Waymo States There Is No Artificial Intelligence Shortcut in Autonomous Driving
Having over 15 years of experience in driverless vehicle technology and a driving history of 200 million miles, the company emphasizes that end-to-end artificial intelligence models and camera-only approaches fail to meet scalable safety standards.
Srikanth Thirumalai, vice president of in-vehicle software at autonomous vehicle company Waymo, stated that there is no artificial intelligence-based shortcut to achieving safe driverless vehicles and that end-to-end systems alone carry serious safety risks.
Safety Risk and Physical Artificial Intelligence
In statements to Axios, Srikanth Thirumalai expressed that even the most advanced artificial intelligence models with billions of parameters can hallucinate. Stating that there is no reboot option when errors are made in artificial intelligence systems operating in the physical world, Thirumalai emphasized that one must face the consequences directly.
In a published blog post, Waymo shared 10 artificial intelligence lessons derived from 200 million autonomous miles, pushing back against the perception in the industry that artificial intelligence offers a shortcut for safe driverless driving.
End-to-End Systems and Sensor Debate
While competitors such as Tesla, Wayve, and Waabi develop end-to-end artificial intelligence networks that convert raw sensor data directly into steering commands, Waymo stated that it has tested these tools but they failed to provide the necessary safety barriers for scaled deployment.
The company also argued that systems relying solely on cameras are insufficient. Stating that the combined operation of cameras, lidar, and radar technologies enhances visibility, Waymo conveyed that high-definition mapping also maintains critical importance for safe autonomous driving.