Sacha Arnoud, Director of Engineering, Waymo - MIT Self-Driving Cars

Waymo became its own company in January 2017, a year before this talk was given. This transition marked a significant milestone, allowing the company to move into a productization phase.

The main motivations include significantly improving safety (as human error causes 94% of crashes), increasing accessibility and affordability of mobility for more people, and enhancing overall efficiency by allowing people to use commute time more productively.

The initial goal was to assemble a vehicle with off-the-shelf sensors and determine if self-driving was a viable possibility. A key milestone was to autonomously drive 1,000 miles in Northern California through complex routes.

A major milestone was reaching a level of confidence and maturity in their system that allowed them to remove the safety driver. This was demonstrated for the first time in November, signifying that the system was deemed safe enough to operate without human oversight.

Waymo addresses this by focusing on '10x' improvements across the board: increasing technology capabilities, team size, sensor performance, and overall system quality and testing practices. This iterative and demanding approach is crucial for achieving the necessary reliability.

Early applications included analyzing Street View imagery to extract crucial mapping data like street numbers and names, improving Google Maps accuracy. A breakthrough was achieved in 2012 with a system that could detect and transcribe street numbers from facades.

Perception is the system responsible for building an understanding of the world around the car, using inputs from sensors and pre-computed map data. It involves detecting objects, understanding their semantics, predicting their behavior, and differentiating true objects from sensor noise or reflections.

Waymo designs its own sensors to ensure they are complementary in localization and capabilities, and to enhance performance beyond off-the-shelf options. This is critical for building a self-driving system they can trust.

Deep learning helps by providing a higher level of semantic understanding to distinguish actual objects from reflections, which can confuse simpler detection systems. By analyzing patterns and context, the system can overcome these sensor imperfections.

Key components include extensive labeling efforts for supervised learning, significant computation and robust infrastructure for training and inference, and sophisticated testing programs involving real-world driving, simulation, and structured testing.

Waymo has accumulated millions of miles in real-world driving, with simulation enabling billions of miles of virtual testing annually. They also operate a dedicated 90-acre testing facility to recreate complex and rare scenarios.