Lecture, Conference Presentation
Sacha Arnoud, Director of Engineering, Waymo - MIT Self-Driving Cars
- Self-driving technology is projected to fundamentally reshape mobility by making it more accessible, affordable, and efficient, while altering perceptions of traffic, parking, and urban environments to allow commuting time to be utilized for other activities.
- Waymo plans to launch a product soon following the removal of safety drivers in November, leveraging a growing fleet in Phoenix to expand the Operating Design Domain (ODD) from constrained areas like Chandler to complex urban cores with slopes and fog, such as San Francisco.
- Technical development aims to transition from merely avoiding collisions to anticipating complex behaviors through deep semantic understanding, enabling the system to navigate roundabouts, interpret visual gestures, and distinguish true obstacles from reflections or sensor errors.
- The path to production involves a "90% down, 90% to go" effort, where the final 90% of work to achieve safety and reliability requires a 10x increase in technology capabilities, team size, sensors, and testing practices beyond a functional 90% prototype.
- Testing infrastructure relies on a three-legged program combining real-world driving, simulation capable of processing 2.5 billion virtual miles with 25,000 concurrent virtual cars, and a dedicated physical facility in Central California to address the "long tail" of rare and adversarial scenarios.
- Data acquisition targets an accelerating pace, moving from 3 million driven miles by May 2017 to 4 million by November 2017, supported by Google's infrastructure including TensorFlow and TPUs to train models on millions of labeled samples equivalent to 300 years of human driving.
- Sensor strategy employs a complementary redundancy of vision, radar, and LIDAR, utilizing cameras for dense semantic information and LIDAR for depth estimation to filter out specific data errors like smoke or reflections that trigger false positives.
- Future scalability depends on deep learning models that learn core principles of safe driving to generalize to an infinite parameter space rather than memorizing specific events, requiring low-latency embedded processing within the vehicle without data center reliance.
- Labeling operations are expected to evolve toward hybrid models using machine learning to generate labels with human oversight, specifically to handle complex visual nuances such as visually impaired jaywalkers or individuals on truck beds who are not standard pedestrians.