Predicting Risks Before They Happen: Vision-Language Reasoning for Autonomous Driving

arXiv:2511.22928 · cs.RO · Submitted 2025-11-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Predicting Risks Before They Happen".

Dev: Here's the gist of the paper. Ensuring safety remains a key challenge for autonomous vehicles (AVs), especially in rare and complex scenarios One critical but understudied aspect is the potential risk situations, where the risk is not yet observable but can be inferred from subtle precursors, such as anomalous behaviors or commonsense violations Recognizing these precursors requires strong semantic understanding and reasoning capabilities, which are often absent in current AV systems due to the scarcity of such cases in existing driving or risk-centric datasets To address these gaps, we introduce PotentialRiskQA, a novel vision-language dataset designed for reasoning about potential risks prior to observation Based on this dataset, we further propose PR-Reasoner, a vision-language-modelbased framework tailored for onboard potential risk reasoning Experimental results show that fine-tuning on PotentialRiskQA enables PR-Reasoner to significantly enhance its performance on the potential risk reasoning task compared to baseline VLMs.

Rosa: Dataset Construction and Annotation The authors introduce PotentialRiskQA, a languagecentric dataset for potential risk reasoning and cognition This dataset is annotated with structured scene descriptions, semantic precursors, and inferred risk outcomes.

Dev: The base information includes the risk type and ID, key frame annotations, and a brief risk summary Key frame annotations identify three critical moments in the video The Precursor Appearance Frame marks the earliest time when the semantic precursors become observable to the ego vehicle, allowing the potential risk to be inferred The Risk Appearance Frame corresponds to the moment when the risk becomes observable and should be detected by AV system Finally, the (Pseudo) Collision Frame denotes either the point of collision.

Rosa: Data collection involved three primary sources: public accident videos, driving accident datasets, and realvehicle collected data.

Paper summary: Rosa: So we're looking at this paper "Predicting Risks Before They Happen: Vision-Language Reasoning for Autonomous Driving," which is about how autonomous vehicles can anticipate dangers that haven't happened yet. The main idea is that there are these potential risks, things you can see coming from subtle signs, like strange behaviors or things not making sense in the scene.

Dev: Right, so the core problem they're tackling is that current autonomous vehicle systems often only deal with risks we can directly observe right now; this paper claims that recognizing those early precursors requires a really strong understanding of context and reasoning, which existing systems struggle with because there aren't enough examples of these potential risk situations in training data.

Taro: I'm interested in the idea of these precursors—the subtle hints that might signal trouble before the actual danger appears. It sounds like they are trying to build a way for the system to think ahead based on what it sees right now, not just reacting to what is immediately in front of it.

Rosa: Exactly. To fix this data gap, they introduce a dataset called PotentialRiskQA, which is language-based and annotated with structured scene descriptions and the inferred risk outcomes—basically showing the precursors and the eventual risk.

Dev: And then they build on that with a framework called PR-Reasoner, which is a vision-language model based system designed to do this potential risk reasoning for onboard systems. The goal is to take what you see at a specific time and any text context, say, and infer the potential risk that might happen later.

Taro: I like the idea of the reasoning chain they propose: it goes from a high-level scene description to identifying a precursor, then inferring the actual potential risk, suggesting something to avoid driving with, and finally proposing an action for the vehicle. That structure seems really organized for complex decision-making.

Rosa: It's structured into five components in that chain, which includes a high level description of the scene first, then finding that specific precursor, then inferring the risk itself, then a driving suggestion to avoid it, and finally an action that is executable and compatible with current autonomous vehicle systems.

Paper summary: Dev: They support this entire reasoning chain by introducing five separate sub-tasks that provide targeted supervision for each step of the chain, which helps train the model effectively on all those parts at once. I want to know how robust this framework is when the world misbehaves in real time.

Taro: That's a fair question, because when things go wrong in real driving, you need to know exactly what that system does next; how gracefully it recovers from that inferred risk scenario is what matters for safety.

Rosa: The experimental results show that when they fine-tune PR-Reasoner on the PotentialRiskQA dataset, it significantly improves its performance on this potential risk reasoning task compared to baseline vision language models.

Dev: And the key finding they highlight is that this improvement in performance comes from identifying these precursors instead of just waiting for the observable risks, and they quantify that gain as a median increase in mitigation time of one point five seconds.

Taro: A point five second lead time is substantial when you're dealing with sudden events; that extra window gives the autonomous system real breathing room to decide on a safer path before things get dangerous.

Rosa: What's interesting is that they found this reasoning chain paradigm helps smaller models catch up to larger models in this specific task, meaning you don't always need massive model scale for complex reasoning when you have the right training data.

Dev: That addresses one of the big concerns with current AI development where everything seems to require a huge model size just to get decent performance on multi step reasoning tasks.

Taro: So, it sounds like this work is really about bridging that gap between simple observation and deep causal inference in driving scenarios. What stops this from just being a cool research concept and makes it practical for real deployment?

Paper summary: Rosa: That's the next question we have to consider, because while the performance on the task is improved, we still need to figure out how reliably this system would work outside of a controlled lab setting when encountering novel scenarios.

Dev: And from a controls engineering standpoint, we have to worry about latency; how fast can this whole inference process run in real time for an actual vehicle loop? The framework relies on taking a frame sequence and text inputs to generate that structured reasoning chain.

Taro: I think the dataset construction itself is a major hurdle because you need these annotations—the precursor appearance frame, the risk appearance frame, and the pseudo collision frame—which are all very specific temporal markers tied to actual video footage.

Rosa: They used a multi-agent procedure involving four stages for auto annotation of PotentialRiskQA, which means they didn't just rely on one person labeling everything manually; they used an LLM based framework to automate it.

Dev: That automation is smart because it helps generate the data needed for training, but we still have to trust that the LLM-based annotations accurately capture those subtle semantic precursors you're looking for.

Taro: If the auto annotation misses a key precursor, or misjudges the potential risk, then even a well-designed model based on PR-Reasoner won't be safe in practice when it encounters that unseen behavior.

Rosa: So, to wrap up on what this paper is about: "Predicting Risks Before They Happen: Vision-Language Reasoning for Autonomous Driving" proposes a way to use vision language models and a new dataset called PotentialRiskQA to teach autonomous vehicles how to reason about potential dangers before they become observable.

Dev: It suggests that by focusing on these precursors, we can gain extra time for risk mitigation, and the PR-Reasoner framework is shown to be more efficient than baseline models when fine-tuned on this specific data.

Taro: Ultimately, it gives us a foundation for building systems that don't just react to immediate threats but can anticipate and proactively manage potential risks in complex driving situations.

Rosa: That’s what we had: a dataset, a reasoning framework, and results showing improved performance on anticipating dangers before they happen.

Conclusion: Rosa: So, this paper is about taking autonomous driving safety up a notch by looking for risks before they even happen, using vision language models to spot those subtle signs in video footage.

Dev: Yeah, it's called "Predicting Risks Before They Happen: Vision-Language Reasoning for Autonomous Driving," and the authors are focusing on building a dataset and a reasoning framework for that foresight.

Rosa: Basically, they introduce this new dataset called PotentialRiskQA, which is language-based, so it has structured descriptions of scenes showing what might go wrong later.

Dev: And then they propose PR-Reasoner, which is this vision language model based system designed to infer that potential risk based on the visual input and any text context at a specific time step.

Rosa: The authors show that fine-tuning this framework on the PotentialRiskQA dataset makes it significantly better at figuring out these risks than standard vision language models.

Dev: They found that detecting those precursors instead of waiting for a risk to become observable can give the system an extra point five seconds of time to react, which is a concrete lead time for avoiding something dangerous.

Rosa: And what's really interesting is that this reasoning chain structure helps even smaller vision language models perform similarly to much larger ones on this specific task.

Dev: That suggests that the way they built the reasoning chain compensates for the lack of sheer model size, which is a big thing when you're trying to deploy these things onboard where resources are limited.

Rosa: It really seems like this work gives us a solid foundation for making autonomous systems not just reactive, but proactive in managing risk.

Dev: It moves the goal from just reacting to what's right there to having some kind of semantic foresight in driving decisions.

Rosa: Now, the question we have is how reliably this works when the world throws something completely unexpected at it outside of a controlled lab setting.

School of Vehicle and Mobility, Tsinghua University

cs.RO

Submitted: 2025-11-28

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: Ensuring safety remains a key challenge for autonomous vehicles (AVs), especially in rare and complex scenarios One critical but understudied aspect is the potential risk situations…

Key concepts

Potential Risk
A danger that is not currently observable but can be predicted based on subtle signs or precursors. For autonomous vehicles, this means recognizing a dangerous situation developing in the near future, rather than reacting only when an accident is already visible.
PotentialRiskQA Dataset
A novel language-based dataset specifically designed for reasoning about potential risks. It contains structured scene descriptions and 'semantic precursors'—early signs of danger—along with inferred risk outcomes, helping train models to anticipate problems.
PR-Reasoner Framework
A vision-language model framework built for onboard risk reasoning. It takes visual input and text context to generate a structured reasoning chain. This chain moves from understanding the scene to identifying precursors, inferring risks, and suggesting actions.
Reasoning Chain Structure
A five-step process within PR-Reasoner: Scene description (s) -> Precursor identification (p) -> Potential risk inference (r) -> Driving suggestion to avoid risk (a1) -> Action compatible with AV systems (a2). This structured chain makes the model's decision-making process interpretable.

Terminology

Summary

Ensuring safety remains a key challenge for autonomous vehicles (AVs), especially in rare and complex scenarios One critical but understudied aspect is the potential risk situations, where the risk is not yet observable but can be inferred from subtle precursors, such as anomalous behaviors or commonsense violations Recognizing these precursors requires strong semantic understanding and reasoning capabilities, which are often absent in current AV systems due to the scarcity of such cases in existing driving or risk-centric datasets To address these gaps, we introduce PotentialRiskQA, a novel vision-language dataset designed for reasoning about potential risks prior to observation Based on this dataset, we further propose PR-Reasoner, a vision-language-modelbased framework tailored for onboard potential risk reasoning Experimental results show that fine-tuning on PotentialRiskQA enables PR-Reasoner to significantly enhance its performance on the potential risk reasoning task compared to baseline VLMs. p. 0

Dataset Construction and Annotation

The authors introduce PotentialRiskQA, a languagecentric dataset for potential risk reasoning and cognition This dataset is annotated with structured scene descriptions, semantic precursors, and inferred risk outcomes. The base information includes the risk type and ID, key frame annotations, and a brief risk summary Key frame annotations identify three critical moments in the video The Precursor Appearance Frame marks the earliest time when the semantic precursors become observable to the ego vehicle, allowing the potential risk to be inferred The Risk Appearance Frame corresponds to the moment when the risk becomes observable and should be detected by AV system Finally, the (Pseudo) Collision Frame denotes either the point of collision. p. 3 Data collection involved three primary sources: public accident videos, driving accident datasets, and realvehicle collected data. p. 3 The auto annotation uses an LLM-based auto-annotate framework with a multi-agent procedure consisting of four stages. p. 4

PR-Reasoner Framework Design

The authors formalize the potential risk reasoning task as inferring a potential risk r based on visual observation Vt and textual context Tt at time step t The goal is to infer the potential risk r that is not currently observable, but may emerge after a future interval ∆t based on identifiable semantic precursors p in the current input. The PR-Reasoner framework comprises a vision encoder, a textual tokenizer, and a VLM backbone. p. 2 The model takes concatenated visual and textual tokens (XV, XT) as input and feeds them into the VLM to generate the structured reasoning chain C = VLM(XT, XV). p. 2 This output C captures the intermediate inference process from scene understanding to actionable decisions, following the chain s → p → r → a. p. 2

Reasoning Chain Structure

The reasoning chain paradigm C consists of five progressive components designed for interpretable and semantically grounded reasoning in the potential risk reasoning task. The chain includes (1) a high-level scene description (s), (2) identification of a semantic risk precursor (p), (3) inference of the potential risk (r), (4) a driving suggestion to avoid risk (a1), and (5) a meta-action executable and compatible with current AV systems (a2). p. 3 To support this, five auxiliary sub-tasks are proposed, aligned with each component of the chain to provide targeted supervision. p. 3

Performance Analysis and Findings

Quantitative analysis shows that fine-tuning on PotentialRiskQA enables PR-Reasoner to significantly enhance its performance on the potential risk reasoning task compared to baseline VLMs The evaluation employs two types of metrics: traditional standard NLP metrics and LLM-scores which use MLLMs with a strong ability to judge the quality of the answers. Table I shows that PR-Reasoner achieves superior results across various VLM backbones, particularly for small- and medium-sized models. p. 6 The impact of model scale is more pronounced on this task compared to the average of all tasks, indicating that larger models benefit more when handling complex, multi-step reasoning without the saturation observed in overall task metrics. p. 6

Conclusion and Safety Implications

The work provides an essential foundation for improving the semantic foresight and proactive safety capabilities of autonomous driving systems. The findings emphasize the importance of precursor identification, as the median gain in mitigation time from detecting precursors instead of observable risks is 1. p. 350 seconds. p. 0 p. 0 Together, our dataset and model provide a foundation for developing autonomous systems with improved foresight and proactive safety capabilities, moving toward more intelligent and resilient AVs. p. 0 The results demonstrate that the proposed reasoning chain paradigm introduced by PRReasoner compensates for the lack of scale, enabling smaller models to match the similar reasoning ability of larger models in the potential risk reasoning task. p. 6 p. 6 The authors believe this work provides an essential foundation for improving the semantic foresight and proactive safety capabilities of autonomous driving systems. p. 6 p. 6 The paper concludes that PR-Reasoner exhibits superior performance in handling potential risk reasoning under challenging scenarios. p. 6

REFERENCES

[1] L. Peng, B. Li, W. Yu, K. Yang, W. Shao, and H. p. 6 Wang, “Sotif entropy: Online sotif risk quantification and mitigation for autonomous driving,” IEEE Transactions on Intelligent Transportation Systems, vol. p. 6 25, no. 2, pp. 1530–1546, 2023

[2] A. D. Lattner, I. p. 6 J. Timm, M. Lorenz, and O. p. 6 Herzog, “Knowledgebased risk assessment for intelligent vehicles,” in International Conference on Integration of Knowledge Intensive Multi-Agent Systems, 2005

[3] R. Takahashi, N. p. 6 Inoue, Y. Kuriya, S. Kobayashi, and K. p. 6 Inui, “Explaining potential risks in traffic scenes by combining logical inference and physical simulation,” International Journal of Machine Learning and Computing, vol. p. 6 6, no. 5

[4] J. Liu, H. Wang, Z. Cao, W. Yu, C. Zhao, D. Zhao, D. Yang, and J. p. 6 Li, “Semantic traffic law adaptive decision-making for self-driving vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. p. 6 24, no. 12

[5] W. Sun, X. Lin, Y. Shi, C. Zhang, H. Wu, and S. Zheng, “Sparsedrive: End-to-end autonomous driving via sparse scene representation,” arXiv preprint arXiv:2405. p. 619620

[6] H. Caesar, V. Bankiti, A. p. 6 H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y Pan, G Baldan and O Beijbom “nuscenes: A multimodal dataset for autonomous driving” in CVPR

[7] X Zhang, L Wang, J Chen et al., “Dual radar: A multi-modal dataset with dual 4d radar for autononous driving” Scientific Data, vol. p. 6 12, no. 1

[8] T Xie, Z Song et al., “Truckv2x: A truck-centered perception dataset” 2025 [Online]

[9] J Fang, D Yan et al., “Dada: Driver attention prediction in driving accident scenarios” IEEE transactions on intelligent transportation systems, vol. p. 6 23, no. 6

[10] S Malla, C Choi et al., “Drama: Joint risk localization and captioning in driving” in Proceedings of the IEEE/CVF winter conference on applications of computer vision

[11] X Li, J Liu et al., “Graph structure-based implicit risk reasoning for long-tail scenarios of automated driving” in 2023 4th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE) IEEE

[12] J Cao, J Fang et al., “Knowledge graph embedding: A survey from the perspective of representation spaces” ACM Computing Surveys, vol. p. 7 56, no. 6

[13] R Wickramarachchi, C Henson and A Sheth, “Knowledge graphs of driving scenes to empower the emerging capabilities of neurosymbolic ai” IEEE Internet Computing, vol. 28, no. 6

[14] C Lv, M Qi et al., “T2sg: Traffic topology scene graph for topology reasoning in autonomous driving” arXiv preprint arXiv:2411. p. 718894

[15] Y Yao, M Xu et al., “Unsupervised traffic accident detection in first-person videos” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) IEEE

[16] J Fang, J Qiao et al., “Traffic accident detection via self-supervised consistency learning in driving scenarios” IEEE Transactions on Intelligent Transportation Systems, vol. p. 7 23, no. 7

[17] Z Shan, Q Zhu and D Zhao, “Vehicle collision risk estimation based on rgb-d camera for urban road” Multimedia systems, vol. 23

[18] W Hu, X Xiao et al., “Traffic accident prediction using vehicle tracking and trajectory analysis” in Proceedings of the 2003 IEEE International Conference on Intelligent Transportation Systems, vol. p. 7 1 IEEE

[19] K.-H. Zeng, S.-H. Chou et al., “Agent-centric risk assessment: Accident anticipation and risky region localization” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

[20] J Achiam, S Adler et al., “Gpt-4 technical report” arXiv preprint arXiv:2303. p. 708774

[21] A Hurst, A Lerer et al., “Gpt-4o system card” arXiv preprint arXiv:2410. p. 721276

[22] S Bai, K Chen et al., “Qwen2. p. 45-vl technical report” arXiv preprint arXiv:2502. p. 713923

[30] W Bao, Q Yu and Y Kong, “Uncertainty-based traffic accident anticipation with spatio-temporal relational learning” in ACM Multimedia Conference, May 2020

[31] L Xu, H Huang and J Liu, “Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021

[32] C Parikh, D Rawat et al., “Roadsocial: A diverse videoqa dataset and benchmark for road event understanding from social video narratives” arXiv preprint arXiv:2503. p. 721459

[33] Y Wang, K Cheng et al., “Drivingdojo dataset: Advancing interactive and knowledgeenriched driving world model” Advances in Neural Information Processing Systems, vol. p. 7 37

[40] A Marafioti, O Zohar et al., “Smolvlm: Redefining small and efficient multimodal models” arXiv preprint arXiv:2504. p. 705299

[41] Y Zhao, J Huang et al., “Swift:a scalable lightweight infrastructure for fine-tuning” 2024 [Online] Available: https://arxiv. p. 7org/abs/2408. p. 705517

[66% of the length is roughly 380 words, this output is around that length.]

--- Page 1 ---

Seeing before Observable: Potential Risk Reasoning in Autonomous Driving via Vision Language Models

Abstract— Ensuring safety remains a key challenge for autonomous vehicles (AVs), especially in rare and complex scenarios One critical but understudied aspect is the potential risk situations, where the risk is not yet observable but can be inferred from subtle precursors, such as anomalous behaviors or commonsense violations Recognizing these precursors requires strong semantic understanding and reasoning capabilities, which are often absent in current AV systems due to the scarcity of such cases in existing driving or risk-centric datasets Moreover, current autonomous driving accident datasets often lack annotations of the causal reasoning chains behind incidents, which are essential for identifying potential risks before they become observable.

I. INTRODUCTION

The complexity of real driving scenarios has become one of the most significant obstacles to AV safety [1], which lies in the variety of driving scenarios, diverse kinds of traffic participants, their abundant interaction, and complicated semantic relations. Among the various risks in driving, a particularly critical yet often overlooked challenge is the presence of potential risks, exemplified in Fig. p. 01 The definition of potential risk is introduced in [2], [3] as a kind of dangerous events, implicit events that will occur within several seconds In other words, it is a kind of risk that is not directly observable at a given moment, especially by typical autonomous vehicle (AV) systems, but with observable precursors and will appear soon. p. 0 By properly interpreting the contextual precursors, autonomous driving systems can anticipate risks before they are fully observable, gaining more time for risk avoidance and enhancing the system’s safety redundancy. p. 0 However, such precursors are usually difficult for AV systems to recognize, as they may involve complex semantic information, require common-sense reasoning, or be embedded in objects of information that are challenging for current perception models to detect. p. 0 Existing risk datasets used for training AV systems primarily emphasize observable risks, while being deficient in potential risk scenarios. p. 0

Pruning redundant reasoning tokens

The authors introduce PotentialRiskQA, a language-based dataset for potential risk reasoning and cognition This dataset is annotated with structured scene descriptions, semantic precursors, and inferred risk outcomes. The main contributions of this paper are:

• We construct PotentialRiskQA, the first language-based dataset for potential risk reasoning and cognition, addressing the data sparsity and deficiency in existing autonomous driving datasets for potential risk identification challenges.

• We propose a novel annotation and reasoning paradigm for the potential risk reasoning task, featuring a structured and interpretable reasoning chain that bridges semantic precursors and inferred risks, enabling effective training of vision-language models. p. 0

How the training signal is built

The authors propose PR-Reasoner, a VLM–based reasoning framework tailored for onboard potential risk reasoning The framework takes visual input Vt and textual context Tt at time step t to infer the potential risk r that is not currently observable. The output C is the structured reasoning chain, which captures the intermediate inference process from scene understanding to actionable decisions, beginning with a semantic description s, identifying a precursor p, inferring latent risk r, and proposing a driving suggestion a. p. 2 To support this task, five auxiliary sub-tasks are introduced to provide targeted supervision. p. 3

Where the gains come from

Quantitative analysis shows that fine-tuning on PotentialRiskQA enables PR-Reasoner to significantly enhance its performance on the potential risk reasoning task compared to baseline VLMs. The median gain in mitigation time from detecting precursors instead of observable risks is 1. p. 350 seconds. This extended temporal window proves the value of identifying semantic precursors, which provide ADS with critical lead time for risk mitigation. p. 3 The results demonstrate that the proposed reasoning chain paradigm introduced by PRReasoner compensates for the lack of scale, enabling smaller models to match the similar reasoning ability of larger models in the potential risk reasoning task.

Conclusion

We present PotentialRiskQA, a language-based dataset specifically designed for potential risk reasoning in autonomous driving systems We introduce PR-Reasoner, a VLM-based framework tailored for potential risk reasoning We evaluate PR-Reasoner alongside several state-of-the-art VLMs with onboard-scale model sizes, and the results demonstrate its superior performance in handling potential risk reasoning under challenging scenarios. The authors believe this work provides an essential foundation for improving the semantic foresight and proactive safety capabilities of autonomous driving systems.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Introduction of PotentialRiskQA dataset will enable vision-language models to perform potential risk reasoning by providing structured scene descriptions, semantic precursors, and inferred risk outcomes.

  2. Development of PR-Reasoner framework will allow autonomous driving systems to infer risks before they are observable by formalizing the task as r = fr(Tt, Vt), where r is a potential risk not in Tt or Vt but in Tt+∆t or Vt+∆t.

  3. Implementation of the structured reasoning chain C (s → p → r → a) will allow models to generate outputs that are interpretable and semantically grounded, providing a sequence from a high-level scene description (s), identification of a semantic risk precursor (p), inference of the potential risk (r), and a driving suggestion to avoid risk (a).

  4. Training on the dataset will enhance model performance by allowing VLMs to learn from sub-task annotations aligned with components like Scenario Description, Unusual Analysis, Potential Risks, Driving Suggestion, and Meta Action, thereby acquiring the ability to reason about risks in complex and ambiguous scenarios.

  5. PR-Reasoner will enable autonomous systems to gain a substantial temporal advantage that can be leveraged for risk mitigation, as data shows that identifying precursors provides a median gain of 1.50 seconds compared to reacting after risks become observable.

Sources

Related papers