On 10 June 2026, a research paper appeared in Nature Communications that represented something genuinely new in the autonomous vehicle industry — not a new safety milestone, not a record ride count, but a scientific tool for evaluating whether a robot drives the way a careful human would. Published jointly by Waymo and researchers at TU Delft University of Technology in the Netherlands, the Reference Driver — or ReD — model is the first computational benchmark designed to assess autonomous vehicle behaviour in collision-avoidance scenarios against a simulated competent human driver. For Australia, where the regulatory framework for autonomous vehicles remains under active development, its arrival is well worth understanding.
What the Reference Driver Model Does
For most of the autonomous vehicle industry’s short commercial history, safety has been measured primarily through outcome statistics — crashes per million miles, airbag deployments per distance driven, incidents per trip. Waymo’s own published safety data shows its vehicles involved in 92 per cent fewer serious injury crashes than average human drivers across comparable driving populations. Those are meaningful numbers, but they share a common limitation: they measure what went wrong, not whether the vehicle was making good decisions across the full range of situations it encountered.
The ReD model takes a different approach. It builds a computational simulation of how a careful, competent human driver responds to a developing collision risk — not at the last moment of impact, but from the moment a risk begins to emerge. That simulation then becomes a behavioural benchmark. An autonomous vehicle’s responses to the same scenarios can be compared against it to assess not just what the vehicle did, but how appropriately it reasoned about the situation. The debate about whether robotaxis are safe has until now rested almost entirely on statistical outcomes — the ReD model adds a second, complementary dimension rooted in behaviour.
Active Inference — The Neuroscience Behind the Model
The theoretical foundation of the ReD model is active inference, a framework drawn from computational neuroscience that describes how the human brain continuously minimises uncertainty by updating its understanding of the world as new information arrives. In driving terms, active inference captures the cognitive process that precedes any physical action: a driver who notices a vehicle pulling unexpectedly into their lane does not simply brake — they update their mental model of the situation, estimate how it might evolve, consider the probable intentions of other road users and select the action most likely to produce a safe outcome. That loop of belief updating and action selection runs continuously, not just at moments of crisis.
Arkady Zgonnikov of TU Delft, one of the paper’s lead authors, described the ReD model as providing “a holistic representation of human collision response” — capturing the cognitive and probabilistic reasoning that precedes any physical evasive action, not just the mechanics of braking or swerving. Karl Friston, the neuroscientist whose foundational work on active inference underpins the model’s architecture, called the paper “a remarkable piece of work — a tour de force in terms of generative modelling.”
Critically, the model is fully automated: it does not rely on manually coded rules to define what a good response looks like. That automation makes it scalable to thousands of distinct scenarios in virtual environments and reproducible, which is essential for any benchmark intended to support regulatory or comparative evaluation across different AV systems. The underlying research is directly connected to the sensors and AI systems that autonomous vehicles already use to perceive and interpret their environment.
A Shared Benchmark for the Industry
One of the most significant aspects of the ReD model is its intended scope. Waymo has released the research code under an academic non-commercial licence, explicitly making it available to other researchers, companies and regulatory bodies. Mauricio Pena, Waymo’s Chief Safety Officer, described the goal as helping “the industry move toward a shared, scientifically grounded approach” to evaluating autonomous vehicle safety — framing the model as a contribution to the field rather than a proprietary competitive tool.
Publication in Nature Communications, a peer-reviewed scientific journal with global reach, places the model within the independent scientific literature rather than an industry white paper. That distinction matters: it means the benchmark’s methodology has been assessed by independent researchers, its assumptions are documented and contestable, and it can be adopted, tested and built upon by the broader scientific and regulatory community.
The autonomous vehicle industry has long faced criticism for the absence of a consistent, comparable safety metric that would allow meaningful comparison across different systems and developers. The most comprehensive published safety statistics available still represent the output of a single company using its own methodology. A shared behavioural benchmark, grounded in peer-reviewed neuroscience rather than ad hoc engineering convention, is a meaningful step toward a more transparent and standardised evaluation framework.
The Relevance for Australian Autonomous Vehicle Regulation
Australia’s National Transport Commission is developing the Automated Vehicle Safety Law framework that will govern how autonomous vehicles operate on Australian roads. That framework focuses on accountability structures — who is responsible when an automated system is engaged and what reporting obligations apply when something goes wrong. As commercial AV services mature in the United States and United Kingdom, Australian regulators face an increasingly specific version of a question that has no straightforward answer: how safe is safe enough, and how do we measure it?
The ReD model offers a language for answering that question. Rather than waiting for the statistical volume that US markets have been able to generate — Waymo has now accumulated more than 170 million autonomous miles across its commercial operations — Australian regulators could evaluate AV systems against behavioural benchmarks in controlled scenarios, comparing response profiles against the competent human driver standard that ReD defines. This kind of simulation-based evaluation complements rather than replaces real-world data, and could allow regulators to assess systems at an earlier stage of deployment than purely statistical approaches permit.
The timeline for autonomous vehicles on Australian roads depends in part on how quickly regulators can build confidence in the evidence base. A scientific benchmark that is reproducible, peer-reviewed and shared across the industry accelerates that process in ways that proprietary safety reports cannot.
What It Means for Public Trust in Australia
Public trust in autonomous vehicles is one of the primary barriers to their broad adoption in Australia. Surveys of Australian attitudes to self-driving technology have consistently shown significant caution among the general population, and that caution is unlikely to dissolve solely on the strength of statistical safety arguments, however compelling the underlying numbers.
One persistent challenge is a mismatch between the statistical record — which, in Waymo’s case, shows dramatically fewer serious injuries per mile than human drivers — and the intuitive human tendency to apply different standards to machine-caused harm than to harm caused by human error. A benchmark that explains not just what happened but how the robot decides — and demonstrates that its decision-making profile matches the cognitive process of a careful human driver — is a different kind of safety argument. It connects statistical performance to the underlying reasoning process in a way that is more accessible to non-specialist audiences and may do more to address the questions Australians actually ask about autonomous vehicles than any mile-count headline.
The Bigger Picture
The ReD model is not a vehicle, a service or a regulatory approval. But it represents a stage of methodological maturity that the autonomous driving field has been working toward — the kind of foundational scientific infrastructure that precedes broad public and regulatory confidence in any safety-critical technology, whether aviation, pharmaceutical approval or structural engineering. As Waymo expands its commercial operations toward London and beyond, and as operators begin to think seriously about markets in the Asia-Pacific region, having a shared and scientifically grounded benchmark for what good robot driving looks like is valuable infrastructure for every participant in the conversation — including Australian regulators, transport planners and the future riders who will ultimately decide how much trust they are willing to extend to a driverless vehicle.
Sources
- Waymo — Introducing the Reference Driver: A New Benchmark for AV Safety (waymo.com)
- Waymo Safety — Published Safety Performance Data (waymo.com)
- National Transport Commission — Automated Vehicle Program (ntc.gov.au)
- Nature Communications — Reference Driver: A Computational Benchmark for Autonomous Vehicle Safety (nature.com)
- Featured image: Melbourne CBD skyline at blue hour via Pexels
