SafeDriveVLA decouples navigation-conflict reasoning from driving actions
A paper accepted at CoRL 2026 and submitted to arXiv on 14 October (arXiv:2610.20609) proposes SafeDriveVLA, a driving vision-language-action model that explicitly predicts a three-level navigation-compliance mode to decouple conflict reasoning from action generation, and uses navigation-conditioned world-model dreaming to determine when following the navigation signal is safe. The authors also present a benchmark suite that evaluates navigation following, conflict awareness and adversarial robustness together, reporting that existing VLAs fall short on both, with SimLingo completing no language-instructed lane change in closed loop. They report SafeDriveVLA reaching 88.16 Driving Score and 68.64% Success Rate on Bench2Drive, lifting closed-loop language following by 49.3% and cutting collisions under unsafe instructions from 60 to 38 versus the best SimLingo variant.
The paper argues that free-form natural-language navigation creates a new vulnerability surface: an inattentive driver may issue instructions unsafe for the current scene, and a malicious user may issue dangerous ones on purpose, so models must explicitly judge when not to follow.
Computer Science > Robotics
arXiv:2610.20609 (cs)
[Submitted on 14 Oct 2026]
Title:SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
Authors:Shaoyuan Xie, Zihan Zhang, Jingxuan Wang, Jiashu Qu, Xiaoqing Liang, Lingdong Kong, Junchi Lu, Henrik I. Christensen, Qi Alfred Chen
View a PDF of the paper titled SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving, by Shaoyuan Xie and 8 other authors