Wayve
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v2 ยท 4,185 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v2 ยท 4,185 words
Add missing citations, update stale details, or suggest a clearer explanation.
Wayve is a British artificial intelligence company, founded in Cambridge in 2017 by Alex Kendall and Amar Shah, that develops end to end "embodied AI" software for autonomous driving and robotics, and in May 2024 raised a 1.05 billion United States dollar Series C led by SoftBank that was the largest artificial intelligence funding round ever announced in Europe.[8][9] Headquartered in London, Wayve is best known for popularising the so called AV 2.0 approach to self driving, which replaces the modular sense plan act robotic stack of earlier autonomous vehicle programmes with a single, end to end neural network trained on raw sensor data. The company has developed two flagship lines of foundation models for driving, the GAIA family of generative world models and the LINGO family of vision language action models, and has raised more than two and a half billion dollars from investors including SoftBank, NVIDIA, Microsoft, Mercedes Benz, Stellantis, Nissan and Uber.[1][2][3]
Wayve does not operate its own robotaxi fleet. Instead it sells its driving foundation model and surrounding software stack to automotive original equipment manufacturers (OEMs) and mobility partners, positioning itself as a horizontal supplier comparable in business model to Mobileye but built on a different technical foundation. Its self driving cars have been tested on public roads in the United Kingdom, the United States, Germany and Japan, and the company has announced commercial pilots and agreements with Nissan, Uber, Asda and Ocado.[12][14]
Wayve Technologies Ltd. was incorporated on 21 August 2017 in Cambridge, England, by Alex Kendall and Amar Shah, two machine learning doctoral students at the University of Cambridge.[11] The company began operating from a residential house outside Cambridge before later relocating its headquarters to King's Cross in London. Co founder Amar Shah departed the company in 2020 and Alex Kendall became sole chief executive officer.[10]
Wayve frames its mission as building embodied artificial intelligence for the physical world, with self driving as its first deployment vertical. The company describes its product as a driving foundation model, a single learned policy that maps camera and other sensor inputs to driving actions, and emphasises that the same underlying approach can in principle be transferred between vehicle platforms with little engineering effort.[1] Announcing the 2024 Series C, Kendall said the round reflected "our team's unwavering conviction that Embodied AI will address long-standing challenges in scaling this technology to everyone, everywhere."[8]
As of early 2026 Wayve employed several hundred staff across offices in London, Mountain View, Vancouver and Yokohama, and reported a post money valuation of approximately 8.6 billion United States dollars following an extension to its Series C funding.[19]
The origins of Wayve trace back to research at the University of Cambridge by Alex Kendall under the supervision of Roberto Cipolla. Kendall, originally from New Zealand, completed his PhD on deep learning for geometric computer vision and simultaneous localisation and mapping. His doctoral work on probabilistic deep learning and Bayesian uncertainty estimation, together with related research on end to end learning to drive, became part of the conceptual foundation for the company. Amar Shah, who studied in Zoubin Ghahramani's Machine Learning Group, brought additional expertise in probabilistic methods and reinforcement learning.[11]
In 2017 the pair founded Wayve and began experimenting with reinforcement learning to control a small electric vehicle. A 2018 paper titled Learning to Drive in a Day, in which a reinforcement learning agent learned lane following in a single afternoon of training on a public road, brought the company to wider attention and helped articulate the central claim that driving could be learned end to end without hand engineered perception, prediction and planning modules.[1]
In 2019 Wayve demonstrated that the same techniques could drive a car through urban streets in central Cambridge that the system had never seen before, using only cameras, a satellite navigation map and a deep neural network.[10] The company subsequently moved its main operations to London and began collecting on road data with a fleet of retrofitted Jaguar I Pace and Ford Mondeo vehicles.
Following Amar Shah's departure in 2020, Wayve concentrated on scaling its data, compute infrastructure and engineering team. In November 2019 it closed a 20 million dollar Series A, and in January 2022 it raised 200 million dollars in a Series B led by Eclipse Ventures with participation from D1 Capital, Baillie Gifford and Microsoft. The Series B was the largest Series B for a European artificial intelligence company at the time.[11]
At this point the company began publicly articulating the AV 2.0 thesis: that the modular software stacks built by first generation autonomous vehicle programmes such as Waymo, Cruise and Argo AI were an evolutionary dead end, and that an end to end neural network trained on diverse driving data was a more scalable path to robust self driving.[1] AV 2.0 was framed as analogous to the transition in computer vision and natural language processing from hand engineered pipelines to deep learning.
In 2023 Wayve announced two new product directions. In June it previewed GAIA-1, a generative world model trained on its UK driving data, and released a technical report in September.[2][3] In September of the same year it introduced LINGO-1, a vision language model that produced natural language commentary on driving scenes.[6] Both releases were aimed at making the company's end to end approach more interpretable, more amenable to evaluation and more efficient to train.
In May 2024 Wayve announced a Series C round of 1.05 billion United States dollars led by SoftBank, with new investor NVIDIA and existing investor Microsoft participating.[8][9] The round was at the time the largest artificial intelligence funding round announced in Europe and one of the twenty largest worldwide.[9] It brought Wayve's cumulative funding to approximately 1.31 billion dollars and added SoftBank to the company's board of directors.[8]
During 2024 and 2025 Wayve opened offices in Mountain View and Yokohama, expanded testing to highways and city streets in the United States and Germany, and announced partnerships with Asda, Ocado and Uber.[14][15][16] In April 2024 it unveiled LINGO-2, the first closed loop vision language action driving model tested on public roads, and in March 2025 it released GAIA-2, a controllable multi view generative world model.[7][4]
In December 2025 Wayve released GAIA-3, a 15 billion parameter generative world model focused on validation and evaluation.[5] On 10 December 2025 Wayve and Nissan signed definitive agreements to integrate the "Wayve AI Driver" into the next generation ProPILOT driver assistance series across a broad range of Nissan vehicles, with Nissan planning to introduce the first ProPILOT model using the technology in Japan in fiscal year 2027.[12] In February 2026 Wayve signed a memorandum of understanding with the United Kingdom government on safety assurance and large scale simulation, joined Uber and Nissan on a robotaxi pilot in Tokyo planned for late 2026, and raised additional funding from Mercedes Benz, Stellantis, Nissan and existing investors that lifted its valuation to roughly 8.6 billion dollars.[13][19]
Alex Kendall serves as chief executive officer and is the public face of Wayve. He completed a PhD in computer vision at the University of Cambridge with a focus on deep learning for SLAM and Bayesian uncertainty, and his early academic work on geometric loss functions and end to end learning informs the company's technical direction.[10] Kendall has become a prominent advocate for the AV 2.0 framing in industry conferences, podcasts and policy discussions and was awarded an MBE in 2024 for services to engineering.[18]
Amar Shah, the co founder who departed in 2020, completed his PhD in Bayesian optimisation and probabilistic machine learning, also at Cambridge. After leaving Wayve he founded a separate machine learning company.
The broader leadership team has included Erez Dagan as president, formerly of Mobileye, Silvius Rus as senior vice president of engineering, formerly of Waymo, and Jamie Shotton as chief scientist, who previously led research at Microsoft Research Cambridge.
The AV 2.0 narrative is the central organising idea of Wayve's technology. In Wayve's framing, traditional autonomous driving programmes from the late 2000s and early 2010s, sometimes referred to as AV 1.0, relied on a modular stack consisting of a perception module that detects objects, a prediction module that estimates their future trajectories, a planner that chooses an ego trajectory and a control module that produces actuator commands. Each module was typically built using a mixture of classical algorithms and supervised deep learning, calibrated against a high definition map of the operating environment and a large library of hand written behavioural rules.[1]
Wayve argues that this approach has three structural problems: it does not scale across cities because every new domain requires fresh HD maps and rule updates; hand written rules struggle with the long tail of unusual but safety critical situations; and the loss functions used to train modules in isolation are only loosely correlated with the true downstream objective of safe and comfortable driving.[1]
AV 2.0 replaces this stack with a single neural network trained end to end on diverse driving data. The network takes raw camera images, vehicle state, route information and optional auxiliary sensors as input, and produces a driving trajectory or low level control commands as output. Training uses a mixture of imitation learning from expert human drivers, reinforcement learning in world model based simulation and self supervised objectives derived from large quantities of unlabelled driving video. Key claimed advantages include generalisation to new cities and vehicle platforms without per location engineering, the avoidance of expensive HD map maintenance, and the ability to leverage data at internet scale.[1][17]
| Programme | Sensor suite | Architecture | HD maps | Business model |
|---|---|---|---|---|
| Wayve | Cameras with optional radar, no lidar required | Single end to end neural network, AV 2.0 | Not required | Software licensed to OEMs |
| Waymo | Cameras, radar, lidar | End to end trained foundation model with explicit object encoder | Used extensively | First party robotaxi service |
| Tesla Full Self Driving | Cameras only | End to end neural network, recent versions | Not used | Consumer subscription on Tesla vehicles |
| Cruise (until 2023) | Cameras, radar, lidar | Modular stack with deep learning components | Used extensively | First party robotaxi service |
| Mobileye | Cameras with optional radar and lidar | Modular stack with deep learning, REM crowdsourced maps | REM map used | Software and silicon licensed to OEMs |
Wayve frequently positions itself as occupying a distinctive square in this matrix, combining a fully end to end model in the Tesla tradition with an OEM licensing business model in the Mobileye tradition, while explicitly rejecting both first party robotaxi operation and reliance on high definition maps.[10]
The GAIA line is Wayve's family of generative world models. World models in the autonomous driving context are deep generative models that learn to predict future video and sensor observations from past observations and candidate actions, and can therefore be used as simulators for training and evaluating driving policies.[2]
GAIA-1, previewed in June 2023 and described in a technical report released in September 2023, is a nine billion parameter generative world model trained on approximately 4,700 hours of Wayve's UK driving data.[2][3] The model, in the words of its authors, "casts world modeling as an unsupervised sequence modeling problem by mapping the inputs to discrete tokens, and predicting the next token in the sequence," with separate encoders for video, action and text and a transformer that predicts the next token.[3] The associated arXiv paper, GAIA-1: A Generative World Model for Autonomous Driving, was posted on 29 September 2023 by Anthony Hu, Lloyd Russell, Alex Kendall, Jamie Shotton and colleagues.[3]
GAIA-1 generates short clips of plausible driving video conditioned on previous frames, ego vehicle actions and natural language prompts such as descriptions of weather or traffic conditions.[2] It is used internally for data augmentation, scenario generation and evaluation.
GAIA-2, released on 26 March 2025, is described as a controllable multi view generative world model.[4] The model is trained on a larger and more diverse dataset including data from the United Kingdom, the United States and Germany, multiple vehicle platforms ranging from passenger cars to vans and multiple camera configurations. GAIA-2 supports fine grained control over ego vehicle actions, dynamic agents and scene properties such as time of day and weather, and generates consistent video across several synchronised cameras.[4]
GAIA-2 was released alongside a technical report and a permissive research preview. The model is intended both as an internal tool for generating corner case scenarios and as a building block for closed loop policy evaluation.
GAIA-3, released on 2 December 2025, is a 15 billion parameter generative world model that doubles the parameter count of GAIA-2 and uses a video tokeniser of roughly twice the capacity.[5] According to Wayve's announcement GAIA-3 is pre trained on approximately ten times the data of its predecessor, spanning multiple continents and a wide range of vehicle types, environments and driving conditions. The release explicitly frames GAIA as a foundation for autonomy evaluation rather than primarily a tool for visual synthesis, with scene generation aimed at safety assurance and regulatory evidence.[5]
| Model | Release | Parameters | Notes |
|---|---|---|---|
| GAIA-1 preview | June 2023 | 9 billion | First generative world model trained on driving data; preview only |
| GAIA-1 technical report | September 2023 | 9 billion | arXiv paper 2309.17080; trained on ~4,700 hours of UK driving |
| GAIA-2 | 26 March 2025 | Approximately 8 billion | Controllable multi view world model; multi country dataset |
| GAIA-3 | 2 December 2025 | 15 billion | Doubled parameters and video tokeniser; ten times more data; evaluation focused |
The LINGO line is Wayve's family of vision language models for driving, used to add natural language interpretation and instruction following to the driving foundation model.
LINGO-1, announced in September 2023, is an open loop driving commentator.[6] It combines the company's vision backbone with an autoregressive language model trained to produce natural language commentary about what the driving system perceives and intends to do. LINGO-1 was trained on a dataset of expert UK drivers narrating their own driving for several thousand hours, allowing the model to associate visual scenes with natural language explanations of road behaviour.[6] Because LINGO-1 generated commentary retrospectively from observed video, it was a useful explanation and evaluation tool but did not itself control the vehicle.
LINGO-2, introduced on 17 April 2024, is described by Wayve as "the first closed-loop vision-language-action driving model (VLAM) tested on public roads."[7] It consists of a vision module that maps consecutive camera frames into a sequence of visual tokens and an autoregressive language model that consumes those tokens together with conditioning information such as the planned route, the current speed and the speed limit. The language model is trained jointly to predict a future driving trajectory and a natural language commentary describing the scene and the chosen action. Wayve tested the model both in its neural simulator, Ghost Gym, and on public roads.[7]
LINGO-2 supports natural language prompts that influence driving behaviour, for example instructing the car to pull over, change lane or take a particular turn. This opens the door to natural language as a control interface for autonomous driving and as a tool for verifying that the model's perception of a scene is consistent with its actions. Wayve has demonstrated LINGO-2 driving in central London while producing real time commentary explaining its decisions.[7]
| Model | Year | Type | Controls vehicle | Notes |
|---|---|---|---|---|
| LINGO-1 | September 2023 | Open loop vision language commentator | No | Trained on UK expert driver narrations |
| LINGO-2 | April 2024 | Closed loop vision language action model | Yes | First closed loop VLAM tested on public roads |
Wayve has also published research on probabilistic deep learning, neural radiance fields, learned simulation, scaling laws for driving, and the integration of foundation models with classical planning and safety mechanisms. The company has invested heavily in fleet learning infrastructure, including cloud scale data curation, automated long tail mining and continuous training pipelines.
Wayve has raised capital across at least seven announced rounds. The table below summarises the principal disclosed rounds.
| Round | Date | Amount | Lead investor | Selected co investors |
|---|---|---|---|---|
| Seed | April 2018 | 3.1 million USD | Compound | Fly Ventures, Local Globe, First Minute Capital |
| Series A | November 2019 | 20 million USD | Eclipse Ventures | Compound, Balderton, First Minute Capital |
| Series B | January 2022 | 200 million USD | Eclipse Ventures | D1 Capital, Baillie Gifford, Microsoft, Linse Capital, Ocado Group |
| Strategic financing | February 2023 | 13.6 million GBP | Ocado Group | Ocado Group strategic investment |
| Series C | May 2024 | 1.05 billion USD | SoftBank | NVIDIA, Microsoft |
| Strategic OEM round | February 2026 | Approximately 500 million USD reported | Mercedes Benz | Stellantis, Nissan, Uber, Microsoft, NVIDIA |
The May 2024 Series C of 1.05 billion dollars was the largest single funding round raised by a private artificial intelligence company in Europe at the time and, according to the company, the United Kingdom's largest artificial intelligence fundraise ever.[8][9] The strategic OEM round in February 2026 brought cumulative reported funding to roughly 2.5 billion dollars and lifted Wayve's valuation to approximately 8.6 billion dollars.[19]
Wayve has structured its commercial strategy around partnerships with vehicle manufacturers, fleet operators and government agencies rather than first party robotaxi operations.
| Partner | Year | Type | Description |
|---|---|---|---|
| Asda | 2021 | Fleet pilot | Twelve month autonomous delivery trial on UK roads, billed as the largest UK self driving grocery delivery trial at the time |
| Ocado Group | 2023 | Strategic investment and pilot | 13.6 million GBP investment and partnership to develop autonomous grocery delivery for OSP retailers globally |
| NVIDIA | 2024 | Investor and silicon partner | Strategic investment as part of the Series C and ongoing collaboration on automotive grade compute |
| Microsoft | 2022 onward | Investor and cloud partner | Series B investor and strategic cloud and AI infrastructure partner |
| SoftBank | 2024 | Lead Series C investor | Board representation and capital commitment |
| Uber | 2024 to 2026 | Mobility partner | Joint plans for Level 4 autonomy trials in the United Kingdom and a Tokyo robotaxi pilot with Nissan |
| Nissan | 2025 to 2026 | OEM partner | Definitive agreements signed December 2025 to integrate the Wayve AI Driver into the next generation ProPILOT driver assistance series across a broad range of Nissan vehicles, with first launch planned in Japan in fiscal year 2027 |
| Mercedes Benz | 2026 | OEM partner and investor | Strategic investment alongside collaboration on automated driving for production vehicles |
| Stellantis | 2026 | OEM partner and investor | Strategic investment as part of the 2026 funding round |
| United Kingdom government | 2026 | Public sector partner | Memorandum of understanding with the Department for Business and Trade covering safety assurance, large scale simulation and integration of self driving technology |
The Asda and Ocado pilots have used Wayve equipped vans on public road delivery routes in the London area, refining the company's narrative around urban driving without high definition maps.[15][16] The December 2025 Nissan agreements represent a significant step toward consumer scale deployment, both as a driver assistance product integrated into Nissan production vehicles and as the basis for an urban robotaxi service in Tokyo; a joint Nissan and Wayve robotaxi prototype built on NVIDIA DRIVE Hyperion was prepared for the Uber trials.[12]
The competitive landscape has shifted significantly since 2017. First generation robotaxi programmes such as Waymo and Cruise have continued to focus on first party operations with sensor rich vehicles in geofenced urban environments. Tesla has continued to develop a camera only consumer driving assistance product and has converged toward an end to end neural network architecture similar in spirit to Wayve's. Mobileye operates an OEM oriented model that uses a modular stack with cameras, radar and optional lidar plus a crowdsourced Road Experience Management map. Wayve's distinctive combination of an end to end driving foundation model, a camera centric sensor suite, no reliance on high definition maps and a third party supplier business model has been positioned as a more capital efficient path to scaling autonomous driving. Independent observers have noted that the broader industry has moved toward end to end learning since 2023, partially validating the AV 2.0 thesis, though Wayve still has less public real world driving exposure than Waymo's US robotaxi fleets.[19]
Wayve's UK testing is conducted under the framework established by the Automated and Electric Vehicles Act 2018 and the Automated Vehicles Act 2024, which created a formal regulatory framework for self driving in Great Britain including a safety standard and a permitting regime for automated passenger services and goods deliveries. The 2026 memorandum of understanding between Wayve and the Department for Business and Trade supports the operationalisation of that act, including the use of large scale simulation, in particular GAIA based generative world models, as evidence in safety cases.[13] In the United States Wayve's testing is conducted under state level autonomous vehicle testing regimes, particularly in California, and in coordination with the National Highway Traffic Safety Administration.
Wayve has received broadly positive coverage for its technical research, with GAIA-1, GAIA-2 and LINGO-2 attracting attention from both the machine learning research community and the automotive press. The AV 2.0 framing has been credited with sharpening industry discussion about the trade offs between modular and end to end approaches to autonomy. Critical commentary has noted that the company's camera centric sensor philosophy understates the difficulty of demonstrating safety to regulators without interpretable intermediate representations, and that Wayve does not yet operate a paid driverless service at scale.[19] Wayve has responded by emphasising the role of GAIA world models in evaluation and LINGO natural language interfaces in interpretability.