NuRisk: A Visual Question Answering Dataset for Agent-Level Risk Assessment in Autonomous Driving
Yuan Gao, Mattia Piccinini, Roberto Brusnicki, Yuchen Zhang, Johannes Betz
2025-10-06
Summary
This paper introduces a new dataset and testing method for evaluating how well AI systems can understand and predict risks in self-driving cars, going beyond just 'seeing' the environment to actually 'thinking' about how situations will unfold over time.
What's the problem?
Current AI systems for self-driving cars, specifically those using Vision Language Models, are good at identifying objects in images and making basic judgments, but they struggle with understanding how risks change as time passes. They can't effectively reason about how different cars and pedestrians will move and interact, which is crucial for safe driving. Existing datasets don't provide the detailed, time-based information needed to train and test this kind of reasoning ability.
What's the solution?
The researchers created a new dataset called NuRisk, which includes nearly 3,000 driving scenarios and over a million individual observations of vehicles and pedestrians. This dataset uses a 'bird's-eye view' to show how things change over time and includes specific measurements of risk for each agent in the scene. They then tested existing AI models on this dataset and found they weren't very good at predicting how risks would evolve. To improve performance, they fine-tuned a specific AI model, making it better at understanding the sequence of events and reducing the time it takes to make predictions.
Why it matters?
This work is important because accurately predicting risk is essential for self-driving cars to be truly safe. The NuRisk dataset provides a challenging benchmark for researchers to develop and improve AI systems that can reason about complex driving situations. The fact that even the improved AI model isn't perfect highlights how difficult this problem is and emphasizes the need for continued research in this area.
Abstract
Understanding risk in autonomous driving requires not only perception and prediction, but also high-level reasoning about agent behavior and context. Current Vision Language Models (VLMs)-based methods primarily ground agents in static images and provide qualitative judgments, lacking the spatio-temporal reasoning needed to capture how risks evolve over time. To address this gap, we propose NuRisk, a comprehensive Visual Question Answering (VQA) dataset comprising 2,900 scenarios and 1.1 million agent-level samples, built on real-world data from nuScenes and Waymo, supplemented with safety-critical scenarios from the CommonRoad simulator. The dataset provides Bird-Eye-View (BEV) based sequential images with quantitative, agent-level risk annotations, enabling spatio-temporal reasoning. We benchmark well-known VLMs across different prompting techniques and find that they fail to perform explicit spatio-temporal reasoning, resulting in a peak accuracy of 33% at high latency. To address these shortcomings, our fine-tuned 7B VLM agent improves accuracy to 41% and reduces latency by 75%, demonstrating explicit spatio-temporal reasoning capabilities that proprietary models lacked. While this represents a significant step forward, the modest accuracy underscores the profound challenge of the task, establishing NuRisk as a critical benchmark for advancing spatio-temporal reasoning in autonomous driving.