Waymo has said it cannot find a shortcut to autonomous driving through advances in AI alone, Axios reported on Aug. 26.

The report said Waymo has logged 200 million miles of driverless driving over more than 15 years, but has concluded that data and model performance alone cannot solve autonomous driving.

Srikanth Sirumala (스리칸스 시루말라이), vice president of onboard software at Waymo, said in an exclusive interview with Axios: "To deploy AVs safely at scale, better AI alone is not enough."

Sirumala said Waymo has aimed for "provably safe AI" since its launch, and that risks are growing as multiple rivals choose an approach that leaves decisions to a single end-to-end AI.

He said even top AI models with trillions of parameters still hallucinate. "Physical AI doesn't have a refresh or reboot button. You have to live with the results," he said.

Waymo previously released 10 AI lessons on its blog on Aug. 10 from building up 200 million miles of driverless driving.

Waymo began in 2009 as Google's self-driving project. Early on, it ran multiple specialised models separately, such as pedestrian detection and traffic light recognition, but gradually shifted to a system based on a small number of large foundation models. By contrast, companies such as Tesla, Wave and Waabi are going a step further, developing end-to-end neural networks that process raw sensor data as-is and output steering commands.

Waymo also tried that approach but concluded the safeguards were not sufficient. Sirumala said: "We confirmed that with a pure end-to-end system alone, we cannot meet safety standards at the scale Waymo operates."

Waymo also confirmed during model training that using cameras, lidar and radar together is much better for securing visibility than removing some sensors.

Sirumala said: "AI can only understand what it can see. If it can't see at all, there's nothing AI can do." Axios said the remarks were interpreted as directly targeting Tesla's strategy of pursuing full autonomous driving with cameras alone.

He stressed that high-definition maps are still needed, contrary to claims by some rivals, and said there is "no single solution" to reach safe autonomous driving.