Waymo and TU Delft Publish the AI World's First Rigorous Human-Driver Benchmark for Robotaxis

Autonomous vehicle companies face a peculiar measurement problem. Their systems are routinely described as safer than human drivers — Waymo’s own data shows its vehicles cause 90% fewer serious crashes than humans over more than 127 million miles of autonomous operation. But the comparison has always been vulnerable to a simple objection: robotaxis operate in carefully mapped, bounded urban environments under favorable conditions. Human drivers operate everywhere, in every condition, navigating roads they’ve never seen before.

On June 10, Waymo took a significant step toward resolving this dispute on scientific terms. The company published research in Nature Communications, co-developed with transportation engineers at TU Delft in the Netherlands, introducing what it calls the “Reference Driver” — a computational model designed to simulate how a careful, competent human driver would behave in the same traffic scenarios that autonomous vehicles encounter.

What the Reference Driver Is

The model is built on a framework called active inference — a theory from computational neuroscience that describes decision-making as a continuous process of imagining possible futures, assigning probabilities to outcomes, and selecting actions that minimize “surprise” — the gap between expected and actual experience.

Applied to driving, this means the Reference Driver doesn’t just model what a human does in the last fraction of a second before a collision. It models the entire behavioral sequence leading up to a conflict: when a driver first perceives a hazard, how their speed and lane position change in response, and what options they evaluate before taking evasive action. The model can simulate a “careful and competent human driver” responding to traffic conflicts in a way that previous computational approaches simply could not.

“The model can simulate the internal ‘surprise’ a driver feels during a conflict, providing a more human-like benchmark for autonomous systems that was previously impossible to automate at scale,” one of the TU Delft researchers explained in documentation accompanying the paper.

The Problem It Solves

Previous human-driver baselines in autonomous vehicle safety research shared a common limitation: they were built from crash data, which records what humans did after they had already lost control of a situation. This produces a benchmark for human failure rather than human competence — a comparison that tells you very little about whether an autonomous vehicle is making good decisions in situations that haven’t yet turned catastrophic.

The Reference Driver inverts this. By modeling how a careful human driver responds from the moment a hazard enters their awareness, Waymo can now compare its system’s behavior not against the average of all human crashes, but against the upper bound of what careful driving looks like.

This has both scientific and regulatory value. Scientifically, it gives AV developers a richer signal about where their systems diverge from human norms — potentially revealing conservative over-caution as much as dangerous under-reaction. Regulatorily, it gives Waymo a defensible methodology for explaining its vehicles’ behavior in any specific incident. When regulators or plaintiffs ask why a Waymo vehicle took a particular action, the company can now point to a peer-reviewed model of what a careful human driver would have done in the same position.

Why This Matters Now

Waymo is in an expansion phase. The company is operating in San Francisco, Los Angeles, Phoenix, and Austin, with additional city launches in various stages of approval. Each new geography brings new traffic patterns, new edge cases, and new incidents that will require public explanation.

The regulatory environment for autonomous vehicles has grown significantly more demanding over the past two years. The National Highway Traffic Safety Administration now requires detailed incident reports from AV operators within 24 hours of any crash involving an injury or airbag deployment. State regulators in California and Texas have imposed additional conditions on commercial robotaxi licenses. The scrutiny is not going away.

Against this backdrop, the Reference Driver is a proactive move. Rather than waiting for regulators to demand a comparison methodology, Waymo is publishing one, open-sourcing the code under an academic non-commercial license, and establishing its preferred framework before others can impose one.

The choice to publish in Nature Communications — one of the highest-impact scientific journals in the world — is itself a statement. This is not a white paper from Waymo’s communications team. It is peer-reviewed research that will be tested, challenged, and potentially adopted by independent researchers and eventually regulators. Waymo is betting its methodology can survive that scrutiny.

The Open-Source Angle

By releasing the research code under an academic license, Waymo is inviting the broader autonomous vehicle research community to use the Reference Driver as a shared baseline. If the model is adopted — by academic researchers, by competing AV companies in their own validation, or eventually by regulatory agencies — it becomes a de facto industry standard rather than a Waymo-proprietary tool.

That outcome would serve Waymo’s long-term interests well. A Waymo-developed benchmark, validated by the scientific community and adopted by regulators, would reflect practices that Waymo already follows. Competitors who have invested less in safety modeling would face a higher bar.

The open-source strategy also reflects a growing industry recognition that the hardest problems in AV safety are not competitively sensitive. How you measure whether an autonomous vehicle drives carefully is not a trade secret. How you build the system that passes the measurement might be.

The 127-Million-Mile Context

Waymo’s safety data is the empirical foundation on which all of this research rests. With over 127 million fully autonomous miles through the end of 2025 — equivalent to more than 150 human driving lifetimes — the company has accumulated enough incident data to train and validate a reference model at scale. The 82% reduction in airbag deployment events and 81% reduction in injury incidents versus human drivers are not projections; they are observed outcomes from a dataset that few organizations in the world have access to.

The Reference Driver paper is, in one sense, Waymo translating those empirical results into a theoretical framework that others can use. The company is done proving to itself that its vehicles are safer. Now it is building the tools to prove it to everyone else.

For an industry that has promised safety for a decade and delivered it inconsistently, a peer-reviewed, open-source benchmark developed with an academic partner and published in a top journal is a meaningful development. Whether it ultimately reshapes regulatory practice — or simply becomes a useful research tool for AV engineers — the Reference Driver is the most scientifically serious attempt yet to answer the question the autonomous vehicle industry has always struggled with: safer than whom, and measured how?