The Road Has Already Told Us Where It’s Dangerous. Are We Listening?

Every fatal crash on a U.S. public road since 1975 is on record. The conditions, the road type, the speed, who was involved, and what happened in the seconds before impact. The National Highway Traffic Safety Administration’s (NHTSA) Fatality Analysis Reporting System (FARS) has been quietly accumulating this data for five decades. Over one million crashes. Freely available. And yet largely untapped by the autonomous vehicle industry.

That is a problem worth fixing.

The Data Is Not Synthetic. That Is the Point.

Synthetic simulation is a powerful tool. It lets engineers explore edge cases and build datasets to test against. Synthetic datasets reflect the assumptions of the engineers who designed them, but we can do better.

The crash database reflects reality and captures the full diversity of U.S. driving conditions: rural two-lane roads at midnight, arterial intersections at school dismissal, highway on-ramps in rain and fog. No synthetic dataset reproduces this distribution.

The variables encoded in each record map directly onto autonomous vehicle development: road geometry, speed, lighting, weather, surface condition, crash configuration, and the presence of pedestrians, cyclists, or animals. Every record is a documented failure mode. And failure modes are exactly what simulation is for.

Combining scenarios drawn from both requirements and real-world events is a powerful asset.

Risk Is Not Evenly Distributed. Neither Should Our Testing Be.

The data surfaces patterns that should directly shape how autonomous systems are tested.

Rear-end crashes are a common occurrence for roughly 1 in 6 fatal multi-vehicle crashes. They are also among the most preventable. Following distance, deceleration prediction, and cut-in detection are well-understood problems. The data tells us they are not yet solved.

Roadside departures dominate single-vehicle fatalities. Loss of lateral control, obstacle avoidance maneuvers, and road-edge departures account for a significant share of deaths on rural roads at highway speeds.

Intersecting-path crashes are the classic T-bone. Cross-traffic detection and intersection behavior remain among the hardest problems in urban autonomy, and the fatality record reflects that.

We have the data and the technology. Using both together we can create a testing strategy that is more effective and develop physical AI systems that will save lives. Sources like the crash database tell us where to concentrate that effort.

Vulnerable Road Users: The Data Is Specific Enough to Act On.

Are all Vulnerable Road Users (VRUs) the same? No, and the data lets us be precise about the differences.

Pedestrian and cyclist fatalities are not randomly distributed. They are concentrated in specific conditions (by age, time of day, lighting etc) that we can use to improve the effectiveness and coverage when testing ADAS and AV systems.

Senior pedestrians (65+) are overrepresented in road-crossings crashes, during daylight hours, at moderate vehicle speeds. This is not a perception problem at the limits of sensor range. It is a yield-and-intent-prediction problem in well-lit, routine conditions.

Children are overrepresented in residential areas in the late afternoon. Lower visual profile, unpredictable movement, and significant surrounding clutter make this a demanding detection scenario despite the low speeds involved.

Cyclists are most at risk on roads when a vehicle overtakes them from behind, the single largest fatal cyclist crash type with side and crossing impacts close behind. These occur across both daylight and darkness, predominantly on higher-speed arterials. Both geometries are detectable. Both have proven fatal when missed.

From Records to Requirements: Parameterized Reconstruction

Each crash record contains enough information to construct a simulation scenario: ego speed, road type, lighting, weather, and participant configuration. Translating a crash record into a test case is an exercise in parametric reconstruction and we can do this at scale with AI and automation.

By using the metadata of a real-world tragedy, we shift the testing philosophy from only “Can we handle a hypothetical?” to also include “Can we handle the documented reality?” The question is no longer whether a system performs in a vacuum, but whether it would have performed differently from the driver whose crash it reconstructs.

And each event can become a template for thousands of variations, answering questions such as “What if the speed was different, or the pedestrian was moving slower?”

Closing the Loop: Data-Driven Validation

This is the approach we have built at Applied Intuition. We treat the crash database not just as a record of the past, but as a blueprint for the future. By ingesting these real-world failure modes, our platform enables engineers to move beyond best-guess synthetic scenarios.

The goal is to use real-world crash data to identify where risk is concentrated and which failure modes are underrepresented in existing test libraries, ensuring that simulation programs reflect the actual distribution of risk on public roads while augmenting rare “long tail” scenarios through parametric simulation.

The Question is Not If. It is When.

An autonomous vehicle operating in the U.S. will encounter a child crossing a residential street in the late afternoon. It will encounter a cyclist on an unlit arterial. It will encounter a stopped vehicle on a dark highway. The crash database does not just predict this; it documents that it has already happened, thousands of times, under conditions we can describe precisely.

The road has already given us the answers. The patterns are clear, and the data is waiting. To build an autonomous future without accounting for the fatal realities of the past is a missed opportunity to save lives. It’s time to stop guessing what’s dangerous and start listening to what the road has been telling us for fifty years.

At Applied Intuition, our mission is to help the industry deploy safe and intelligent machines. We built this approach to serve our internal autonomy program and all our customers as we have one common goal: improve safety on the road and create a better future for all of us.