Waymo has developed a new computer model designed to better assess how its robotaxis perform compared with human drivers.
The Alphabet-owned company, working with TU Delft, published research on the system Wednesday in Nature Communications.
According to Waymo, the model should provide a more accurate picture of how attentive drivers respond to dangerous situations.
Based on “active inference”
The new framework is built around a theory known as active inference, which assumes drivers continuously predict possible outcomes and act to achieve the safest and most predictable result.
Waymo said the approach improves its understanding of how humans behave in crash scenarios encountered by autonomous vehicles.
“For decades, the automotive industry has used physical and virtual crash dummies to evaluate a car’s safety features, including its hardware and structural integrity,” the company wrote in a blog post.
Waymo said the new model “evolves this concept” by providing a behavioral benchmark for how “a careful and competent human driver responds to traffic conflicts.”
New “Reference Driver” model
Waymo calls the system the Reference Driver.
Unlike earlier models, which focused on last-second reactions, the new framework aims to replicate driver behavior leading up to a collision.
Previous systems, according to the company, mainly modeled “last-second, reactive” maneuvers.
The Reference Driver seeks to capture earlier decision-making and anticipation.
Simulating “surprise”
Arkady Zgonnikov, an assistant professor at TU Delft, said the model can simulate the internal “surprise” experienced by a driver during a conflict.
That capability provides “a more human-like benchmark for autonomous driving systems that was previously impossible to automate at scale,” he said.
Waymo said the model can be used to analyze:
- Collision avoidance.
- Other road user behaviors.
- Large test sets involving thousands of scenarios.
According to the company, the system can evaluate complex crashes in virtual environments more quickly and efficiently.
Increased scrutiny
The research comes as Waymo expands into additional cities and faces growing regulatory attention.
In January, one of the company’s robotaxis struck a child near a school in Santa Monica, California.
Using its previous benchmark, Waymo said an attentive human driver would likely have struck the child at roughly 14 mph.
The robotaxi slowed from 17 mph to 6 mph before impact, and the company said the child suffered minor injuries.
The incident remains under investigation by:
- The National Highway Traffic Safety Administration (NHTSA).
- The National Transportation Safety Board (NTSB).
Research code released
Waymo said it wants researchers to help improve the model.
The company is making the Reference Driver code available under an academic, non-commercial license.
Permitted uses include:
- Research.
- Teaching.
- Personal experimentation.
- Scientific publications.
TL;DR
Waymo has introduced a new “Reference Driver” model that it says provides a more realistic way to compare robotaxis with human drivers. Developed with TU Delft, the framework models how drivers anticipate and respond to traffic conflicts.
AI summary
- Waymo and TU Delft published research on a new driver model.
- The “Reference Driver” is based on active inference theory.
- It models driver behavior before crashes, not just last-second reactions.
- The work comes as Waymo expands and faces greater scrutiny.
- Waymo is releasing the research code for academic use.
