Crossing the Pond and Beyond: Generalizable AI Driving for Global Deployment

Wayve’s expansion to the US is more than just a geographic milestone—it is a key demonstration of our AV2.0 approach to autonomous driving. Powered by a foundation model trained on globally diverse data, our embodied AI adapts to new geographies and vehicle platforms with minimal additional data, making it far more scalable than conventional self-driving systems.

This blog presents real-world results on how Wayve’s foundation model generalizes across new geographies and vehicle platforms. Specifically, we present findings on:

• Rapid adaptation in the US: Our model successfully adapted to driving on the right side of the road with 500 hours’ worth of incremental US-specific data collected over 8 weeks.
• Learning country-specific driving behaviors: Our AI quickly adapted to new road signs, traffic flow patterns, and intersection rules, showing strong improvements with 100 hours’ worth of additional data. 
• Strong zero-shot performance in Germany: The new model performed 3X better in Germany than how the initial deployment performed in the US, demonstrating that generalizability improved after diverse market exposure.
• Seamless transition between vehicle platforms: With 100 hours’ worth of new vehicle-specific data, the model adjusted to a different platform.

Figure 1. Illustrates our model’s adaptation to US driving, reaching UK-level performance parity after training on 500 hours’ worth of new US-specific data.

This rapid learning highlights how our foundation model generalizes to new geographies. As a result, we can derive a single driving model capable of deploying in both the UK and US. The videos below show the same AI model driving our vehicles in London and the Bay Area.

Learning New Driving Behaviors

Beyond overall driving performance, we also examined how rapidly our model learned new US-specific driving behaviors, such as 4-way stops, right turns on red, unprotected left turns, and freeway merging on short on-ramps. These behaviors reflect key differences in road infrastructure and traffic patterns between countries. For example, stop signs are rare in the UK, and unprotected left turns follow the opposite traffic flow. By fine-tuning our model with incremental US-specific data, it adapted not only to differences in road infrastructure but also to implicit norms like local driving culture, which are often too nuanced to program manually.

We tested how much additional sampled data was needed for the model to learn these new behavioral competencies using virtual ‘offroad evaluation.’ As new training data volume increased from 10 to 500 hours, offroad performance improved, as indicated by higher scenario pass rates in Figure 2. Notably, the model showed strong improvements within 100 hours’ worth of additional sampled data, with continued improvement up to 500 hours.

Figure 2. Illustrates how new, learned behavioral competencies improved rapidly with new domain-specific data.

What Sets Wayve Apart: Extracting Value from Many Sources of Data

Wayve’s foundation model uniquely learns from a broad spectrum of unlabelled data. This enables us to move beyond the limitations of solely relying on sensor-rich AV data, which is expensive to collect and thus scarcer. Instead, we augment our internal fleet data with readily available third-party datasets from fleet partners, automakers, and other sources of lower-fidelity driving videos. This flexibility allows us to build a global “data ocean” without collecting it all ourselves, exponentially increasing our data coverage. 

Crucially, this diverse data corpus fuels a continuous improvement flywheel. Each new dataset from every new domain strengthens our foundation model’s performance across all driving scenarios, creating a powerful network effect. As the model builds on prior experience (i.e. data it has already trained on), it learns faster in new situations and generalizes more efficiently. 

To illustrate this, consider the following results: adding data from two markets (US and UK) improves model accuracy, as shown by the lower loss curve of the orange line, indicating better model accuracy. Notably, this improvement extends to both regions and a newly introduced third market. These offline results predicted a performance boost in Germany before we even opened our new testing and development hub there last week!

Figure 3. Models trained with a mix of UK and US data (orange line) show consistently lower loss curves, indicating improved model accuracy across all regions, including previously unseen environments like Germany. 

We saw this prediction also play out in real-world testing. On-road testing in the UK showed a similar network effect: adding 10,000 hours’ worth of data from UK, US, and Germany to our foundation model resulted in a 3x performance increase compared to adding the same volume of UK-only data.

Figure 4. Compares the relative improvement in model on-road performance when training with more geographically diverse data.

Early Results from Germany: Strong Zero-Shot Performance

As we begin testing in our third market, Germany, early results show a considerable leap in zero-shot performance. Compared to our initial deployment in the US, our model performed 3X better in Germany from the start without any fine-tuning (Figure 5). 

Figure 5. Compares the model’s zero-shot performance when we went from Market 1 to 2 (UK to US zero-shot) and then from Market 2 to 3 (UK and US to Germany zero-shot) without additional market-specific training data.

These strong zero-shot results pave the way for the next phase of learning. Below, you can see footage of the same model—already driving in the UK and US—now navigating German roads. We’re excited to refine its performance further by training on new German-specific data, including snowy conditions and high-speed Autobahns.

Beyond Geographies: Generalization Across Vehicle Platforms

Adapting to new markets is only one dimension of generalization. We tested our foundation model’s ability to adapt to new vehicle platforms with different sensor configurations—an essential step for working with various automakers. In transitioning from our development vehicle to a new automotive platform, we found that 100 hours of vehicle-specific data led to an 8X performance improvement. This aligns with our geographic adaptation findings, suggesting that only incremental training may be needed for cross-platform deployment.

Figure 6. Illustrates our model’s generalization to a new vehicle platform after training on 100 hours’ worth of new vehicle-specific data.

From Generalization to Automotive-Grade AI

While our embodied Al’s ability to rapidly generalize across geographies and vehicles is a key breakthrough, developing a safe, automotive-grade product requires rigorous validation across a wide range of driving scenarios. To do this efficiently and accelerate deployment, we take a data-driven approach—leveraging our AI science innovations in synthetic data generation and neural rendering, like GAIA, PRISM, and Ghost Gym. These state-of-the-art techniques create high-quality, photorealistic simulations, allowing us to extensively test our AI against rare but safety-critical edge cases that would take years to encounter in the real world. By combining virtual and real-world testing, we ensure our AI is not just adaptable, but provably safe and ready for global deployment.

Scaling AV2.0 Globally: A Roadmap for the Future

With these robust validation capabilities in place, we are scaling AV2.0 globally. Our successful expansion from the UK to the US demonstrates how our AI adapts across markets and vehicle platforms, paving the way for broader deployment. 

As we expand, our goal remains the same: to develop embodied AI that can operate any vehicle, anywhere. By leveraging diverse data, we are building a highly generalizable AI that enables safe, adaptable, and scalable autonomy. The network effect of data is accelerating this progress—enhancing model performance, driving faster adaptation, and enabling broader generalization. 

Our expansion is only getting started. Stay tuned as we continue to advance AV2.0 and make embodied AI a reality for advanced assisted and automated driving.