Zero-Shot ODD Perception with Vision-Language Models

A new arXiv paper (2605.07649) explores using vision-language models (VLMs) as zero-shot sensors for Operational Design Domain (ODD) perception in autonomous systems, particularly Automated Driving Systems (ADS). The research addresses the need for adaptable perception that can handle evolving ODD definitions without task-specific training data. VLMs integrate visual recognition and language reasoning, making them suitable for this role. The study contributes to safety regulation compliance in autonomous driving.