AUTOPILOT-VQA: incident-centric VQA benchmark for dashcam video understanding

Researchers present AUTOPILOT-VQA, a new visual question answering benchmark designed to evaluate how well vision-language models reason about safety-critical incidents in dashcam videos. The dataset utilizes structured questions based on real-world driving scenarios to assess model capabilities beyond simple object recognition.

• The benchmark covers diverse safety-relevant categories including weather, lighting, road layout, signage, and accident avoidability.
• It requires models to answer grounded questions about both contextual scene properties and specific event-level incident details.
• This approach shifts the focus toward temporally grounded, safety-aware reasoning for autonomous driving systems.
• The dataset is released as part of the AUTOPILOT CVPR 2026 competition to provide a standardized assessment tool.

The authors consider this important because it supports the development of more interpretable and robust vision-language systems for real-world autonomous driving.