Keep pulling the thread on Vincent Vanuc.
At Waymo, strict safety and regulatory constraints are handled by an explicit verification system that operates outside of the primary AI model, checking the AI's proposed driving plan for compliance.
The primary challenge for Waymo in scaling its autonomous vehicle operations is solving the "long tail" of rare and exceptional driving events, which become common occurrences at the scale of millions of miles driven.
Vincent Vanuc believes that the development of reliable, physically realistic world models would be a technical advance that could completely change the landscape for autonomous driving.
A fundamental challenge in creating controllable, physically realistic world models is injecting a deep understanding of causality, a problem that machine learning has historically struggled with.
Waymo's strategic approach to autonomous driving was to initially "over-sensorize" its vehicles to solve the difficult Level 4 problem first, with the plan to reduce costs later, contrasting with competitors who started with cheaper Level 2 systems.
Vincent Vanuc is convinced that the required performance bar for commercial Level 4 autonomous driving is not human-level, but rather superhuman-level safety and capability.
According to Waymo's safety reports, its autonomous vehicles are currently safer than the average human driver, showing a significant margin of fewer collisions and fewer reported injuries.
Waymo has started collecting data in Tokyo, Japan, marking its first international expansion and its first test of driving on the left side of the road.
A key unsolved problem in generalist robotics is the ability to generalize motion and skills, as current systems are not good at performing tasks that are fundamentally different from their training data.
A key conceptual breakthrough in robotics was reframing robot actions as a different form of language, which allows researchers to leverage the same architectures and techniques used for large multilingual models.
Vincent Vanuc states that "third-party imitation"—the ability for robots to learn by watching videos of humans performing tasks—is a critical capability that has not yet been solved by the research community.
Waymo has observed that scaling laws similar to those found in large language models also apply to its large models for autonomous driving behavior and perception, showing comparable log-linear growth with respect to data and model size.