TLDR: NuRisk is a new Visual Question Answering (VQA) dataset for autonomous driving that enables quantitative, agent-level risk assessment by providing spatio-temporal data. It combines real-world and simulated safety-critical scenarios. Benchmarking shows existing Vision Language Models (VLMs) struggle with spatio-temporal reasoning, achieving low accuracy. However, a fine-tuned 7B VLM agent, developed using NuRisk, significantly improves accuracy to 41% and reduces latency, demonstrating explicit spatio-temporal reasoning capabilities crucial for safer autonomous systems.
Autonomous driving technology has made significant strides, with robotaxi services now operating in defined urban environments. However, a major challenge remains: handling the unpredictable and often rare safety-critical situations that fall outside typical operational boundaries. Ensuring the trustworthiness of self-driving cars requires systems that can reason beyond pre-programmed rules or common training data.
Traditional Vision Language Models (VLMs) have shown promise in understanding driving scenarios, but they often fall short in two key areas: they primarily ground agents in static images, missing how risks evolve over time, and they provide qualitative judgments (e.g., “This is dangerous”) rather than precise, quantitative risk assessments. This limitation means they can’t offer the detailed insights needed for autonomous vehicles to make precise, informed decisions.
Introducing NuRisk: A New Benchmark for Autonomous Driving Safety
To bridge this critical gap, researchers have introduced NuRisk, a comprehensive Visual Question Answering (VQA) dataset designed for agent-level risk assessment in autonomous driving. NuRisk is built on a massive collection of 2.9K scenarios and 1.1 million agent-level samples. It combines real-world data from popular datasets like nuScenes and Waymo with safety-critical scenarios from the CommonRoad simulator, ensuring a diverse and challenging set of situations, including rare collision risks.
What makes NuRisk unique is its focus on providing Bird-Eye-View (BEV) based sequential images, coupled with quantitative, agent-level risk annotations. These annotations allow for spatio-temporal reasoning, meaning models can understand not just what’s happening now, but how risks develop and change over time for each individual vehicle or pedestrian.
How NuRisk Works
The dataset construction involves a multi-stage process. First, data from Waymo, nuScenes, and CommonRoad is extracted, including ego vehicle states and obstacle trajectories. Then, BEV images are generated, providing a consistent overhead view of the driving environment, visualizing all dynamic agents with distinct shapes and colors. Crucially, physics-based risk annotations are computed for each agent in every BEV image. These include metrics like Time-to-Collision (TTC) and Distance to Collision (DTC), which are essential for quantitative risk assessment.
The data is then converted into a conversation format, compatible with existing VLM training frameworks. This format includes sequential images and human queries requesting agent-specific risk analysis, with VLM responses providing structured JSON output containing numerical risk scores, spatio-temporal distance tracking, and chain-of-thought explanations. This allows models to explicitly articulate their reasoning process.
Key Findings and the NuRisk VLM Agent
Initial benchmarking of well-known VLMs (both proprietary like Gemini and open-source) on NuRisk revealed significant limitations. With vision-only inputs, even the best proprietary models achieved a peak accuracy of only 33% and suffered from high latency. They struggled to perform explicit spatio-temporal reasoning, often scoring zero on these critical metrics.
However, the research showed that incorporating physics-based information (like position, velocity, and acceleration) into the input dramatically improved proprietary model performance, boosting accuracy to an impressive 92%. This highlights the importance of explicit physical data for accurate risk assessment. Open-source models, however, struggled to process this additional textual data due to context window limitations.
To overcome these shortcomings, the researchers developed the NuRisk VLM Agent by fine-tuning an open-source 7B VLM (Qwen2.5-VL-7B-Instruct) using the NuRisk dataset. This specialized training proved highly effective. The fine-tuned agent achieved a peak accuracy of 41.1%, significantly outperforming the best pre-trained proprietary models (33%). More importantly, it demonstrated explicit spatio-temporal reasoning capabilities, achieving strong longitudinal and lateral spatial and temporal accuracies, indicating it learned the causal relationship between vehicle dynamics and risk. Furthermore, the NuRisk VLM Agent operates four times faster than leading proprietary models, making it more practical for real-world deployment.
Also Read:
- Improving Robot Navigation with Contextual Textual Descriptions in LLMs
- New Dataset and Evaluator Enhance Safety Perception in Multimodal AI
Conclusion
NuRisk establishes a critical benchmark for advancing spatio-temporal reasoning in autonomous driving. While the task remains profoundly challenging, the NuRisk VLM Agent represents a significant step forward, demonstrating that specialized training can enable VLMs to perform the quantitative, spatio-temporal risk assessment necessary for truly safe and reliable autonomous vehicles. The research paper can be found here.
