Wayve: the end-to-end driving bet, and what it has actually shown

Wayve is a London company building what its technology page calls a "robot brain" for cars: a single neural network trained on driving data that turns raw sensor input into steering and speed, with no HD map underneath and no modular perception-plan-control stack. It calls this AV2.0, in contrast to the AV1.0 "sense-plan-act" architecture used by most of the industry. The company is registered as Wayve Technologies Ltd (England No. 10924127) at 230-238 York Way, London N7 9AG, and its site reports $1.5 billion secured to deploy its autonomy platform, with Alex Kendall as CEO.

The interesting thing about Wayve for a physical AI reader is that its thesis is testable in a way most robotics theses are not. If one model with one set of weights can drive in a city it has never seen, that is a claim you can put a number on. Wayve has run exactly that experiment and published the scale of it. What it has not published is the outcome distribution: how often the safety operator took over. Almost everything below comes from pages on wayve.ai, and every claim is flagged as marketing language, a self-reported result, or a genuinely peer-reviewable artifact.

Companion reading: self-driving cars and autonomous vehicles, foundation models and VLA for robotics, robot simulation and digital twins, sim-to-real transfer.

Table of contents

• Key takeaways
• The architecture bet
• The product portfolio
• GAIA: the world model line
• Demonstrated versus claimed
• What is open, and what is not
• What to watch
• Frequently asked questions
• Changelog

The architecture bet

Wayve describes AV2.0 as replacing the modular pipeline with "a single neural network trained on diverse data to convert raw sensor inputs into safe driving outputs" (marketing language, from the technology page). Four advantages are claimed on that same page. All four are Wayve's own marketing bullets: it eliminates labeled data by learning from raw unlabeled data with self-supervised learning, it allows a lean sensor suite that OEMs can choose from, it is mapless ("AV2.0 doesn't rely on HD maps"), and it is vehicle agnostic from passenger cars to delivery vans. None of the four carries a published number.

The sensor position is more hedged than the popular framing suggests. The AI Driver product page has a "Camera-first Sensing" heading with automotive radar as a cost-effective complement, then adds that "Additional sensors, such as LiDAR, can be integrated as needed by the OEM." Lidar is left optional per customer. Wayve also uses lidar itself, as ground truth: GAIA-3's validation aligns lidar point clouds from original recordings against generated frames to check for misalignment. The bet is about the cost and labor of maps and region-specific logic, which the multi-country page calls "costly and labor-intensive to expand."

The research arm, Wayve Labs, lists six focus areas: representation learning, spatio-physical intelligence, decision-making, learning systems, cross-embodiment learning, and world and reward modeling. Cross-embodiment is described as using "the data and systems we built for driving to accelerate deployment to diverse robotic platforms across mobility and manipulation," which is the company signalling that it sees driving as one instance of a broader embodied AI programme.

The product portfolio

The Wayve AI Driver is a tiered portfolio. L1 is "Wayve AI Sense," L2 is "AI Driver Assist" (eyes-on, hands-off, "designed for the world but will be first launched in Europe, the US, and Japan"), L3 is "AI Driver CoPilot" (eyes-off, hands-off), and L4 is "AI Driver," a turnkey tier described as optimized for goods delivery and robotaxis. The product page itself gives the tiers no operational design domain document, no launch date and no fleet size. Those details appear only in press releases, which announce intent.

Named external relationships, with the dates Wayve's own press index gives them:

• Uber. "Wayve autonomous rides, available on Uber soon," with a London interest list. Wayve and Uber announced L4 autonomy trials in the UK on 10 June 2025, and on 5 August 2026 announced that Transport for London had granted Private Hire Vehicle licences to Wayve's autonomous vehicles, with early rider access planned for later that summer.
• Nissan. On 10 April 2025 Nissan announced it would launch its next-generation ProPILOT technology featuring Wayve AI Driver software, targeted at FY2027. Nissan and Wayve signed definitive agreements on 10 December 2025.
• Uber, Nissan and Wayve jointly. A memorandum of understanding announced 12 March 2026 targets a robotaxi pilot deployment in Tokyo by late 2026 using the Nissan LEAF. A robotaxi prototype built on NVIDIA DRIVE Hyperion was exhibited at NVIDIA GTC 2026 on 16 March 2026.
• Stellantis. A strategic technology partnership announced 21 May 2026 to integrate the Wayve AI Driver into the STLA AutoDrive platform, followed on 17 June 2026 by a Stellantis, Wayve and Uber partnership to explore global L4 robotaxis.
• Qualcomm. Announced 10 March 2026, introducing the Wayve AI Driver as a software option for customers using Qualcomm's Snapdragon Ride Platform.
• AMD, Arm and Qualcomm. A $60 million Series D extension announced 15 April 2026.
• NVIDIA. Wayve's Gen 3 robot platform will be built on NVIDIA DRIVE AGX Thor, announced 19 September 2025.
• Microsoft Azure. Supercomputing collaboration for foundation model training.
• S.RIDE. A Tokyo taxi app partnered with Wayve on 12 May 2025 to collect public road data.
• UK Government. A partnership signed 12 May 2026 ahead of UK passenger service trials.

Wayve also states it helped shape a UNECE-adopted regulatory framework for assisted and automated driving (10 February 2026), and that GAIA-3 validation runs through DriveSafeSim, a UK government grant-funded project.

GAIA: the world model line

GAIA is Wayve's generative world model family, and it is where the company publishes real numbers. Wayve's own blog headlines GAIA-1 (2023) as a "9-billion parameter generative world model"; the self-hosted GAIA-1 paper breaks that down as a 6.5B-parameter world model plus a 0.3B-parameter image tokenizer. GAIA-2 (March 2025) switched to latent diffusion over whole sequences, conditioning on ego action (speed, steering curvature), 3D bounding boxes, weather, time of day, and road attributes, with latent conditioning from CLIP embeddings and a proprietary model optimised for driving. It trained on UK, US, and Germany data with specific geographic regions held out for validation.

GAIA-3 (2 December 2025) is stated as a 15-billion-parameter latent diffusion world model, trained with five times more compute than GAIA-2 and roughly ten times more data spanning nine countries across three continents, with a video tokenizer twice the size of GAIA-2's. It generates NCAP-style tests such as CCFTAP and CCRS virtually. Wayve says correlation studies between GAIA-3's synthetic interventions and on-road experiments indicate the model can reliably predict relative policy performance, and gives no correlation coefficient.

GAIA-4 (3 August 2026) adds radar to the simulation alongside cameras, which Wayve calls an industry first for autonomous driving simulators. Its headline mechanism is "world-on-rails": Wayve describes the ego vehicle's behavior changing while every other agent replays the exact behavior from the real-world log, with "reactive agents" as the relaxation. Wayve reports one number for it, self-measured: training GAIA for this task "improves how faithfully it preserves the recorded world by 2.5x." The GAIA-4 post also contains the sharpest statement of Wayve's own problem: "end-to-end driving models are not separable. There is no intermediate perception output to score and no planning module to test in isolation," and "as a model's performance improves, the miles needed to validate safety sharply increase." That is an honest framing of why the company needs the simulator at all.

Demonstrated versus claimed

The AI-500 Roadshow (published 17 December 2025) is the substantive demonstration. From June to December 2025 Wayve ran the AI Driver software in 506 cities across Europe, North America, and Asia, covering 1.45 million kilometres with the software engaged over 46,347 hours of testing operation. In 219 cities, 43% of the total, the system operated with no prior local data at all. In 67% of cities the training set held fewer than 100 kilometres of Wayve-collected or third-party driving data. It ran on a single foundation model with one set of weights globally, configured only by a country embedding and a driving-side embedding. Wayve states plainly that this ran with safety operators behind the wheel.

That is a real result for the generalization claim. On the safety claim it is silent. No intervention rate, no disengagement rate, no miles per intervention, no contact or collision statistics, no human baseline. 1.45 million kilometres were driven and the outcome distribution is not published. For a reader trying to compare Wayve with any mapped or lidar-based stack, this is the gap that matters most.

The cross-geography adaptation post (10 March 2025) has the same shape. UK to US transfer: 500 hours of incremental US data collected over eight weeks brought the model near its UK benchmark, described as a 40X improvement, with a fivefold improvement at 100 hours. In Germany, Wayve reports the model performing 3X better from the start than the initial US deployment had, with no fine-tuning. Adding 10,000 hours from UK, US, and Germany gave a 3x performance increase over the same volume of UK-only data. A new vehicle platform needed 100 hours for an 8X improvement. Every one of these is a ratio against a base metric Wayve never names, measured by an internal virtual "offroad evaluation" reported as scenario pass rates. The relative trends are informative, and the absolute level is unknowable from outside.

One result is placed next to public baselines, though Wayve is the party doing the placing. LA-Pose (5 May 2026) is a self-supervised camera pose estimator trained on 10.2 million unlabelled driving clips using an inverse-dynamics latent action, with a post-training head fitted on a small amount of labeled data from Waymo, nuScenes, and Argoverse. Wayve claims over 10% higher pose accuracy than state-of-the-art feed-forward methods with orders of magnitude less labeled data, and says that on PandaSet, which the model never saw in training, "it still outperforms all baselines." Those benchmark numbers are self-published on wayve.ai with no third-party verification. An ablation reports a 50-dimensional latent bottleneck beating a 1,536-dimensional one. The task is camera pose estimation, a perception sub-task, and it is the one place where Wayve puts itself next to public baselines at all.

Wayve is also candid about specific failures. LA-Pose "currently degrades for reversing motion" because backing up is rare in the post-training data. The AI-500 post warns that in countries with no local data "the model may exhibit behaviours learned in other markets." The GAIA-2 technical report's future work section says the model "occasionally produces temporal or semantic inconsistencies, particularly in long-horizon or complex scenarios," and that "real-time or near-real-time video synthesis remains computationally intensive." GAIA-3 says work continues on "efficiency and real-time generation," which means GAIA-3 is not real-time either. These are limitations Wayve chose to disclose, and they are the company's own account of its own systems.

What is open, and what is not

One artifact is genuinely open: WayveScenes101, a novel-view-synthesis benchmark of 101 driving scenarios of 20 seconds each, 101,000 images from five time-synchronised cameras at 1920x1080, recorded in the US and UK. It ships COLMAP camera poses and reconstructed 3D points, holds out the central forward-facing camera for off-axis evaluation, carries weather, road type, and time-of-day metadata, works with the NerfStudio framework, and is reached through a GitHub repository that carries the download instructions. The licence is non-commercial, research purposes only.

Beyond that, 17 self-hosted PDFs sit on the science page, running from Learning to Drive in a Day (2018) through FIERY (2021), Driving with LLMs and LingoQA (2023), CarLLaVA (2024), and SimLingo and GAIA-2 (2025). LA-Pose points to la-pose.github.io. No model weights, no inference code, no training code, and no commercial licence terms appear anywhere. GAIA-1 through GAIA-4, LINGO-1 and LINGO-2, PRISM-1, Rig3R, and the AI Driver are all closed. Every headline figure in the preceding section is self-published and self-measured, with no reproduction path.

What to watch

A safety number. Any per-kilometre intervention or disengagement figure from the roadshow fleet, or a quantitative safety case under the referenced Safety 2.0 framework, would change the assessment more than another order-of-magnitude scale claim.

Absolute metrics replacing multipliers. The moment Wayve names the base metric behind the 40X and 8X figures, external comparison becomes possible.

GAIA-4 fidelity values. The post defines outcome fidelity, closed-loop fidelity, and component fidelity as its validation scheme and publishes no value for any of the three. The one figure it does give, a 2.5x improvement in preserving the recorded world, measures the generator's training, not the simulator's accuracy against on-road outcomes.

A fare-carrying L4 service. London has PHV licences and Tokyo has a stated late-2026 pilot target with the Nissan LEAF. A first paying rider, a published ODD, and a fleet size would move L4 from an announcement to a product.

Third-party evaluation. A leaderboard placement, an independent audit, or a head-to-head against a mapped lidar stack on a shared route. Wayve's thesis invites this comparison and no published version of it exists.

Per-vehicle compute. "Lean and adaptable compute requirements" with no TOPS, wattage, or bill-of-materials figure, despite cost being the core argument against maps and lidar.

Frequently asked questions

Does Wayve refuse to use lidar? No. The AI Driver is camera-first with radar complementing it, and the product page states that lidar can be integrated as needed by the OEM. Wayve also uses lidar point clouds as ground truth when validating GAIA-3's generated frames. Nissan's April 2025 FY2027 announcement pairs the Wayve AI Driver with Nissan's own next-generation lidar.

What is the single strongest published result? The AI-500 Roadshow: 506 cities, 1.45 million kilometres, 46,347 hours, one set of weights globally, 219 cities with no prior local data. It demonstrates geographic reach and configuration simplicity. It reports nothing about how well the system drove.

Has Wayve published a disengagement rate? Not on any page reviewed here. No intervention rate, miles per intervention, collision statistics, or human baseline appears anywhere on the site.

What can I actually download? WayveScenes101, under a non-commercial research licence, plus the self-hosted papers. No weights, no inference or training code, no eval harness.

How large is the driving model? Undisclosed. The 15-billion-parameter figure belongs to GAIA-3, a world model used for simulation. The AI Driver's parameter count, latency, compute, and training corpus size are not published. The closest characterization is the March 2025 multi-country post, which describes training on "vast petabyte datasets from our internal fleet and data partners."

Is GAIA a product or a training tool? A training, simulation and validation tool. It generates driving scenarios, including NCAP-style tests, for closed-loop evaluation. The GAIA-4 post calls GAIA "one building block of Simulation 2.0, our broader framework for evaluating Wayve AI Driver for commercial deployment." It does not drive the car.

Who are the named partners? As listed on Wayve's press index: Uber, Nissan, Stellantis, Qualcomm, NVIDIA, AMD, Arm, Microsoft, S.RIDE, and the UK Government. Wayve also cites UNECE regulatory work and the UK-funded DriveSafeSim project. This is what the company has announced, and announcements of collaboration are not evidence of shipped product.