SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
SafeDriveVLA uses world-model dreaming to refuse unsafe driving instructions, cutting collisions from 60 to 38.
SafeDriveVLA is a vision-language-action driving model aimed at unsafe or malicious natural-language navigation instructions. It predicts a three-level navigation-compliance mode, informed by a world model that imagines strictly following the instruction, so conflict reasoning is separated from action generation. On Bench2Drive it reports 88.16 Driving Score and 68.64% success rate, raises closed-loop language following by 49.3%, and cuts collisions under unsafe instructions from 60 to 38 versus the best SimLingo variant. The paper also introduces a benchmark covering instruction following, conflict awareness, and adversarial robustness.
- Open language navigation lets inattentive or malicious users issue unsafe instructions.
- SafeDriveVLA predicts a three-level compliance mode before generating actions.
- A navigation-conditioned world model dreams of strictly following the instruction.
- Bench2Drive score is 88.16 DS and 68.64% success; unsafe-instruction collisions fall from 60 to 38.
- Products
- Bench2Drive
- AI models
- SafeDriveVLASimLingo
From the original article
1 min read at huggingface.co
Vision-Language-Action (VLA) models are gaining attention in autonomous driving, inheriting commonsense priors from large language models (LLMs) and widening the navigation interface from a fixed command set to free-form natural-language instructions.
By Shaoyuan Xie, Zihan Zhang, Jingxuan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.