The autonomous vehicle industry spent a decade architecting complex, modular systems with separate teams for perception, prediction, and planning. That approach produced brittle systems that worked in geofenced, sunny suburbs but crumbled under the diversity of real-world driving. Alex Kendall and Wayve took the opposite bet: replace the entire stack with a single, end-to-end deep learning model trained on vast, diverse data. In this talk, Kendall traces the journey from Cambridge research to commercial partnerships with Nissan and Uber, making a rigorous case for why embodied intelligence demands a general-purpose solution. He explains how their AV2.0 model learns to drive directly from sensor input to motion plan, sidestepping the hand-coded rules and HD maps that limited previous systems. The result is a system that has demonstrated safe driving across 500 cities globally. For practitioners building in robotics or autonomy, this is a masterclass in why simplicity at the architectural level—one model, one objective—can unlock robustness that modular engineering cannot match.
Key Takeaways
- The AV2.0 architecture collapses perception, prediction, and planning into a single neural network that outputs motion plans directly from sensor data, eliminating hand-coded interfaces between modules.
- General-purpose driving capability requires training on diverse, global data rather than optimizing for a constrained operational design domain like Phoenix suburbs.
- End-to-end learning compresses intelligence into a single inference stack efficient enough to run on mass-scale automotive hardware, making deployment economics viable.
- The same data-driven paradigm that powers next-token prediction in large language models applies to driving: a unified model trained on broad data outperforms task-specific narrow AI.
- Wayve's commercial path with Nissan and Uber validates that OEMs and fleet operators are ready to integrate learned driving models rather than modular rule-based systems.
Who should watch: Autonomy engineers and robotics architects evaluating end-to-end learned systems versus modular pipelines for safety-critical physical deployment.
Why This Matters
Wayve's end-to-end thesis mirrors the consolidation happening across AI—from monolithic transformers in language to unified models in robotics—suggesting that the era of modular, hand-engineered autonomy stacks is ending.
