Overview
- Waymo published a blog and senior software interviews Wednesday summarizing 10 lessons from roughly 200 million driverless miles and explicitly rejected the idea that bigger AI models alone can deliver safe full autonomy.
- The company says large foundation models still 'hallucinate' in the real world and that pure end-to-end systems that map raw pixels straight to steering lack the transparency and guardrails Waymo requires for operational safety.
- Waymo detailed its sensor and software approach, saying fused inputs from cameras, lidar, and radar plus an independent onboard validation layer are needed to detect geometry, track velocity in poor weather, and check proposed trajectories against physics and traffic law.
- The statement functions as a calibrated critique of camera-only, supervised fleet strategies promoted by rivals and comes as Tesla is reported to hold an invitation-only Cybercab event on Sept. 3 that highlights the commercial pressure to scale lower-cost robotaxis.
- Waymo’s stance could reshape regulator and industry expectations because the company pairs technical claims with a large real-world record—over 15 years of development, more than 200 million autonomous miles, and hundreds of thousands of weekly driverless trips—which it says reveal edge cases simulations and supervised miles miss.