How End-to-End Learning Created Autonomous Driving 2.0: Wayve CEO Alex Kendall

Alex defines the fundamental shift from classical robotics (perception, planning, mapping, control components with HD maps and LiDAR) to Wayve's end-to-end deep learning approach. He explains how this was contrarian in 2017 but represents building intelligent machines with onboard decision-making rather than infrastructure dependence.

• •AV 1.0 relied on hand-engineered components, HD maps, and heavy infrastructure vs. AV 2.0's single end-to-end neural network
• •Wayve's approach focuses on generalization - one foundation model that works across multiple vehicles, sensors, and use cases
• •The architecture must be both functionally safe and run real-time onboard with limited compute
• •Early objections included safety concerns, lack of interpretability, and skepticism about AI capabilities

Kendall explains how world models serve as internal simulators that enable the AI to reason about what happens next. Starting with a tiny 100K parameter model in 2018, Wayve now uses GAIA, a full generative world model that simulates multiple cameras and sensors, enabling emergent behaviors like nudging forward at blind turns.

• •World models act as physical reasoning engines - simulating how the world works and predicting outcomes
• •GAIA can simulate multiple camera views and be prompted to control different agents in the scene
• •Emergent behaviors include slowing in fog, nudging forward at unprotected turns until visibility improves
• •World models are critical for handling complex multi-agent scenarios safely and smoothly

The discussion covers Wayve's approach to training data, emphasizing diversity across vehicles, sensors, countries, and scenarios. Kendall explains how world models generate synthetic data that doesn't replace real-world data but recombines it to improve learning efficiency, while unsupervised learning techniques identify anomalies and poor performance scenarios.

• •Training data reaches tens or hundreds of petabytes from multiple megapixel cameras, radar, and LiDAR
• •Diversity comes from partnerships across dashcams, fleets, manufacturers, and robot operators globally
• •Unsupervised learning clusters unusual experiences and identifies scenarios where the system underperforms
• •World models magnify learning by generating synthetic understanding and recombining real data in new ways
• •Data efficiency is critical - unlimited data would just be a lookup table, but efficiency drives faster time to market

Wayve demonstrates unprecedented generalization by deploying in 500+ cities globally without HD maps. The system adapted to Nissan's vehicle in Tokyo in just 4 months from first drive to media demonstrations, showcasing the ability to handle new countries, vehicles, and sensor configurations rapidly.

• •Deployed in 500+ cities in one year including London, highways, Europe, Japan, and New York
• •Tokyo deployment: 4 months from first drive to public media demonstration with Nissan
• •No HD maps required - the AI generalizes to new environments through learned intelligence
• •Each new market requires exponentially less data (hundreds of hours for US, much less for subsequent markets)
• •Safety and flow metrics remain uniform globally; utility (navigation, road signs, languages) requires more localization

Kendall explains Wayve's go-to-market pivot to partnering with automotive OEMs rather than building vertically integrated robotaxis. This approach targets the 88 million non-Tesla vehicles built annually, enabling native software integration without hardware retrofits for faster, lower-cost deployment at massive scale.

• •90M cars built annually - Tesla builds ~2M, leaving 88M opportunity for Wayve's OEM partnerships
• •Software-defined vehicles now have the GPU compute and sensor infrastructure needed for autonomy
• •Native integration avoids expensive hardware retrofits, enables global homologation, and reduces costs
• •OEMs want to move from driver assistance to eyes-off autonomy with liability transfer
• •Path to tens/hundreds of thousands of affordable robotaxis through manufacturer partnerships

While acknowledging camera-only can reach human-level performance, Wayve advocates for camera-radar-front LiDAR stacks (under $2000) to achieve superhuman safety. The industry has coalesced around this architecture with automotive-grade components, providing necessary redundancy to eliminate accidents beyond human error.

• •Industry standard: surround camera + surround radar + front LiDAR costs under $2000 (automotive grade)
• •Camera-only can reach human level, but additional sensors enable superhuman performance
• •95%+ of accidents caused by human error can be eliminated, but long tail requires perception beyond vision
• •Same neural network family adapts to different sensor configurations across OEM vehicles
• •99%+ of training cost is in the base model; personalization to specific sensors is highly efficient

Starting in 2021, Wayve developed Lingo, the first vision-language-action model for autonomous driving. This enables the car to not only drive but converse about its decisions, creating a chauffeur experience where users can talk to the system, request driving styles, and regulators can interrogate reasoning.

• •Lingo (2021) was first vision-language-action model - can see, drive, and converse simultaneously
• •Language pre-training improves representation power beyond imagery alone
• •Product features: chauffeur experience, natural language commands for driving style preferences
• •Introspection tool for engineers and regulators to understand and diagnose system reasoning
• •Runs on next-gen automotive compute like NVIDIA Thor (current onboard compute too limited)

Kendall sees mobility maturing faster than manipulation due to data access, supply chains, and hardware challenges like tactile sensing. Wayve's navigation foundation model will enable manufacturers building any mobility application, with benefits from experiencing multiple verticals making the model more general-purpose over time.

• •Mobility is 5-10 years ahead of manipulation (manipulation at 2015 self-driving maturity level)
• •Automotive provides largest robot data supply chain - competitive advantage for foundation models
• •System becoming general-purpose navigation agent: arbitrary viewpoint + goal → safe trajectory
• •Near-term expansion: consumer automotive, robotaxis, trucking, other mobility applications
• •Foundation model benefits from multi-vertical experience, improving generalization across use cases

Kendall identifies three key research areas: measurement systems with simulators that close the real-world gap, building generality through multi-modal alignment, and engineering efficiency for training/serving. He references AlphaGo's perfect simulator enabling Monte Carlo tree search as the goal for robotics.

• •Measurement is critical: fast regression detection, efficient simulators that close real-world gaps
• •Chicken-and-egg: perfect simulator solves self-driving (like AlphaGo with perfect game simulator)
• •Generative world models are compute-intensive but essential for measurement at scale
• •Multi-modal alignment for generality, especially for human-robot interaction and navigation
• •Infrastructure efficiency for training/serving is competitive advantage, not just 'sexy' algorithms

Speculating on AV 3.0, Kendall envisions taking intelligence outside individual cars through vehicle communication. With majority autonomous vehicles, they could coordinate to eliminate traffic lights, share sensor data to see around corners, though this raises cybersecurity and latency challenges.

• •AV 2.0 = intelligence on the car without infrastructure; AV 3.0 = intelligence beyond the car
• •Vehicle mesh networks could eliminate traffic lights through coordination
• •Shared sensing: vehicles communicate to see around corners, reducing onboard sensor needs
• •Opens cybersecurity, communication protocol, and latency challenges
• •Potential future: humans restricted from driving or relegated to recreational driving zones