The dream of self driving cars hinges on more than clever sensors and slick dashboards. It rests on a quiet, stubborn challenge: how do we test a system that learns from oceans of data, across three big fronts called intelligent cockpits, autonomous driving, and roadside networks? The traditional path has been to gather huge libraries of driving footage and replay them in simulators, hoping to catch the rare misstep. But the real world is messy, and edge cases don’t fit neatly into boxes. A collaboration of researchers from Peking University, CATARC, Geelys ZEEKR, Tsinghua University, and the Chinese Academy of Sciences proposes a radically different approach: treat testing as an on demand question that can be answered by a unified, logic grounded representation of scene data. They call it Query as Test, QaT, built on Extensible Scenarios Notation, ESN. Lead authors for the project are Yilun Lin, Shengyue Yao, and Xiangbin Meng, and the team emphasizes the collaboration of many institutions to bridge cockpit data, vehicle data, and road data into one coherent language.
In the paper published in the IEEE Transactions on Intelligent Transportation Systems in 2025, the authors show how ESN can transform a sprawling, multimodal dataset into a declarative knowledge base. The goal is not just to store data, but to reason with it. This is a shift from data stacking to semantic fusion, from brittle test scripts to flexible logical queries, and from opaque black boxes to human readable explanations. The work is notable not only for its technical ambition but for its explicit aim to make validation more trustworthy, privacy aware, and adaptable to the pace of AI driven development. The project is a cross institutional effort across Peking University International Innovation Center in Shanghai and several industry partners, including CATARC and ZEEKR, as well as universities like Tsinghua and CAS. The combined voice of academia and industry signals a shift toward practical, scalable, and auditable validation in autonomous systems.
From raw data to a logical map
ESN is a way to tame the complexity of an integrated cockpit vehicle road scene by turning heterogeneous data into a single language of facts and rules. The backbone is a declarative programming approach called Answer Set Programming, which supports nonmonotonic reasoning — the ability to handle defaults and exceptions — in a way that mirrors how humans often reason about plans and risk. In ESN, a traffic scene is encoded with two core predicates: holds and occurs. Holds captures the persistent state of a thing at a moment in time, such as a car’s position or speed. Occurs captures instantaneous events, like a brake activation or a lane change. When you string these facts and rules together, you get a dynamic, explainable story of what happened, why it happened, and what could have happened under different rules.
Why go to such lengths? Because logic offers interpretability by design. If a vehicle slows down because its distance to the car ahead is too small, the ESN representation can produce a human readable trace that explains the cause in terms of the rules that were active at that moment. This is a big departure from many current approaches that rely on statistical correlations and opaque model decisions. The ESN approach promises not only to fuse data across cockpit, vehicle, and road domains but also to make the reasoning behind safety decisions legible and auditable, a crucial feature for regulatory confidence and public trust.
The authors also lay out how ESN can be built to work with existing public datasets. They describe how to transform data from the Waymo Open Motion Dataset and nuScenes into ESN facts, a process they call ASP-ification. This transformation creates a cross domain knowledge base that can be queried to discover high level events, run cross domain analyses, and reason about complex scenarios. In short, ESN is a universal translator that lets drivers, cars, and road networks talk to one another in a single logical dialect.
QaT turns testing into a conversation with logic
The heart of the QaT paradigm is to change tests from a fixed collection of scripted scenarios into a living dialogue with the data. QaT uses an LLM as a friendly front end that translates natural language test ideas into structured specifications, but the heavy lifting happens inside a deterministic ASP engine operating on the ESN database. This neuro symbolic loop lets testers pose questions like whether the ego vehicle would violate a safety property under a given set of circumstances, and then the system searches for actual, verifiable counterexamples in the data. If a counterexample exists, the ESN oracle returns a trace that pinpoints the exact time, place, and context of the failure, along with the reasoning that led to it.
An especially powerful feature is counterfactual reasoning. testers can ask what would have happened if the braking policy had been more aggressive, or if rain had been heavier, without running new simulations from scratch. The QaT framework allows such what-if exploration by simply adjusting the rules that define how the vehicle should act, and then re querying the ESN database. It is testing not only for how the system behaves, but for how changes to the decision rules would reshape that behavior. This kind of on the fly counterfactual analysis is a kind of exploration that used to require time consuming re-simulation, now accomplished by a logical reconfiguration and a fast, deterministic search.
Privacy is another pillar. ESN offers native privacy protection through semantic abstraction. Instead of sharing raw trajectories, data owners can express rules that elevate abstracted facts such as regional density or risk levels. Under controlled conditions, refinement rules can permit deeper insights, but outside observers never access sensitive, identifying details. The idea is not to blur with noise but to elevate the level of the data being shared, preserving utility while protecting privacy. This is a practical feature given the growing regulatory emphasis on data protection in AI systems and connected infrastructures.
What the numbers say and what it means for the road ahead
To test whether QaT delivers on its promise, the authors assembled a corpus of 30 driving scenarios spanning urban streets, highways, and unusual edge cases. They compared ESN QaT against three baselines: a traditional SQL based search, a retrieval augmented generation system using embeddings, and a direct LLM based analysis of scenario descriptions. Across 600 test runs per system, the results were striking. ESN QaT achieved a 97 percent success rate, and the average query time was about 0.04 seconds, delivering a high level of semantic understanding with an accuracy of 0.951. These are not small numbers; they indicate a robust, deterministic, and fast reasoning engine at scale that can handle diverse, real world driving conditions.
The SQL baseline, representative of traditional data processing, fared poorly on semantic and reasoning tasks, with a 9.2 percent success rate. The RAG based approach did better on some tasks, achieving about 49.7 percent, but still fell short on deeper interpretability and complex reasoning. A direct LLM approach stood out from the baselines by offering broader language based reasoning, but it delivered only 84.5 percent success and an accuracy of about 0.733, with less determinism and reproducibility. In other words, the data driven logic in ESN QaT combined with a disciplined search engine is not just faster; it is measurably more reliable and explainable than the current wave of purely neural approaches.
The cross domain capability is another strong point. The ESN QaT framework demonstrated cross domain query support in 19 of 20 cases, a 95 percent success rate, with data fusion accuracy around 0.891 and a privacy preservation score near 0.972. In practice, that means engineers can ask questions that require integrating cockpit, vehicle, and road data, and still get precise, privacy minded answers without wrestling with sprawling, error prone data pipelines. The QA metrics used to gauge QaT’s performance — translation fidelity, semantic interpretation, violation detection, and expressiveness — all exceeded defined targets, reinforcing the claim that QaT is not just clever, but genuinely usable in practice.
These results carry more than technical significance. They point toward a shift in how we build and validate AI powered transportation systems. The authors argue for Validation driven Development, a philosophy that uses logical validation to steer the pace and direction of development. In an era where large language models can generate impressive, human like reasoning but struggle with reliability and reproducibility, QaT provides a practical bridge: keep the creativity and flexibility of neural models, but anchor them with symbolic validation that can be read, checked, and audited. The paper also envisions parallel systems architectures in which symbolic ESN QaT and neural LLMs run side by side, each learning from the other, to achieve safer, more reliable autonomy on the road.
In a field that often feels like software engineering on fast forward, the ESN QaT approach offers a quiet but powerful counterweight to fragility. It does not pretend to replace data driven learning or end the need for big simulations. Instead, it gives safety validation a sharper voice, a clearer map, and the ability to ask the right questions at the right time. The collaboration behind the study, spanning top universities and industry players across China, signals a broader trend toward principled, auditable AI systems in transportation — a trend that has real implications for regulation, consumer trust, and the pace at which autonomous technologies can safely scale to everyday life.
