Wayve Leverages Automaker Fleets to Scale Physical AI Past Robotaxis
Ai4
Wayve argues passenger vehicles offer the only hardware footprint large enough to generate the real-world operational data required to develop physical foundation models.
## Background
Published on YouTube by the Ai4 channel on August 21, 2026, this interview features Financial Times San Francisco correspondent Rafe Uddin in conversation with Wayve VP of AI Vijay Badrinarayanan. According to the description, the discussion explores why Physical AI is emerging as the industry's next major frontier, examining how end-to-end learning and world models empower AI to understand, reason, and act in the physical world.
## TL;DR
Vijay Badrinarayanan outlines Wayve's strategic vision to develop generalizable physical AI foundation models for the 100 million passenger vehicles produced annually, rather than building custom, vertically integrated robotaxi hardware. Tracing Wayve's roots from Cambridge University's SegNet to major commercial partnerships with Nissan, Stellantis, and Uber, he explains how modern software-defined vehicles provide the vast on-policy data and compute base necessary to train embodied intelligence. The discussion emphasizes how generative world models like GAIA enable safety verification across hundreds of global cities, illustrating that cross-hardware generalization in autonomous vehicles serves as the foundational stepping stone toward widespread industrial and embodied robotics.
## Atmosphere & style
The conversation has the measured, authoritative cadence of an elite industry fireside chat, balancing journalistic rigor with deep architectural insight. Rafe Uddin poses incisive questions about commercial viability, public safety, and industry competition, while Badrinarayanan responds with steady, conviction-driven clarity backed by over a decade of deep learning research. The tone is pragmatic yet distinctly visionary, creating a compelling insider look at how physical AI is expanding beyond software chatbots into real-world machines.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=kvO6egx_rRY&t=4s) Wayve was founded on the insight from SegNet that end-to-end, data-driven deep learning outperforms classical computer vision, scaling autonomy directly with data and compute.
- [03:47](https://www.youtube.com/watch?v=kvO6egx_rRY&t=227s) Rather than manufacturing custom vehicles or captive fleets, Wayve positions itself as a pure software intelligence layer aimed at the 100 million passenger vehicles produced globally each year.
- [03:47](https://www.youtube.com/watch?v=kvO6egx_rRY&t=227s) Wayve's unified foundation model architecture spans from L2++ supervised driving up to L5 full autonomy, supported by major OEM partnerships with Nissan and Stellantis.
- [07:16](https://www.youtube.com/watch?v=kvO6egx_rRY&t=436s) Wayve rejects narrow geofenced operational design domains (ODDs), validating its models across roughly 500 cities worldwide to solve autonomy and geographic generalization simultaneously.
- [08:49](https://www.youtube.com/watch?v=kvO6egx_rRY&t=529s) Modern software-defined passenger vehicles provide the largest deployed base of high-compute robots, serving as the essential data flywheel for collecting on-policy physical AI data.
- [13:08](https://www.youtube.com/watch?v=kvO6egx_rRY&t=788s) Wayve plans to utilize automotive partner manufacturing facilities and existing industrial robots for robotics foundation model training rather than building artificial test warehouses.
- [15:11](https://www.youtube.com/watch?v=kvO6egx_rRY&t=911s) Training on London's dense, complex street layout created robust learned representations that transferred to US driving conditions within months.
- [15:11](https://www.youtube.com/watch?v=kvO6egx_rRY&t=911s) Wayve's global footprint pairs the UK's foundational AI research strengths with Silicon Valley's specialized autonomous vehicle systems and engineering talent.
- [18:05](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1085s) Testing in San Francisco allows Wayve to benchmark its foundation model approach directly against vertically integrated incumbents like Waymo and Zoox.
- [20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) Seamless transfer between diverse vehicle platforms like the Ford Mustang Mach-E and Nissan demonstrates the hardware-agnostic portability of Wayve's core intelligence.
- [20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) Generative world models like GAIA are crucial for policy evaluation, data augmentation, and closed-loop safety verification before real-world deployment.
- [20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) Physical AI and embodied intelligence are advancing rapidly with massive automaker demand, representing a substantially larger technological frontier than text-based chatbots.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=kvO6egx_rRY&t=4s) The Roots of Wayve: From Cambridge and SegNet to Physical AI
Financial Times San Francisco correspondent Rafe Uddin introduces Vijay Badrinarayanan, VP of AI at Wayve, noting the FT's long-standing interest in the London-born autonomous driving company. Badrinarayanan leads Wayve's Bay Area team focusing on learned intelligence, representation learning, and multimodal foundation models. Before joining Wayve, Badrinarayanan led a team at Magic Leap and collaborated closely with Wayve's founder and CEO, Alex Kendall, during Kendall’s PhD at Cambridge University. Together in the 2013–2014 era, they created SegNet, an early deep learning system for end-to-end road scene segmentation.
Reflecting on the company’s inception, Badrinarayanan explains that Wayve was founded in 2017 at the tail end of Kendall’s doctorate. The fundamental insight behind SegNet was that a data-driven, end-to-end learned approach delivered a "night and day" leap in performance compared to the prior ten to fifteen years of classical computer vision. This became Wayve’s guiding conviction: building generalizable intelligence for autonomous systems requires an end-to-end, data-driven architecture where capability scales directly with compute and data. Starting with passenger cars before expanding into other physical embodiments, Wayve has grown from foundational research into a commercial enterprise building physical AI foundation models alongside world models like GAIA for safety verification, evaluation, and data augmentation.
```mermaid
flowchart LR
accTitle: Wayve Core Intelligence Architecture
subgraph Inputs["Data & Sensors"]
A["Diverse Sensor Stack<br/>Cameras, Radar, etc."] --> B["On-Board Compute"]
end
subgraph Core["Learned Intelligence Layer"]
B --> C["End-to-End Foundation Model"]
D["World Models<br/>(e.g., GAIA)"] -. Simulation & Eval.-> C
end
subgraph Output["Embodied Action"]
C --> E["Driving Policy & Control Actions"]
C --> F["L2++ Supervised to L5 Autonomy"]
end
```
- **Founding Date:** 2017 out of Cambridge University by Alex Kendall (mentored by Vijay Badrinarayanan).
- **Early Precedent:** SegNet (2013–2014), pioneering end-to-end deep learning for pixel-wise semantic segmentation of urban scenes.
- **Core Principle:** Autonomy is fundamentally an intelligence problem that must scale with compute, data, and learned representations rather than hard-coded heuristics.
---
[03:47](https://www.youtube.com/watch?v=kvO6egx_rRY&t=227s) The Intelligence Layer: Wayve's Commercial Strategy vs. Vertically Integrated AVs
Uddin contrasts Wayve's positioning with high-profile competitors like Waymo scaling robotaxi fleets in San Francisco, Tesla pursuing supervised autonomy through full vertical integration, and Zoox deploying custom vehicles in Las Vegas. Badrinarayanan frames Wayve’s mission around the approximately 100 million passenger vehicles manufactured globally each year. Rather than manufacturing hardware or deploying its own captive vehicle fleets, Wayve positions itself strictly as the generalizable software intelligence layer capable of distilling foundation models down to run across disparate sensor configurations, vehicle platforms, and embedded compute chips.
This strategy allows Wayve to monetize across the entire spectrum of autonomy—from L2++ point-to-point supervised highway and urban driving up to fully autonomous L5 robotaxis—using a unified intelligence core. On the commercial front, Wayve has secured major OEM partnerships, expanding its agreement with Nissan to cover nearly 90% of their production volume (numbering in the millions of vehicles) alongside a recent partnership with Stellantis. Concurrently, Wayve obtained a commercial testing license in London, partnering with Uber to launch a supervised autonomous ride-hailing service where the public can directly experience the technology.
- **Market Focus:** 100 million new vehicles produced globally per year.
- **OEM Partnerships:** Nissan (scaling to ~90% of vehicle production) and Stellantis.
- **Service Deployment:** Supervised autonomy ride-hailing service in London launched on the Uber network.
---
[07:16](https://www.youtube.com/watch?v=kvO6egx_rRY&t=436s) Scaling Strategy: Broad Real-World Testing vs. Constrained ODDs
Addressing Uddin's question about public perception, friction in urban deployment, and safety concerns, Badrinarayanan emphasizes that Wayve’s deployment methodology diverges sharply from competitors who restrict themselves to narrow operational design domains (ODDs). While existing robotaxi operators typically scale city by city within tightly mapped, geofenced boundaries, Wayve believes the only path to true safety and generalizability is addressing both the level of autonomy and the geographic diversity of deployment at the same time.
To validate its models against global edge cases, Wayve has already tested its autonomous driving systems across approximately 500 cities worldwide. The supervised ride-hailing deployment with Uber represents the initial phase of public rollout, designed to ensure rigorous safety while rapidly expanding coverage across diverse metropolitan environments through partnerships with mobility networks and vehicle manufacturers.
---
[08:49](https://www.youtube.com/watch?v=kvO6egx_rRY&t=529s) Software-Defined Cars as the Stepping Stone to Physical AI
When asked why Wayve began with autonomous vehicles rather than other robotic form factors, Badrinarayanan explains that cars represent "the largest stable base of robots which are available in the world today." Modern vehicles are increasingly software-defined platforms outfitted with high-bandwidth sensors and serious on-board compute capacity capable of executing frontier AI models in real time.
Crucially, developing physical AI requires vast amounts of "on-policy data"—data collected while the model is actively interacting with the physical environment, logging its safe actions, edge cases, and failure modes. Collecting this data at scale requires a viable, high-value commercial vehicle market. Once the software layer generalizes across varying sensor suites and partner compute stacks, it creates a self-reinforcing data flywheel. While Wayve conducts deep research into bipeds, humanoids, and robotic manipulators, Badrinarayanan notes that those hardware platforms have a tiny deployed footprint and rely heavily on synthetic data. However, Wayve’s core breakthroughs—simulation engines, world models, and end-to-end training algorithms developed over a decade—translate directly to broader robotics.
| Dimension | Automotive Physical AI (Wayve Core) | Non-Automotive Robotics (Humanoids, Arms) |
|:--- |:--- |:--- |
| **Installed Hardware Base** | Tens of millions of software-defined cars globally | Very small, fragmented fleets in labs and early pilots |
| **On-Board Compute & Sensors** | Automotive-grade chips and rich camera/radar suites | Highly heterogeneous, often constrained payloads |
| **Data Collection** | Scaled on-policy collection via production passenger fleets | Heavily reliant on passive data, synthetic sims, or small warehouses |
| **Current Readiness** | Commercial production deployments (L2++ to Supervised AV) | Exploratory research phase; frontier scaling in progress |
---
[13:08](https://www.youtube.com/watch?v=kvO6egx_rRY&t=788s) Synergies with Industrial Robotics and Manufacturing Partners
Uddin raises skepticism regarding robotics startups that lack real-world physical data and resort to scraping passive video from YouTube or setting up isolated training warehouses with robotic arms. In contrast to those approaches and the efforts of labs like OpenAI, Anthropic, or Travis Kalanick's venture Atma, Badrinarayanan highlights Wayve's structural advantage through its existing automotive manufacturing client relationships.
Major automotive partners operate extensive factory footprints equipped with thousands of industrial robots, gantry systems, and manipulators. Rather than building artificial test warehouses from scratch, Wayve plans to leverage its enterprise relationships to deploy and evaluate foundation models directly in real-world industrial environments, tapping into practical workflows and high-fidelity operational data.
---
[15:11](https://www.youtube.com/watch?v=kvO6egx_rRY&t=911s) From London's Chaos to Silicon Valley: Engineering and Talent Expansion
Reflecting on Wayve's geographic journey, Badrinarayanan defends the decision to start in Cambridge and London. Navigating London's chaotic, narrow, and unpredictable urban core—supported by permissive testing regulations—provided an exceptionally rich and dense training signal. Training end-to-end models on complex European city streets bootstrapped Wayve’s generalizability far faster than testing in structured, low-density suburbs would have, allowing the system to adapt to US driving conditions within months.
Wayve subsequently expanded its presence to the US West Coast, Germany, and Japan. The Bay Area office was established to access deep pools of multi-disciplinary engineering talent, particularly in robotics software engineering and automotive hardware integration, developed through decades of autonomous vehicle investment in California. This dual-presence strategy combines the UK’s strong foundational AI research culture with Silicon Valley’s specialized AV systems expertise.
- **London as a Testbed:** Provided high-complexity, unstructured urban driving data under supportive UK regulatory frameworks.
- **US Transferability:** Wayve’s learned models adapted from UK left-hand traffic to US roads within just a few months.
- **Global Footprint:** Engineering and operational hubs in the United Kingdom, the United States (Silicon Valley), Germany, and Japan.
---
[18:05](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1085s) Benchmarking in San Francisco and the Multi-City 2025–2026 Roadmap
Operating test vehicles in San Francisco allows Wayve to benchmark its end-to-end intelligence directly against incumbent AV players like Waymo and Zoox on the same streets. Demonstrating robust performance in both London and San Francisco serves as empirical proof to global automakers that Wayve's generalizable architecture can handle fundamentally different urban environments without brittle, hand-crafted ODD rulebooks.
Badrinarayanan details the near-term deployment roadmap across two distinct commercial streams:
1. **Supervised L2++ / Hybrid Urban Autonomy:** Integrated into consumer production vehicles for Nissan and Stellantis across North America, Europe, and Japan.
2. **Supervised Robotaxi Mobility on the Uber Network:** Rolling out public commercial trials in London using modified Ford Mustang Mach-E vehicles, and in Tokyo utilizing Nissan platforms, with plans to expand the service to 10 additional cities the following year.
[20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) Cross-Hardware Generalization and the Reality of Physical AI
Reflecting on Wayve's rapid hardware expansion, Rafe Uddin highlights how remarkably fast the company's autonomous driving technology has transferred across completely different vehicular platforms, pointing specifically to successful deployments across Ford Mustang Mach-E and Nissan vehicles. Vijay Badrinarayanan affirms that this cross-hardware generalizability is one of the most encouraging indicators of the foundation model approach's scalability.
When asked for his final thoughts, Vijay emphasizes a core message for the audience: "physical AI is real." While building intelligence for the physical realm introduces deep, unique engineering challenges, rapid breakthroughs in world modeling and simulation are revolutionizing development. He points directly to Wayve's GAIA world model and simulation engine, noting how transformative it is for evaluating autonomous policies and rigorously validating driving behavior for safety and reliability. While much of the tech industry remains fixated on pure large language models (LLMs) and software-based agentic workflows, Vijay warns that Physical AI is advancing at breakneck speed, driven by intense demand from global automotive manufacturers. He concludes that embodied physical intelligence will ultimately prove to be "way bigger than the chatbot."
- **Hardware Agnosticism:** Wayve’s end-to-end models successfully generalize across diverse OEM platforms, including Nissan and Ford Mustang Mach-E vehicles, proving software portability without needing platform-specific redesigns.
- **Simulation and Safety Validation:** Advanced generative world models like GAIA enable comprehensive policy testing, closed-loop evaluation, and safety validation in simulated environments before real-world deployment.
- **Automaker Demand:** Tier-1 automotive manufacturers are demonstrating intense commercial and technical interest in deploying embodied AI into consumer vehicles.
- **Beyond the Chatbot Frontier:** While text-based generative AI popularized the current AI wave, the integration of generative modeling with physical action represents a substantially larger technological and industrial frontier.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=kvO6egx_rRY&t=4s) The fundamental insight behind SegNet was that a data-driven, end-to-end learned approach delivered a "night and day" leap in performance compared to the prior ten to fifteen years of classical computer vision. — Vijay Badrinarayanan
> [00:04](https://www.youtube.com/watch?v=kvO6egx_rRY&t=4s) Autonomy is fundamentally an intelligence problem that must scale with compute, data, and learned representations rather than hard-coded heuristics. — Vijay Badrinarayanan
> [07:16](https://www.youtube.com/watch?v=kvO6egx_rRY&t=436s) the only path to true safety and generalizability is addressing both the level of autonomy and the geographic diversity of deployment at the same time. — Vijay Badrinarayanan
> [08:49](https://www.youtube.com/watch?v=kvO6egx_rRY&t=529s) the largest stable base of robots which are available in the world today — Vijay Badrinarayanan
> [08:49](https://www.youtube.com/watch?v=kvO6egx_rRY&t=529s) Crucially, developing physical AI requires vast amounts of "on-policy data"—data collected while the model is actively interacting with the physical environment, logging its safe actions, edge cases, and failure modes. — Vijay Badrinarayanan
> [15:11](https://www.youtube.com/watch?v=kvO6egx_rRY&t=911s) Navigating London's chaotic, narrow, and unpredictable urban core—supported by permissive testing regulations—provided an exceptionally rich and dense training signal. — Vijay Badrinarayanan
> [20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) physical AI is real. — Vijay Badrinarayanan
> [20:15](https://www.youtube.com/watch?v=kvO6egx_rRY&t=1215s) way bigger than the chatbot. — Vijay Badrinarayanan
## Mentioned
Vijay Badrinarayanan, Rafe Uddin, Alex Kendall, Travis Kalanick, Wayve, Financial Times, Magic Leap, Cambridge University, SegNet, GAIA, Waymo, Tesla, Zoox, Nissan, Stellantis, Uber, Ford Mustang Mach-E, OpenAI, Anthropic, Atma, London, San Francisco, Tokyo, Silicon Valley
## Video description
Vijay Badrinarayanan - VP of AI, Wayve.ai
Rafe Uddin - San Francisco Correspondent, Financial Times
Generative AI transformed how machines understand and generate information. The next challenge is enabling AI to understand, reason and act in the physical world. In conversation with Financial Times San Francisco Correspondent Rafe Uddin, Wayve VP of AI Vijay Badrinarayanan explores why Physical AI is emerging as the industry's next frontier, how advances in world models and end-to-end learning are making it possible, and why Wayve chose to pursue this approach long before it entered the mainstream.
Wayve argues that physical artificial intelligence will eclipse text-based chatbots by using end-to-end foundation models that treat autonomy as a software scaling challenge rather than a hardware or geofenced mapping problem, according to VP of AI Vijay Badrinarayanan.
End-to-End Learning Over Geofenced Heuristics (0:04)
Wayve's technical philosophy stems from the 2013–2014 development of SegNet at Cambridge University by CEO Alex Kendall and Badrinarayanan, which demonstrated that data-driven deep learning delivered a major leap over classical computer vision. Founded in 2017, the company rejects the narrow operational design domains (ODDs) and rigid HD-mapping used by traditional robotaxi operators. Instead, it relies on an end-to-end foundation model architecture where system capabilities scale directly with compute and diverse real-world data.
To ensure its policy generalises globally, Wayve has tested systems across approximately 500 cities. It pairs real-world driving data with generative world models, such as GAIA, which provide closed-loop simulation, safety verification, evaluation, and synthetic data augmentation before real-world deployment.
The Software Layer for Mass-Market Fleets (3:47)
Rather than manufacturing bespoke vehicles or operating captive fleets like Waymo or Zoox, Wayve targets the roughly 100 million passenger vehicles produced globally each year. Positioning itself strictly as an intelligence software layer, the company distills its foundation models to execute across varied sensor suites, vehicle architectures, and on-board compute platforms, spanning levels from L2++ supervised driving to L5 autonomy.
This platform-agnostic strategy has led to production agreements with major original equipment manufacturers (OEMs). Wayve expanded its partnership with Nissan to cover close to 90% of its production volume—representing millions of vehicles—and added Stellantis as a partner. Demonstrating cross-hardware transferability, the software has run across disparate platforms, including Ford Mustang Mach-E and Nissan vehicles.
Cars as the Foundation for Physical Robotics (8:49)
Wayve selected passenger vehicles as its entry point because modern software-defined cars represent the largest existing installed base of sensor-rich, compute-capable robots. Developing robust physical AI requires high volumes of 'on-policy data' gathered through direct environmental interaction, logging edge cases, and recording safety critical actions. Mass-market automotive fleets provide this commercial data flywheel at scale, an advantage missing in non-automotive robotics.
While Wayve conducts exploratory research into humanoids, bipeds, and robotic manipulators, Badrinarayanan points out that those platforms remain hampered by small physical footprints and heavy reliance on synthetic training data. However, the simulation engines and training pipelines built for automotive physical AI are designed to translate directly into broader robotics. Wayve also plans to test models inside its OEM partners' automotive factories, accessing thousands of deployed industrial robots and manipulators.
Multi-City Testing and the Deployment Roadmap (15:11)
Wayve initially leveraged the dense, narrow, and irregular streets of London under permissive UK testing frameworks to build a rich training signal. This complexity enabled its models to adapt to US road conditions within months. To bridge foundational UK AI research with Silicon Valley's autonomous vehicle systems engineering, Wayve expanded into the Bay Area, alongside establishing hubs in Germany and Japan. Operating in San Francisco allows the company to benchmark directly against local players like Waymo and Zoox on the same streets.
Wayve's commercial rollout splits into two main streams: integrating supervised L2++ urban autonomy into consumer passenger cars for Nissan and Stellantis across North America, Europe, and Japan; and expanding public ride-hailing trials. Supported by a commercial testing license in London, Wayve partnered with Uber to deploy supervised autonomous Ford Mustang Mach-E vehicles, with plans to trial Nissan platforms in Tokyo and scale supervised robotaxis to 10 additional cities.
Badrinarayanan maintains that physical AI is rapidly moving into real-world production due to aggressive demand from automakers and advancements in generative world models. However, the operational transition from supervised L2++ driving and supervised ride-hailing pilots into fully uncrewed, unsupervised commercial operations across diverse regulatory jurisdictions remains an ongoing milestone.
[0:04] Vijay Badrinarayanan: The fundamental insight behind SegNet was that a data-driven, end-to-end learned approach delivered a "night and day" leap in performance compared to the prior ten to fifteen years of classical computer vision.
[0:04] Vijay Badrinarayanan: Autonomy is fundamentally an intelligence problem that must scale with compute, data, and learned representations rather than hard-coded heuristics.
[7:16] Vijay Badrinarayanan: the only path to true safety and generalizability is addressing both the level of autonomy and the geographic diversity of deployment at the same time.
[8:49] Vijay Badrinarayanan: the largest stable base of robots which are available in the world today
[8:49] Vijay Badrinarayanan: Crucially, developing physical AI requires vast amounts of "on-policy data"—data collected while the model is actively interacting with the physical environment, logging its safe actions, edge cases, and failure modes.
[15:11] Vijay Badrinarayanan: Navigating London's chaotic, narrow, and unpredictable urban core—supported by permissive testing regulations—provided an exceptionally rich and dense training signal.
[20:15] Vijay Badrinarayanan: physical AI is real.
[20:15] Vijay Badrinarayanan: way bigger than the chatbot.
Original transcript
[0:04] Uh I'm one of the Financial Times San Francisco
correspondents cover autonomous vehicles,
[0:08] data centers, robotics um for us from the
west coast and well VJ is a VP of AI at Wave
[0:17] uh who we're big fans of at the FT having both
been based in London. uh he leads their Bay Area
[0:23] team uh sort of focused on learn intelligence and
he's behind which is fundamentally behind a lot of
[0:29] the company's autonomous vehicle technology.
So his work spans representation learning,
[0:35] multimodal modal foundation models. Um and before
he joined Wave, he led a team uh working at Magic
[0:42] Leap um and previously worked with Waves founder
Alex Kendall uh at Cambridge on SegNet, which was
[0:49] a influential early system for using deep learning
to understand road scenes. now well he's based
[0:55] on the west coast for Wave building out a team in
the Bay Area and you know I mean VJ like kick off
[1:03] tell us tell us more about Wave tell us more about
what you guys are doing thank you thanks for the
[1:08] introduction Rafe um Wave started in 2017 um at
Cambridge our founder CEO Alex Kendall uh the tail
[1:17] end of his PhD is when he decided to start Wave
um and the mission of Wave is to build the most
[1:25] generalizable intelligence to serve autonomous
systems starting with cars and then expanding into
[1:32] other forms of embodiment when the time is right.
Um, an interesting anecdote. I was an academic at
[1:39] Cambridge University and the mentor for his PhD.
And Alex and I, we got working in the 2013 2014
[1:47] era of deep learning quite quite early on. And
we worked on this neural network architecture,
[1:54] an end-to-end neural network architecture called
Signet, which uh gained a lot of popularity at
[1:59] that point in time to solve the image segmentation
problem. And the core thesis and learning from
[2:05] that was when we when we got the results for
segmentation it was really night and day between
[2:11] a datadriven end to end learn segmentation
approach and what we all did in the past 1015
[2:18] years. So that really laid the foundational
principle to start wave. If you want to build
[2:25] generalizable intelligence for autonomous systems
it has to be end to end. It has to be datadriven
[2:30] and intelligence needs to scale with compute and
data and the right architecture print methodology
[2:36] of course and that's how wave started and to
this day that's the principle which guides us
[2:43] um today wave is not just um a leading technology
uh provider for generalizable intelligence for
[2:51] for various levels of autonomy from the very
early L0 to L5 stages uh but also a business we
[2:59] serious commercial traction at this point. Um, and
we've built we we built not just the foundation
[3:05] models for physical AI but also all of the other
important models which surround it which allow you
[3:13] to evaluate um proof safety um you know augmentata
etc. So these are world models like Gaia which
[3:21] perhaps seen this week and I mean what fascinates
me often about Wave I mean we we look at Whimo in
[3:28] in San Francisco and across the US we've got Tesla
testing out robo taxis on the road as well zuks
[3:34] here in Vegas I'm sure people if they can manage
to sort of hail the zuks which is quite tricky
[3:40] um wave is about to deploy commercial services
as well um so talk us through that so this week
[3:47] wave secured a license in London. It's planning to
deploy vehicles on the road there with Uber. Walk
[3:56] us through the commercial picture for Wave. We'll
do um so we need to take a step back and and and
[4:03] look at where Wave sits amongst all of these um
players. So as I mentioned there's various levels
[4:11] of autonomy uh from you know no driver assistance
to um supervised ADAS to robot taxis and beyond.
[4:19] Um now companies like Whimo they're deep into
the robot taxi deployment and scaling city
[4:25] after city. Um Tesla has you know launched its FSD
products. These are supervised autonomy products
[4:33] um fully vertically integrated for their cars
for their sensor stack etc. But if you look at
[4:39] what happens in the real world, there are 100
million cars which are produced every single
[4:43] year. And Waves has positioned itself to be
the the intelligence layer, the generalizable
[4:48] intelligence layer for all of these automakers.
So we don't build the hardware ourselves,
[4:55] but we build intelligence because we've always
believed autonomy is an intelligence problem.
[5:00] You need to you need to build a foundation model
and distill it so it works on various sensor
[5:05] configurations, various robot software types,
various types of onboard compute. Right? So
[5:10] this is is really the the distinction between what
Wave is um what Waves mission is versus um these
[5:17] um other companies. And one of the things um you
know from from a commercial traction perspective,
[5:23] we've gone deep into supervised autonomy. We
have announced partnerships with Nissan last
[5:28] year uh and now expanding to nearly 90% of their
um production which is you know in the millions
[5:34] and with Stellantis more recently and there's
more more to come this year. Um but equally we've
[5:40] always believed that you know this intelligence
one single intelligence needs to be able to serve
[5:47] various levels of autonomy. You just tap into this
based on scale of compute data and your ability to
[5:53] evaluate and prove safety. you could be building
um you could tap out and have an L2++ system
[6:00] uh a point-to-point supervised driving autonomy or
if you go further along uh it'll be a more robot
[6:05] taxi kind of service. So with this particular
announcement uh about having a license so we
[6:14] are partnering with Uber um and launching
an autonomy service a supervised autonomy
[6:19] service where the general public can actually
come and experience a technology um and this
[6:26] you know as as I pointed out we we don't build
cars ourselves we we're always in a partnership
[6:30] model with major automakers and who have this
ambition to build robot taxi platforms um and
[6:36] eventually this technology will get into those
high levels of autonomy platforms. I'm curious,
[6:42] I mean, you obviously you mentioned that a large
part of that deployment is going to be sort of,
[6:47] you know, bringing the public uh with you in a
sense. Obviously, we've had lots of focus on LLMs,
[6:54] but it feels like autonomous vehicles really do
sort of confront the public in the streets that
[7:00] they live in. Um, we've seen problems with
Whimo in San Francisco and other cities.
[7:05] I'm just curious, I mean, when you're thinking
about public perception around this technology,
[7:10] you know, what what is the approach there? How
how are you going to sort of scale appropriately
[7:16] in some of these markets? Yeah. Um so this this
deployment with Uber is um is a first step for
[7:25] us and of course safety comes first for us. Um so
we will go through um you know this is much more a
[7:33] supervised um autonomy experience uh you will you
will find um and one of the things we have always
[7:42] believed in is to build the most generalizable
safe intelligence you need to address both the
[7:49] scale of deployment and levels of autonomy
simultaneously. uh if you look at what other
[7:54] players are doing, it's much more on a you know
OD constraint and operating domain constrained
[7:59] um deployment but what we believe is the safest
thing to to achieve the level of safety required
[8:03] you need to deploy very broadly. That's something
we did uh for example um the last last year or so
[8:09] where we tested in about 500 cities in the world.
Um so our our foray into higher levels of autonomy
[8:17] will be supervised but we will quickly scale to
many many many cities through this partnership
[8:23] model um with a network operator like Uber
and various auto auto manufacturers and it
[8:30] also feels like autonomous vehicles are almost
a stepping stone I mean I think a large part
[8:35] of ways focuses it's leveraging this ability to
actually deploy technology in markets in the form
[8:41] of AVs and then looking at other domains as well.
And so I'm curious how you'll straddle the two.
[8:49] Yeah. So this is a really interesting and very
fundamental question, right? Um when people think
[8:55] about robots, you never think of cars as robots.
Um but the question is why not and I think people
[9:01] should and we took that very seriously. you know
about the 2017 2018 era when we started working
[9:07] on you know end to-end models which take um sensor
data in and actions out the first types of robots
[9:15] we wanted to address were cars but for the very
simple reason that they're the the largest stable
[9:21] base of robots which are available uh in the
world today um everywhere and increasingly all
[9:27] of these are softwaredefined vehicles you have the
right sensors you have the right compute where you
[9:32] can actually run serious sized models um on board
right so that's and the other thing is one of the
[9:43] one of the you know the the critical things you
need to build physical AI is data and to collect
[9:51] data it's not just about you know passive data
collection you need to collect what we call on
[9:55] policy data where you're actually deploying
models in the real world in a safe way and
[9:59] then collecting you know uh both the successes
and failures of models But to do that at scale,
[10:05] you need to intersect this with serious business
value. Um, and all of this comes together in the
[10:11] automotive world really well, except that you need
to have a very generalizable intelligence. Not not
[10:17] from just a pure model perspective, but also
from the robot software perspective. It has to
[10:22] generalize across sensors, across various types of
interfaces which partners have, various types of
[10:28] embedded compute which partners have. Once you do
that, you create this virtuous data flywheel which
[10:35] then powers your intelligence. So all of this kind
of comes together nicely in the automotive world.
[10:40] Um that said, there's been some serious progress
in beyond automotive other types of embodiment,
[10:47] humanoids or bipeds or arm manipulators, etc. So
we're we're definitely invested quite deeply in
[10:55] that as well, but we don't think that is yet ready
for prime time. the you know deployed base of
[11:01] these kinds of robots are still very very small.
Um and while there's quite a bit of research
[11:08] going on on let's say synthetic data etc to train
they're still in it's still in the research phase
[11:14] but we are working on these things and we believe
that you know there's a lot of carry in terms of
[11:20] our knowhow into these. Yeah, it's interesting,
right? We're in a market at the moment where huge
[11:26] amounts of capital are flowing into, you know,
a select range of startups, startups that are
[11:32] talking about generalized robots, but they're
lacking in lots of data that they're really
[11:37] struggling to procure that. I mean, some of
them are talking about using data from YouTube,
[11:41] uh, which fundamentally [snorts] is probably
not going to get them very far. And so you
[11:46] know when you talk about the transferability
of the data from the AV deployments into say
[11:52] robotics or other other form factors I I just
wonder I mean how much augmentation will you
[11:58] need to do with say you know will you need to to
effectively start up warehouses of robots that
[12:03] you're training on on tactile tasks as well?
I mean and how distant is that for wave? I
[12:10] think I think there's much more synergy. I mean
from from a data perspective yes they're quite
[12:16] distinct and you probably have to create this the
right um you know the right baseline models which
[12:23] enable you to deploy this you know in in the
field at some scale and then create the data
[12:27] flywheel just like what we did with um with with
the auto manufacturers. Um but where the carry
[12:35] is much more significant is the core technology
itself the training algorithms the simulation
[12:41] etc etc where we've invested a good part of a
decade into building something very specific
[12:47] and very very essential for evaluating end to end
models which is what all of the physical AI today
[12:53] the frontier um scale um the frontier level is
uh today so we need to so that part is where I
[13:02] there's a lot of carry for us. I mean, do you feel
like in a way that that that road map that you
[13:08] sort of laid out here is distinct from other
operators in the space? Is it different from
[13:15] Whimo's approach and and what they plan to do? Is
it different from some of what we're seeing come
[13:20] out of the frontier labs like Open and Antropic?
OpenI obviously have their own robotics team.
[13:25] uh so does enthropic or different from say like
Travis Kalanick and atoms his new uh venture which
[13:32] is also exploring I think industrial physical AI.
Yeah, that's a that's a very interesting question.
[13:39] It's um it's still kind of early stages when it
comes to general physical AI. So it's it's hard to
[13:46] get you know certainly um say where this will all
sort of meet. Um people have you know people have
[13:55] gone into um just buying a ton of robot setting
up a warehouse and and start doing this kind of
[14:00] experiments uh which I guess is an obvious
start but how does that really translate to
[14:06] um a realworld application I don't think anyone
yet made that leap and one of the things we
[14:13] believe we are in an advantageous position to to
exploit is already existing customer relationships
[14:18] and they are big manufacturers uh with huge huge
number of you know um robots, manipulators etc. in
[14:25] their factories. So I think we have an edge that
way to actually broaden the scope of our offering
[14:32] and of course data collection comes there as well.
Interesting. Obviously you guys are deploying in
[14:38] Japan at the moment, right? Um I mean I wondered
actually if you could walk us through some of the
[14:44] team building that you're doing on the West Coast
as well and and also why that's necessary. I feel
[14:48] like for for people who aren't familiar with
Wave, a lot of the focus and and emphasis for
[14:53] for several years has been, you know, this is
a a European startup that raised, you know, a
[14:59] remarkable amount compared to many other European
startups and also uh luckily for us in London is
[15:04] not a fintech uh which was like the only startup
type that we would get for like the past decade.
[15:11] But talk to us about what you're doing on the west
coast. Yeah. So this is kind of um rewinding back
[15:19] um you know I think it was it was a really
great idea to start wave in in London um in
[15:26] Cambridge and then quickly we moved to London
for this very simple reason that he's you know
[15:31] we focused a lot on urban autonomy first which is
which is really the very very challenging problem
[15:36] to solve this you want to be able to solve for
very generalizable intelligence in very chaotic
[15:42] environments and London is a great test bed for
it um from both a testing perspective and the you
[15:51] know the the regulatory policy which allows us to
do that testing um and data collection. So the the
[15:57] diversity the richness of the data we collected
really bootstrapped us very quickly um compared
[16:03] to other um competitors who you know who tested in
much more much less chaotic environments. So from
[16:10] an end to-end model development perspective that
was a great first step. Um and from that point on
[16:18] so if you if you you know you ask me why you know
why the US obviously one is the the business case
[16:25] um huge number of um cars are manufactured in
this country. Um and one of the interesting thing
[16:31] was how quickly we could actually transfer our
technology from the UK to the US. In within months
[16:35] we were able to adapt it and get to a stable uh
performance here. Um but one of the advantages of
[16:43] expanding into the US is talent u both hardware
and AI. Um and given the US you know even from
[16:52] the early 2000s has been invested in autonomy
there's just a huge amount of you know general
[16:58] u multi-disiplinary talent just available here
and that's how waves expanded into US and then
[17:04] into Germany and Japan uh for obvious reasons
um and that's how we we have now no longer a
[17:11] UK company but a global company yeah I mean walk
me through some of the pools of talent that exist
[17:17] in the US that you were struggling to f find in
the UK and elsewhere as well. Yeah, I think so.
[17:23] the two there's very specific talent like you know
uh robot software engineering or um just hardware
[17:31] um which is plentiful in in in the Bay Area um
due to past investments in autonomy of course
[17:39] um but AI as well I think you know when it comes
to AI talent in general it's my own personal view
[17:46] is it's best to expand as much as possible there's
great AI talent everywhere UK is really a hot bit
[17:51] of it uh but we were able to complement it um and
you know deal with the pressures the competitive
[17:58] pressures as we expanded that vi to the US. Yeah.
And then obviously another element of this is
[18:05] you're able to deploy vehicles on exactly the
same roads as some of your leading competitors,
[18:11] right? We now see Wave on the road in San
Francisco testing vehicles right alongside Whimo's
[18:17] deployments right alongside Zuk's deployments. How
important is that as well in terms of providing
[18:22] a barometer of comparison with with folks who are
either more established or are attached to some of
[18:28] the you know larger tech companies? Yeah, I think
I think it is very important uh one to actually
[18:35] you know um benchmark our technology in some um
implicit way to to know where we stand of course.
[18:44] Um, and it's also really important for for our
customers to actually experience this technology
[18:49] because we're so focused on generalizable
intelligence. We have to demonstrate this across
[18:55] many different operating domains. Um, and each
of these cities are, you know, quite unique. Um,
[19:01] London is um, different to San Francisco I
would say in terms of the challenges. Um,
[19:07] yeah. And then I mean thinking about some of the
other markets that you might look to deploy and
[19:12] I mean what what are the next steps? So you've
got London coming this year. Um you're obviously
[19:17] testing in Japan because you have a partnership
with Nissan testing in San Francisco. I mean where
[19:22] else are you currently sort of eyeing where do you
feel like there is some value ad to wave in terms
[19:28] of deployments? Yeah. So the the two different
streams. So the the more supervised L2++ urban
[19:36] um and hybrid driving technology which will go
into the Nissan and Stellantis vehicles, they're
[19:42] focused on North America, Japan, and Europe. Um
when it comes to experiencing our intelligence
[19:50] on Uber's network uh in the supervised robot taxi
service, uh we're slated as you pointed out to to
[19:57] launch that in London and Tokyo uh this year. In
London, it'll be on our Maki powered Maki vehicles
[20:03] and in in Tokyo, it's going to be the Nissan
vehicles uh this year and but we expect to get
[20:09] to 10 more cities next year. Yeah. Uh we haven't
yet we are not yet ready to release the names, but
[20:15] it's it's going to expand pretty quickly. Yeah. I
mean, it's interesting. I mean, you mentioned the
[20:19] Mackie vehicles and the Nissan vehicles. I mean
it feels like the technology is quickly sort of
[20:25] transferring between different hardware platforms
as well. Yes. Um which is really encouraging. I
[20:32] mean I think the the sort of final question for
me I I don't know if there is anything you want
[20:37] to sort of just finish on. Yeah. I think I would
sort of you know I don't know if people really um
[20:45] have got this but physical AI is real and I think
the it has its own unique challenges but there's
[20:54] been a number of breakthroughs in terms of both
modeling um simulation etc just you know I would
[21:02] encourage people to look at uh Gaia 4 which is our
latest simulation engine it's really fascinating
[21:07] what you can do in terms of eval EV evaluating
these uh models, policies, validate them for
[21:14] safety, reliability um and the frontier of and
it's moving so quickly. Uh I think people are
[21:23] focused so much on LLMs and agentic AI and things
like that, but this is coming and this is real
[21:30] and the the interest from auto manufacturers for
example is just intense. Um so it's it's a it's a
[21:37] great space to be in. Um, watch out. Great. It's
gonna be way bigger than the chatbot. Awesome.
[21:43] Right. Well, thanks VJ very much. Thanks to all
of you for sisting in on it. Great. Thank you.
