Exclusive: Waymo says there's no AI shortcut to self-driving

Waymo, the autonomous vehicle pioneer spun off from Google in 2009, has spent over 15 years and amassed 200 million miles of driverless driving experience. Despite the rapid advancements in artificial intelligence, Waymo's vice president of onboard software, Srikanth Thirumalai, asserts that there is no shortcut to achieving safe, large-scale deployment of self-driving vehicles.

In contrast to competitors embracing end-to-end AI systems, Waymo believes that such models are not yet capable of meeting their safety standards at the scale they operate. Waymo's approach has been to gradually transition from numerous specialized models to fewer, larger foundation models. The company emphasizes that their success is built on a foundation of demonstrably safe AI, rather than relying on the latest AI breakthroughs.

Waymo's blog post highlights 10 key lessons learned from their extensive autonomous driving experience, including the importance of incorporating multiple sensor types, such as cameras, lidar, and radar, to provide comprehensive visibility for their AI models. Thirumalai dismisses the notion that superior AI models can compensate for insufficient sensing capabilities.

He also criticizes the notion of high-definition mapping as a necessary component, arguing that it is not essential for achieving safe autonomy. The article underscores the fact that Waymo's extensive experience gives its claims considerable weight, but it does not definitively resolve the ongoing debate about the most effective path to safe self-driving technology.