How Waymo Tests Self-Driving Cars for Crash Safety

Waymo has created a testing tool it calls the Reference Driver model — a kind of virtual crash test dummy designed to see how well its self-driving system handles sudden, dangerous situations on the road.

Think of it this way: when car makers test how well a vehicle protects passengers, they use crash test dummies that simulate a human body. Waymo's Reference Driver is similar, except instead of testing what happens when a car hits something, it tests how well the car can avoid the hit in the first place.

The model is not based on a real person. Instead, it is a made-up "ideal driver" that Waymo uses as a benchmark. When something unexpected happens on the road, Waymo asks: would my self-driving car respond as well as, or better than, this reference driver?

How the Reference Driver Works

The Reference Driver model tests how the Waymo system reacts in reconstructed crash scenarios. Waymo does not just ask "did the car crash?" Instead, it asks a more detailed question: given the exact same conditions that caused a crash in real life, would the Waymo Driver have avoided it?

The reference driver itself is based on what Waymo calls NIEON — a level of performance that does not exist in real humans, and is deliberately set higher than how most actual people drive. This is an important choice. If Waymo used an average human driver as its benchmark, the self-driving car could pass the test while still making dangerous mistakes that matter when thousands of these vehicles are on the road. By aiming higher, Waymo forces its system to meet a standard that no single person could reliably achieve.

Specific Crash Patterns

Waymo has also published research on which types of crashes at intersections matter most for testing. Intersections are some of the highest-risk places in driving, so Waymo focuses on crash patterns that happen there — different approach angles, speeds, traffic signal states, and blocked views.

This is the same logic that car makers use when they test passive safety features like airbags. You cannot test every possible crash, so you select a representative set that covers most of the real-world injuries that happen. Waymo's approach is similar: instead of testing against an endless number of scenarios, engineers test against a defined set that represents the crashes that actually cause harm.

Waymo has also tested its system on real fatal crashes from the past. The team reconstructs the exact conditions of a crash that killed someone, then asks whether the Waymo Driver would have prevented it.

What the Safety Numbers Show

According to Waymo's published data, its self-driving system records about 0.02 serious injuries or deaths per million miles driven. For comparison, human drivers record about 0.22 per million miles — roughly ten times higher.

That gap is worth taking seriously. But it also comes with an important caveat: Waymo's number applies only to specific cities with clear maps and known road conditions, not to all the different types of roads across America. Still, the difference is large enough that even if the measurement methods are not perfect, something real is being measured. The point is not that the system is flawless, but that under the conditions where it operates, it outperforms human drivers by a significant amount.

There is a deeper question underneath these numbers. Human driving data exists everywhere, but it is messy and mixed with many different factors — road type, weather, time of day, how busy the roads are. Waymo's testing method, with its reference driver and specific crash types, tries to give a clearer picture than a simple overall rate can.

A Conversation That Started Decades Ago

Anyone who has covered car safety over the past thirty years will recognize this pattern. Back in the 1990s, when cars started getting electronic stability control — a system that automatically corrects skids — the industry had the same debate. Should a computer override what the driver is doing in an emergency? What is the right standard to judge whether a machine's decision is safe?

Electronic stability control won that argument. It is now required on all new cars sold in the United States and Europe. The benchmark that justified it was not how the average person drives, but what a car could theoretically do if everything went perfectly. Waymo's Reference Driver follows the same logic, just applied to self-driving cars.

Why This Matters Beyond Waymo

The U.S. government has not set strict safety rules for self-driving cars. Instead, it has mostly let companies develop and publish their own testing methods. This means that what Waymo and other companies say about how they test safety carries more weight than it normally would.

If Waymo's Reference Driver approach becomes the standard that the industry, regulators, and insurance companies all use, it creates a shared language. That kind of agreement has historically been the stepping stone from companies testing themselves to independent third parties and government agencies doing the testing.

For engineers and safety teams working on self-driving cars, this means the virtual crash test dummy concept is becoming less of an internal tool and more like a standard the whole industry might adopt. How quickly that happens, and whether regulators accept it, will be one of the important safety questions for self-driving cars over the next few years.