PaperlayerView exact revision on arXiv

Source-linked AI summary

NuRisk: A Visual Question Answering Dataset for Agent-Level Risk Assessment in Autonomous Driving

Yuan Gao, Mattia Piccinini, Roberto Brusnicki, Yuchen Zhang, Johannes Betz

arXiv:2509.25944v2cs.AI

TL;DR

Current VLM risk assessment is largely qualitative and lacks agent-level spatio-temporal risk reasoning. NuRisk pairs sequential BEV scenes with quantitative risk annotations and benchmarks pretrained models alongside a fine-tuned 7B agent. The fine-tuned agent achieves higher accuracy than the reported pretrained baseline and best proprietary model, while responding faster.

Abstract

from arXiv · show

Understanding risk in autonomous driving requires not only perception and prediction, but also high-level reasoning about agent behavior and context. Current Vision Language Model (VLM)-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 2.9K scenarios and 1.1M agent-level samples, built on real-world data from nuScenes and Waymo, completed with safety-critical scenarios from the CommonRoad simulator. The dataset provides Bird's-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. More information can be found at https://github.com/TUM-AVS/NuRisk.

I. INTRODUCTION

VLMs could complement autonomous-driving systems, but existing risk assessment remains largely qualitative and underexplores agent-level spatio-temporal reasoning. NuRisk frames quantitative assessment across time as a VQA capability to evaluate.

I. INTRODUCTION

  • VLMs may complement autonomous-driving stacks, which struggle with real-world variability and rare safety-critical corner cases outside their operational design domains.The paper describes both modular stacks and end-to-end learning approaches as facing this limitation.Source: §I · p. 1
  • Existing VLM risk assessment has focused on hazard detection [7] and qualitative assessment, while lacking quantitative reasoning for rigorous safety evaluation.The motivating question is whether pretrained VLMs can assess agent-level risk using spatio-temporal reasoning about TTC and DTC; such insights matter for motion planners’ decisions.Source: §I · p. 1

1) VLMs in Autonomous Driving:

  • Surveys describe VLM applications across autonomous-driving tasks, and scenario analysis areas including VQA datasets, scene understanding, benchmarks, and risk assessment.The paper places risk assessment within a broader VLM-for-driving literature.Source: 1) VLMs in Autonomous Driving · p. 1

2) VLMs-based Risk Assessment in Autonomous Driving:

  • Prompting-based work applies instructions and contextual cues to pretrained models, [7], [12], while fine-tuning approaches include LoRA and other risk-reasoning systems –[17].The paper characterizes these methods as focused on qualitative agent risk assessment, lacking quantitative labels for explicit spatio-temporal reasoning.Source: 2) VLMs-based Risk Assessment in… · pp. 1–2

3) VQA datasets in Autonomous Driving:

Existing autonomous-driving VQA datasets cover perception, prediction, planning, and some risk-aware reasoning, but do not systematically evaluate quantitative agent-level risk across time. The paper positions NuRisk as addressing this gap with safety-critical coverage and a VLM evaluation and fine-tuning pipeline.

3) VQA datasets in Autonomous Driving:

  • No existing VQA benchmark systematically evaluates quantitative agent-level risk assessment that combines spatio-temporal reasoning with safety-critical metrics.Prior datasets address spatial relations, spatio-temporal grounding, or qualitative risk-aware reasoning [23]–[27].Source: 3) VQA datasets in Autonomous Driving · p. 2

B. Critical Summary

  • Existing approaches largely assess risk at individual time instances, leaving temporal dynamics, agent trajectories, and evolving risk scenarios underexplored in safety-critical contexts.The paper also notes that real-world datasets contain limited safety-critical situations, while existing risk-oriented VQA lacks systematic quantitative evaluation.Source: B. Critical Summary · p. 2

C. Contribution

  • NuRisk contributes 2.9K scenarios and 1.1M agent-level samples, systematic evaluation of pretrained VLM prompting strategies, and a LoRA fine-tuning pipeline for a 7B model.The dataset includes synthetic safety-critical collision scenarios and is intended for explicit spatio-temporal risk reasoning.Source: C. Contribution · p. 2

II. METHODOLOGY

NuRisk combines scenarios from nuScenes, Waymo, and CommonRoad to provide quantitative agent-level risk annotations with sequential BEV scenes. Its construction uses consistent BEV representations and temporally aligned physics-based labels.

II. METHODOLOGY

  • NuRisk fills the absence of quantitative agent-level risk ground truth in nuScenes and Waymo by augmenting their scenarios with CommonRoad safety-critical collisions.The resulting framework supports quantitative risk analysis for individual traffic participants.Source: §II · p. 2

A. NuRisk Dataset Construction

  • The dataset combines 1,000 Waymo scenarios, 850 nuScenes scenarios, and 1,000 CommonRoad safety-critical scenarios, with the real-world sources emphasizing normal driving and CommonRoad emphasizing corner cases.CommonRoad scenarios are generated through motion planning with road networks, traffic signs, and realistic traffic participants.Source: A. NuRisk Dataset Construction · pp. 2–3
  • The pipeline extracts agent and ego trajectories, renders dataset-specific maps into BEV images within a 30-meter ego-centered radius, and preserves distinct agent shapes and colors.BEV representations are used to isolate spatio-temporal risk reasoning from perception uncertainty such as occlusion and lighting variation.Source: A. NuRisk Dataset Construction · p. 3
  • Scenarios are resampled uniformly at 2 Hz, transformed to an ego-centric frame, and annotated with weighted risk scores based on longitudinal and lateral DTC and TTC.The BEV images and annotations are stored together as temporally aligned data.Source: A. NuRisk Dataset Construction · p. 3

B. VQA Dataset Creation

The VQA data pipeline turns each scenario into sequential BEV images paired with multi-agent risk annotations, then organizes the data for VLM training and benchmarking. It also supports per-agent sequence organization and optimizes images for model input constraints.

1) Preprocessing Pipeline:

  • For every scenario and resampled timestep, preprocessing produces a BEV image and a corresponding multi-agent risk set for the unified dataset.The pipeline jointly generates rasterized scene representations and structured risk annotations for traffic participants.Source: 1) Preprocessing Pipeline · pp. 3–4
  • The sequential data pairs BEV representations with per-agent risk annotations, and can be reorganized into per-agent sequences.The paper describes the resulting sequence dataset before introducing an agent-centric organization.Source: 1) Preprocessing Pipeline · pp. 3–4

B. VQA Dataset Creation

2) Conversation Format Conversion:

NuRisk converts its VQA samples into an instruction-following conversation format for sequential, agent-specific risk analysis. Responses use structured JSON to express quantitative assessments and reasoning across time.

2) Conversation Format Conversion:

  • Each sample pairs a sequential image path with alternating human and GPT messages requesting agent-specific risk analysis.Human queries include an <image> token; GPT responses provide structured JSON risk scores and spatio-temporal distance tracking.Source: 2) Conversation Format Conversion · p. 4
  • The responses include chain-of-thought explanations intended to articulate agent-behavior prediction and risk assessment across temporal sequences.The format follows the LLaVA conversation structure and is stated to support compatibility with existing VLM training frameworks.Source: 2) Conversation Format Conversion · p. 4

3) Dataset Quality Assurance:

NuRisk applies automated and human checks during preprocessing, then filters invalid entries. The resulting dataset contains 1.1M agent-level VQA samples with risk scores across levels 0–5.

3) Dataset Quality Assurance:

  • Automated checks validate risk-score ranges, image integrity, conversation completeness, JSON structure, agent fields, and timestep alignment.Sampled data also undergoes human checks of target-agent visibility and risk assessments against human annotations; invalid entries are filtered.Source: 3) Dataset Quality Assurance · pp. 4–5
  • 1.1M agent-level VQA samples make up the final dataset, with 617K from nuScenes, 482K from Waymo, and 64K from CommonRoad.Risk scores are reported as balanced across levels 0–5.Source: 3) Dataset Quality Assurance · p. 5

C. Outcome Analysis

The evaluation framework benchmarks pretrained VLMs with prompting strategies and fine-tunes a NuRisk VLM agent. The fine-tuning pipeline uses LoRA-based adaptation and validation-based checkpoint selection.

1) Benchmarking for pre-trained VLMs:

  • The pretrained-VLM benchmark uses zero-shot inference with prompting techniques to adapt models for autonomous-driving risk analysis.Contextual prompting adds task instructions and safety metrics, chain-of-thought prompts step-by-step spatio-temporal analysis, and in-context learning supplies risk-scenario examples.Source: 1) Benchmarking for pre-trained VLMs · p. 5

2) Fine-tuning NuRisk VLM Agent:

  • Fine-tuning a VLM agent is the proposed approach for enhancing spatio-temporal reasoning in agent-level risk analysis.Qwen2.5-VL-7B-Instruct is adapted using LoRA while keeping the vision encoder and base language-model weights frozen; the fusion layer is also LoRA-adapted.Source: 2) Fine-tuning NuRisk VLM Agent · p. 5
  • Causal language modeling optimizes cross-entropy for risk-assessment conversation completion using structured JSON annotations as supervision.Validation runs every 150 training steps, and the checkpoint with the lowest held-out validation loss is selected.Source: 2) Fine-tuning NuRisk VLM Agent · p. 5
  • The LoRA-based pipeline is designed to transfer to other LLaVA-like VLM architectures without full model retraining.The LLaVA conversation format serves as a standardized training interface.Source: 2) Fine-tuning NuRisk VLM Agent · p. 5

III. RESULTS & DISCUSSION

The results section outlines three evaluation stages: pretrained-model benchmarking, physics-based information, and NuRisk fine-tuning. It also identifies the hardware and VLMs used for experiments.

III. RESULTS & DISCUSSION

  • The evaluation proceeds from pretrained VLMs and prompting, to physics-based information, and finally to fine-tuning a 7B VLM on NuRisk.The stages examine whether domain-specific training can close the performance gap.Source: §III · p. 5

A. Experimental Setup

2) Evaluation Metrics:

The evaluation measures risk-assessment quality, spatio-temporal prediction, and response time, then benchmarks pre-trained VLMs under zero-shot and prompting strategies. Gemini variants lead the reported comparisons, but proprietary models take longer to respond than open-source models.

2) Evaluation Metrics:

  • Risk assessment uses MAE, QWK, accuracy, precision, recall, and F1; spatio-temporal reasoning uses spatial and TTC-based temporal accuracy, while efficiency uses average response time.Spatial accuracy counts longitudinal and lateral predictions within 0.5 meters; temporal accuracy uses a 0.5-second tolerance.Source: 2) Evaluation Metrics · pp. 5–6

B. Experiment 1: Benchmarking for pre-trained VLMs

1) Proprietary vs. Open-Source Performance:

  • Gemini variants achieve the highest QWK scores (0.88-0.89) and peak accuracy (0.33); open-source maxima are QWK 0.72 and accuracy 0.22.The reported proprietary response times are 45.88-107.32 s, compared with 6.55-19.64 s for open-source models.Source: 1) Proprietary vs. Open-Source… · p. 6

2) Prompting Strategy Effectiveness:

The section compares prompting strategies and input configurations, including physics information and sequential images. Physics-enhanced inputs substantially improve proprietary-model performance, whereas sequential images add only modest benefit and open-source models face context-window and latency constraints.

2) Prompting Strategy Effectiveness:

  • Contextual prompting’s task-specific information can enhance performance, whereas chain-of-thought and in-context learning yield negligible improvements while increasing latency.The authors report that even the best proprietary models have moderate accuracy of at most 0.33 with visual information alone.Source: 2) Prompting Strategy Effectiveness · p. 6

1) Impact of Input Modality:

  • The physics-enhanced input experiment compares proprietary and open-source models across single-image and multi-image sequences and contextual, chain-of-thought, and in-context prompting.Table II covers physics-enhanced and sequential-image inputs.Source: C. Experiment 2: Physics-Enhanced… · p. 6, Table II · p. 7, 1) Impact of Input Modality · p. 7
  • 0.92 accuracy and 0.19 MAE are reported for Gemini-2.5-Flash with physics-enhanced input, outperforming its vision-only counterpart and the text-only Gemini-1.5-Pro baseline.Sequential images raise Gemini-2.5-Flash accuracy only from 0.91 to 0.92.Source: 1) Impact of Input Modality · p. 6, Table II · p. 7
  • Textual physics information yields substantial gains particularly for proprietary models, while open-source models do not realize comparable improvements.The section evaluates inputs incorporating position, velocity, and acceleration from the ego and nearby vehicles.Source: 1) Impact of Input Modality · p. 6, C. Experiment 2: Physics-Enhanced… · p. 6
  • Open-source models’ limited context windows hinder processing of token-intensive physics text; InternVL3-8B response time rises from 1.36 s to over 32 s.Multiple images further increase token load and latency for open-source models.Source: 1) Impact of Input Modality · p. 6
  • Sequential visual context offers only marginal gains, and open-source models show no clear advantage from adding sequential images.Gemini-2.5-Flash accuracy increases slightly from 0.91 to 0.92.Source: 1) Impact of Input Modality · pp. 6–7

2) Impact of Prompting Strategy:

Fine-tuning improves quantitative risk accuracy and response time, while the LoRA configurations differ in classification and spatio-temporal performance. Contextual prompting helps pre-trained reasoning models over zero-shot prompting, but advanced strategies add latency without reliably improving performance.

2) Impact of Prompting Strategy:

  • Contextual prompting performs better than zero-shot prompting for pre-trained reasoning models, while other advanced strategies do not improve performance well and substantially increase response time.The authors attribute this to prompting strategies not enhancing internal reasoning when models already have physical information and rely on visual spatio-temporal capabilities.Source: 2) Impact of Prompting Strategy · p. 7

D. Experiment 3: Fine-tuning the NuRisk VLM Agent

  • The two fine-tuned agent variants use LoRA ranks 64 and 256, trained for two epochs on 50,000 NuRisk samples.Source: D. Experiment 3: Fine-tuning the… · p. 7
  • Fine-tuned agents significantly outperform all proprietary models across nearly every metric, according to the comparison normalized to the NuRisk VLM Agent.The comparison inverts response time and MAE scales for interpretability.Source: D. Experiment 3: Fine-tuning the… · p. 7, Figure 5 · p. 7
  • 41.1% accuracy, MAE 1.01, and QWK 0.28: LoRA-256 surpasses Gemini-2.5-Pro’s 33% accuracy and the 14% pre-trained baseline, with a 10.2-second response time, four times faster than the leading proprietary model.Figure 5 compares the proprietary models and both fine-tuned configurations, inverting response-time and MAE scales for interpretability.Source: D. Experiment 3: Fine-tuning the… · p. 7, Figure 5 · p. 7
  • LoRA-256 achieves 34.1% longitudinal and 26.0% lateral spatial accuracy, plus 27.0% longitudinal and 26.4% lateral temporal accuracy; LoRA-64 has the highest QWK (0.304) and stronger temporal reasoning.The authors characterize LoRA-256 as optimized for classification accuracy and LoRA-64 as offering a more balanced profile.Source: D. Experiment 3: Fine-tuning the… · p. 7

IV. CONCLUSION AND FUTURE WORK

  • The conclusion reports that physics-enhanced inputs boost accuracy to 92%, while the NuRisk VLM Agent achieves 41.1% accuracy and runs four times faster than the best proprietary model.Source: §IV · p. 8
  • The conclusion presents NuRisk as a dataset of 2.9K scenarios and 1.1M agent-level samples drawn from nuScenes, Waymo, and CommonRoad.Source: §IV · p. 8
Loading 2509.25944v2…