Four Months to Drive Tokyo: How Wayve Bet the Company on One End-to-End Network

Pat Grady opens by asking Alex Kendall to define AV 1.0 versus AV 2.0, the shift Wayve has been arguing since 2017. Kendall’s seed-deck pitch was that classical robotics broke autonomy into perception, planning, mapping, and control, then hand-engineered each piece and leaned on infrastructure such as high-definition maps. Wayve’s bet was the opposite: intelligent machines that make their own decisions onboard, built the only way he thinks you build an AI system, with end-to-end deep learning. Ten years later he still describes AV 2.0 as replacing that stack with one neural network, now talked about as a foundation model that can handle many cars, sensor layouts, and use cases without onerous infrastructure.

Grady boils it down: sensor in, motion out, a gigantic net in the middle. Kendall laughs and says that is right at a simple level, then names two problems language models do not have in the same way. The first is safety. The system has to be safe by design. “We can’t just pump more data in and hope that hallucinations go away.” The architecture still has to be end-to-end and data-driven, but it also has to be functionally safe and support a robust behavioral safety case. The second is compute: the model has to run in real time on a vehicle, inside the hard limits of onboard chips and sensors. The narrative, though, is the one already playing out in language and gameplay: learned end-to-end systems beat hand-coded ones.

In 2017 the industry finished that sentence with “it’ll never work because…” Kendall says he could count hundreds of those meetings. Typical lines: it is not safe, it is not interpretable, you cannot understand what it is doing, or simply that they had not heard of this AI thing. Five or ten years ago, he grants, the interpretability complaint was fair. Today there are tools for seeing how deep-learning systems reason. Deeper than the tools is a philosophical objection: if your ambition is an intelligent machine, it is naive to expect a single line of code you can point to as the cause of an outcome. The beauty of these systems is that they are wonderfully complex, and the way to design and understand them is data-driven. He is equally unimpressed by hybrids that glue a rules stack to an end-to-end stack. Those designs, he says, often get the worst of both worlds, plus cost and complexity. Sonya Huang asks whether Tesla’s FSD 12 flipped the consensus. Kendall credits large language models for making the world curious, plus Wayve’s own jump from central London onto highways and into Europe, Japan, and North America, with cars in New York the week before. The morning of the taping he was still in a meeting with people who wanted “end-to-end AI” plus hard constraints. The market, he says, is still split between teams leaning in and teams catching up.

Huang asks why Wayve can launch in hundreds of cities while AV 1.0 companies still go out and build an HD map. Kendall’s one-word answer is generalization: reasoning about something you have never seen. Every drive produces a first. That morning it was a road worker rolling something out on a pedestrian crossing without stepping into the lane, and the car had to decide whether it could pass without yielding. You will never see every experience in training data, so a system that cannot generalize is neither safe nor useful around the world. That requirement, he says, motivated the entire approach.

Nissan in Tokyo is the exhibit. In September, Wayve unveiled a vehicle to media. Four months earlier was the first time the team had driven in Tokyo or gotten hands on that car. A new country, a new vehicle, a new sensor distribution, then a media ride-along. The model held up because it had been trained on diverse data, diverse sensors, and complex urban driving, including the mess of central Tokyo. If you are building a vertically integrated robot you can go deep on one stack. Wayve’s ambition is different: to be the embodied-AI foundation model for the best fleets and manufacturers in the world. Unless they want a separate network for every application, they have to amortize cost over one large intelligence and adapt quickly to whatever the customer cares about.

Later he splits driving into three scores. Safety is safety-critical behavior. Flow is whether the drive is smooth and enjoyable. Utility is navigation and road semantics: signs, language, driving culture. Safety and flow, he says, generalize almost uniformly across countries. Utility is the hard one. Going from the UK to the US took hundreds of hours of data to get within 10 percent of their frontier. Germany was cheaper on some axes because the model had already learned to drive on the right in America, but it had to unlearn right-on-red and learn the autobahn, where they drive up to 140 km/h. Each new market, in his telling, takes exponentially less data because some of the structure has been seen before. This year the cars are in 500 cities. Corner cases at that scale are rarely a clean category; they are confounding factors stacked on each other, plus a road sign in a language the model has not read yet.

Huang asks whether reasoning is deliberately trained or an emergent behavior, the way inference-time compute became its own track in language models. In the physical world, Kendall says, reasoning is well expressed as a world model, and Wayve has been training for it on purpose. In 2018 they put a first version on the road: a 100,000-parameter network that could simulate a 30-by-3-pixel strip of the road ahead, used as an internal simulator for model-based reinforcement learning. Today that line of work is GAIA, a generative world model that can simulate multiple cameras and sensors, be controlled, and be prompted to change agents in the scene. Train the system to simulate what happens next, then put that representation on the road, and you get behavior you did not have to script.

The ride that morning offered examples. On an occluded unprotected turn, the car nudged forward until it could see, then completed the turn. In London fog it slows down and drives only as far as it can reason about. Those habits, Kendall says, grow out of training the model to predict what happens next, especially in messy multi-agent scenes. Huang asks whether world models are the key to teaching new-scenario reasoning. “100%,” he says. They are also a data-efficiency tool. In the limit, stacking real miles might produce generic level-five driving, but that is a gigantic lookup table, neither economic nor technically feasible. Efficiency comes from the data curriculum, from learning algorithms, and from world models that recombine real experience rather than replace it. Synthetic data, in his framing, magnifies rare scenes so they can be practiced; it is not a substitute for the real world.

The data problem is unlike language. A typical self-driving car has a dozen multi-megapixel cameras, plus radar and maybe LiDAR; aggregated, that is tens or hundreds of petabytes. Diversity matters more than the raw pile. Wayve’s two moves are to become a trusted partner and pull data from dashcams, fleets, manufacturers, and robot operators, then to use unsupervised clustering to find anomalies and the scenarios where the system is weak, and to drive the learning curriculum from those. Fifty model developers work on one main production model. The company’s data infrastructure, simulation, and pre-road safety licensing were built from scratch around end-to-end deep learning, not as a hedge. Under resource constraints, Kendall says, everything is about iteration speed.

Wayve began as passionate embodied-AI researchers. In the last couple of years Kendall says he has grown to respect the automotive culture they now sell into. Building millions of cars demands a level of reliability and mean time to failure that is extraordinary. What he has learned from OEMs is efficiency, quality, and the gap between a technology and a product, which is reliability. They also care about brand: a car’s driving personality should match the badge. From human driving data, Wayve can draw a clean line between “helpfully assertive” and “unhelpfully aggressive,” and sit a model on the right side of it. A sports brand, in that logic, should drive like a sports brand. All fleets share one model family that is retrained at very large scale and iterated monthly; specializing to a sensor set or an embedded chip is a thin layer on top. More than 99 percent of the cost and time sits in the base model. That is how they serve more than one manufacturer at once.

The go-to-market has settled on partnering with consumer automotive OEMs rather than retrofitting robotaxis city by city. Software-defined vehicles now ship with GPUs, surround cameras, surround radar, and sometimes a front LiDAR, which is the infrastructure autonomy needs. About 90 million cars are built each year. Tesla builds a couple of million with its own stack. The rest, on the order of 88 million, is the partnership market. Manufacturers, Kendall says, do not want to stop at driver assistance; they want eyes-off systems that can take liability and give the user time back. Native software integration, not aftermarket hardware, is the path to homologation worldwide and to tens or hundreds of thousands of affordable robotaxis. People still get in skeptical. Without exception, he claims, they get out smiling. He wants autonomy to stop being a robotaxi tourism product and become something people use in every city.

The yearly Twitter fight over cameras versus radar and LiDAR is, for him, the wrong debate and not the frontier question. Outside Tesla, he says the industry has coalesced on surround cameras, surround radar, and a front-facing LiDAR. That automotive-grade kit costs under $2,000 and, with a frontier GPU, is a solid platform for L3 and L4 eyes-off driving. It gives redundancy. Camera-only, he agrees, can get you to human level, and because 95 percent or more of accidents are human error, eliminating distraction is already a huge step. The goal is to go beyond human level and kill the remaining long tail, which he thinks needs perception that vision alone does not provide. Wayve trains on all the permutations: camera-only, camera plus radar, camera plus LiDAR. The car they had just driven was camera-only. Other partner vehicles carry the extra sensors. The point is native integration into the OEM’s own car, not a retrofit hat.

In 2021 Kendall’s team asked to start a language project. His first reaction was no: a startup is about focus. They talked him into it, and a year or so later Wayve released Lingo, which he calls the first vision-language-action model in autonomous driving. It can see, drive, and converse. You can ask what looks risky, what happens next, or let it commentate the ride. He names three benefits. Language in pre-training makes the representation stronger than imagery alone. Aligning with language unlocks product features, a chauffeur you talk to instead of a system that requires a robotics PhD, including telling it to hustle through a commute. Third, and easy to miss, language is an introspection tool. Regulators and engineers can ask the model why it did something and get an explanation of its reasoning, a direct answer to the old “uninterpretable” attack. Demos still run offboard. Next-generation development vehicles with Nvidia Thor, he says, will have enough compute to run it on the car.

Zoom out from cars and Kendall’s claim is that mobility will become possible well before manipulation. Manipulation is still hard: data access, hardware supply chains, even tactile sensing. He puts that stack at roughly the maturity of self-driving in 2015. Wayve’s system, meanwhile, is becoming a general-purpose navigation agent. Give it an arbitrary sensor view and a goal and it produces a safe trajectory, useful for passenger cars, robotaxis, trucking, and any mobility robot a fleet wants to build. Automotive, he notes, has the largest robot and data supply chain, which is luck they intend to spend. Working a narrow application with expensive infrastructure, Phoenix-style, has limited runway. The recipe for scale is general-purpose, lean, low-cost hardware and a system that is as intelligent and robust as you can make it. Humanoids excite him; so do other forms of locomotion. The same story, he thinks, will play out there.

Asked what research is still needed for physical AGI, Kendall does not pretend scaling current methods is enough, even though there is plenty of room to scale. He talks about four factors that all have to be pushed: data, compute, algorithmic capability, and embodiment, meaning the robot’s hardware. On algorithms he cares most about measurement. How do you quantify the system, catch regressions fast, and build a simulator that closes the real-world gap at scale and still runs efficiently? Generative world models, he notes, are no secret and very compute-hungry. People say it is a chicken-and-egg: if you have a perfect simulator you have solved self-driving, and vice versa. He believes that. AlphaGo showed that with a perfect simulator you can solve a problem with search. He wants the same in robotics. Other pillars are more generality across modalities, especially for human-robot interaction, and unglamorous engineering efficiency in training and serving. Getting that infrastructure right, he says, is a real competitive advantage or disadvantage.

Grady asks, deadpan, what AV 3.0 might look like in five, ten, or fifteen years. AV 2.0’s premise was intelligence on the car, without a ton of infrastructure and overcooked hardware. If Kendall speculates, one idea is taking intelligence *outside* the car. Once autonomous vehicles are the majority, they can communicate. Why keep traffic lights if they can coordinate? Why carry every sensor if you can borrow the view from the AV in front and see around a corner? He flags cybersecurity and latency as new problems and admits the company has not thought it through in depth. Huang riffs that humans might be boxed out of the mesh, or kept to recreational tracks. Kendall’s hiring pitch is more immediate. The team is the product. Wayve, with sites in London, Stuttgart, Tel Aviv, Vancouver, Tokyo, and Silicon Valley, is trying to be a frontier AI lab with a near-term automotive product, the ChatGPT-scale impact for robotics. Anyone who wants to pioneer embodied AI and actually ship it, he says, should come talk.