Interview, Fireside Chat
How End-to-End Learning Created Autonomous Driving 2.0: Wayve CEO Alex Kendall
- Wave's Strategic Positioning: The company aims to become an "embodied AI foundation model" for global fleets and manufacturers rather than a vertically integrated robotics solution, seeking to amortize costs over a single large intelligence capable of rapid adaptation across diverse applications.
- AV 2.0 Architecture Definition: Wave defines "AV 2.0" as a paradigm shift from AV 1.0's hand-engineered, component-based stacks (perception, planning, mapping, control) reliant on HD maps and C++ codebases to a singular end-to-end deep learning neural network that operates without onerous infrastructure.
- Core Technical Philosophy: The approach prioritizes generalization and onboard intelligence, enabling robots to reason about unseen scenarios and edge cases (e.g., road workers, double-parked cars) without pre-defined rules or exhaustive manual coding of every possibility.
- Data Strategy and Diversity: To achieve generalization, the model is trained on diverse data from multiple sensor architectures (camera-only, camera-radar, camera-radar-lidar) across 500+ cities globally, utilizing unsupervised learning to cluster anomalies and drive curriculum learning.
- Sensor Fusion Stance: While acknowledging camera-only can reach human-level performance, Wave advocates for a "surround camera, surround radar, and front-facing lidar" stack (costing under $2,000) to achieve superhuman safety, redundancy, and the ability to resolve long-tail edge cases.
- World Models as Reasoning: The company utilizes "generative world models" (previously a 100k parameter network in 2018, now GAIA) to simulate multi-sensor environments, allowing the AI to learn emergent reasoning behaviors such as nudging forward on unprotected turns or adjusting speed in fog.
- Integration with Large Language Models: Wave released "Lingo," a vision-language-action model in 2022/2023 that integrates natural language for pre-training improvements, enables chauffeur-style interaction, and provides an introspection tool for regulators and engineers to query system reasoning.
- OEM Partnership Model: Rather than building proprietary hardware fleets, Wave partners with major OEMs like Nissan to integrate its software stack natively into mass-produced vehicles, aiming to scale to the 90 million annual car production volume to enable "eyes-off" autonomy.
- Regional Adaptation Performance: The system demonstrates strong safety and flow metrics globally but requires varying data volumes for "utility" adaptation (e.g., 100 hours of new data to reach 90% performance in the US vs. minimal data for Europe due to shared driving rules).
- Efficiency and Scale: The model leverages synthetic data generated by world models to magnify learning and improve data efficiency, allowing for rapid iteration and deployment without the prohibitive costs of purely real-world data collection.
- Future Roadmap (AV 3.0): Speculation points toward "AV 3.0" involving vehicle-to-vehicle communication and coordination to eliminate traffic lights and reduce hardware dependency, potentially creating a mesh network of autonomous intelligence.
- Talent and Culture: Wave recruits by combining a frontier AI research environment with the near-term product scale of the automotive industry, operating teams globally in hubs including London, Stuttgart, Tel Aviv, Vancouver, Tokyo, and Silicon Valley.