Xinyue
Yang, Linlin
Sun, Bingchuan
Wang, Yan
Zhang, Baochang
contents 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 introduce a Feasible Trajectory Expansion (FTE) strategy to generate multiple physically valid trajectories and a step-wise normalized trajectory representation to adapt this diverse data. In the RL stage, we present Adaptive Diversity-Aware Sampling (ADAS) that prioritizes high-diversity samples and introduce Spanning Driving Reward (SDR) with a focal style weighting to amplify reward's value span for improving sensitivity to driving quality. On the Navsim benchmark, Curious-VLA achieves SoTA results (PDMS 90.3, EPDMS 85.4) and a Best-of-N PDMS of 94.8, demonstrating its effectiveness in unlocking the exploratory potential of VLA models. Code: https://github.com/Mashiroln/curious_vla.git.
format Preprint
id arxiv_https___arxiv_org_abs_2603_06049
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models
Chen, Canyu
Yang, Yuguang
Tan, Zhewen
Wang, Yizhi
Zhan, Ruiyi
Liu, Haiyan
Mao, Xuanyao
Bao, Jason
Tang, Xinyue
Yang, Linlin
Sun, Bingchuan
Wang, Yan
Zhang, Baochang
Computer Vision and Pattern Recognition
Robotics
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 introduce a Feasible Trajectory Expansion (FTE) strategy to generate multiple physically valid trajectories and a step-wise normalized trajectory representation to adapt this diverse data. In the RL stage, we present Adaptive Diversity-Aware Sampling (ADAS) that prioritizes high-diversity samples and introduce Spanning Driving Reward (SDR) with a focal style weighting to amplify reward's value span for improving sensitivity to driving quality. On the Navsim benchmark, Curious-VLA achieves SoTA results (PDMS 90.3, EPDMS 85.4) and a Best-of-N PDMS of 94.8, demonstrating its effectiveness in unlocking the exploratory potential of VLA models. Code: https://github.com/Mashiroln/curious_vla.git.
title Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models
topic Computer Vision and Pattern Recognition
Robotics
url https://arxiv.org/abs/2603.06049