The AI-500 Roadshow: 500 Cities and What We Learned

Earlier this year, we introduced the AI-500 Roadshow to test how well the Wayve AI Driver, derived from a single foundation model, generalizes to unfamiliar environments. Generalization is a core advantage of our AV2.0 approach and is essential for building driving intelligence that can scale without region-specific engineering. To assess this, we tested the AI Driver in more than 500 cities with safety operators behind the wheel and observed its behavior when deployed in locations that were often entirely unseen in training.

The roadshow took place on public roads across Europe, North America and Asia. It spanned a wide range of geographies, climates, road layouts and driving cultures, enabling us to observe system behavior in unfamiliar settings and identify the conditions in which a single global model can operate reliably.

Why real-world diversity matters

Reliable generalization requires exposure to varied environments. Differences in the operational design domain strengthen the model’s internal representations and improve its ability to interpret scenarios outside the distribution of the training dataset. This is central to enabling a single global AI Driver to operate without re-engineering the system for every new deployment region.

Our earlier market expansions to the US, Germany, and Japan demonstrated the feasibility of this approach. The AI-500 Roadshow extended that evaluation on a significantly larger scale.

Testing generalization on a global scale

We selected cities and towns that offered meaningful real-world complexity and enabled us to observe system behavior in zero-shot and few-shot settings, where the model had no or very little local data.

By December, we surpassed our 500-city target, operating the AI Driver software in 506 cities worldwide. Each location provided additional evidence of how the system handled unfamiliar environments.

In total, the roadshow covered 1.45 million kilometres with the AI Driver software engaged, providing a broad real-world dataset for evaluating generalization performance.

The key learning: reliable zero-shot generalization

The most important result from the AI-500 Roadshow is that the AI Driver software demonstrated reliable zero-shot operation in a significant portion of the cities tested. Across all deployments, the system operated with no prior local data (zero-shot) in 219 cities, representing 43% of the total. In addition, 67% of the cities visited had sparse data in the training dataset, defined as fewer than 100 kilometres of Wayve-collected or third-party driving data.

Together, these distributions indicate that a single global model can operate in many unfamiliar environments without region-specific engineering or extensive local data preparation. This is the clearest validated outcome of the roadshow.

Initial behavior and adaptation in new regions

Wayve AI Driver is powered by a single foundation model with one set of weights running globally. Instead of maintaining country-specific model variants, we provide two inputs, a country embedding and a driving-side embedding, to guide early behavior when the system is first deployed in a new region.

During initial operation in countries with no local data, the model may exhibit behaviours learned in other markets. These diminish as native data is collected and incorporated into training. Using domain priors helps stabilize early behavior and provides a consistent starting point across regions.

Data exposure across environments

The roadshow accumulated 46,347 hours of testing operation, the equivalent of 5.3 years of continuous driving. These hours spanned:

• dense urban centres such as Paris, New York and Tokyo
• smaller towns, mountain regions, and coastal roadways
• highways and autobahns
• varied traffic patterns and local signage
• distinct cultural driving norms

Each location surfaced different operational design domain characteristics, providing insight into how a single global model responds across a wide range of real-world conditions.

Implications for scalable deployment

Deploying driving intelligence across global markets requires a system that can operate effectively without region-specific engineering. The AI-500 Roadshow demonstrated that a single global model can reliably operate across many cities, even in places never seen before, validating the core premise of our foundation-model approach.

Generalization at this scale is essential for building AI Driver software that can be integrated by OEMs across markets with minimal adaptation. It establishes a practical path toward global deployment and informs the next phase of development and evaluation.

👉 Watch the final video from our drive from Tokyo to Mount Fuji, as we conclude the AI-500 roadshow.