Interview, Fireside Chat
How Waymo Is Using GenAI to Build a Better Driver
- Waymo prioritizes a gradual, deliberate, and safety-guided expansion of autonomous driving technology to ensure broad access.
- The company plans to broaden commercial applications beyond right-hailing to include deliveries, trucking, and personally owned vehicles.
- Future partnership structures will be explored to increase access, leveraging existing models such as integration with Uber and Uber Eats in Phoenix.
- Go-to-market strategies will continue to iterate as the service moves beyond its early deployment phase.
- Deployment of the sixth-generation sensor stack is underway following the fifth generation, with a continued projection to utilize all three sensor modalities (cameras, LiDAR, and radar) for redundancy and capability.
- Improvements to the customer experience are being driven by user feedback on pick-up and drop-off locations to address the complexity of determining convenient spots in dense urban environments.
- Operations and fleet size aim to scale by an order of magnitude or multiple orders of magnitude to enhance service performance.
- The company intends to publish additional operational data consistently as services scale and grow.
- Scaling laws are expected to apply to autonomous driving models, favoring an approach of training large models and distilling them into smaller onboard versions rather than training small models from the start.
- Full autonomy is predicted to require more than just end-to-end or foundation models due to current limitations in handling hallucinations, explainability, and 3D spatial planning.
- A remaining "point zero zero zero one percent" of hard problems exists that requires additional work beyond state-of-the-art AI to ensure confidence for removing the driver.
- Performance standards are set by comparing autonomous performance against models of attentive, very good human drivers in specific scenarios.
- Simulation software continues to be built with realistic models of pedestrians, cyclists, and drivers to ensure statistical realism at macro and behavioral levels.