Waymo's investment in data mining, training cycle, and automation of the feedback loop is key to building a robust and generalizable autonomous driving system.
- Investment in frameworks and infrastructure for closing the loop on data mining training cycle
- Investment in feedback loop as a first-class object in the development life cycle
- Automation and low human engineering cost in the ML infrastructure
- Discovery of interesting long-tail examples through data mining and hard data example mining strategy
- Challenges in optimizing for both long-tail and average case distributions
- Unification and simplification of technology development and team organization to build a robust and generalizable core stack
Waymo has encountered various interesting long-tail examples such as construction zones, vehicles running stop signs or red lights, jaywalking pedestrians or cyclists, and even a drunk cyclist weaving through traffic with a stop sign on his back. These examples are brought into their data sets and expanded into a family of similar scenarios that guide the training and evaluation of their system. Waymo's investment in a simulator enables them to expand these examples and build a more robust and generalizable core stack.