top of page
-04.png){kind=small}
Dataset
Global Coverage and Behavioral Normalization
Our database provides over 500 million kilometers of real-world accident and incident coverage, with an additional 10% of comparative data from Europe, the US, and Australia.
Research confirms that human avoidance behavior—how people interpret complex traffic situations and act to prevent accidents—is consistent worldwide, enabling accurate cross-regional modeling.
Using more than 9 Billion kilometers of accident statistics from Singapore, we support normalizing data to regional conditions, adapting exposure and event likelihood across different regions, city zones, countries, and environments to ensure realistic and comparable safety analysis.
Resembler Icon Query Language (RIQL)
The Resembler Icon Query Language (RIQL) makes complex data exploration simple and visual.
Using hundreds of proprietary semantic elements, RIQL allows users to search, filter, and compare accidents and incidents across our 500M+ km AV dataset with intuitive icon-based logic — taking the complexity out of deep analysis.
Built on insights from the AV Human Benchmark Dataset, RIQL enables rapid understanding of how autonomous systems perform in real-world conditions, supporting faster validation, safer algorithms, and more transparent performance assessment.
ODD Comparison
Resembler.ai offers an ODD Comparison Tool to verify and compare Operational Design Domains (ODDs) and deployment areas across regions.
Based on real accidents and incidents, the tool integrates human accident avoidance behavior into the selected infrastructure and environment.
The tool highlights differences in road design, traffic dynamics, and environmental factors that influence autonomous driving performance — enabling city planners and developers to make data-driven improvements and reduce deployment risk for automated systems.
Annotated Risk Frames
Resembler.ai’s dataset, enriched with expert-led annotations, includes bounding boxes, risk progression data, and contextual tags derived from real-world accidents and incidents.
The database highlights frames where risk emerges before an incident occurs, carefully selected from the best-of-best and worst-of-worst examples of defensive driving.
These annotations capture how humans perceive emerging risk and often intelligently circumvent situations where others fail to follow traffic rules.
The data can be fed directly into ADAS and AD vision systems, or — through OpenDRIVE and CARLA — into the planning and intention-prediction stack, helping developers verify, benchmark, and enhance safety performance across diverse environments.
bottom of page