It’s All Simulated

8 million. That’s the number of miles Waymo (formerly part of Google) cars drive every day [1]. No, not in real-life — in a simulation!

Miles are important. The more miles driven, the better. Take, for example, a new driver who’s just gotten started on the road. In California, they have to drive a minimum of 50 hours before applying for a license; generally speaking, the more hours driven, the better they get at driving.

Self-driving cars work in an almost-identical fashion. Waymo (the DMV) makes sure their self-driving car software (the new driver) has gotten as much experience driving as possible, a safety precaution to improve the car’s driving abilities before it is released into the public without supervision. The more miles of data Waymo can collect on roadways, the better its cars become, beating out the rest of the competition (Uber, Cruise, Zoox, etc.). In a neural network-based world in pursuit of full autonomy, data reigns supreme.

This brings us to 3D, virtual simulations. A previously underutilized tool, simulations are now key components to quickly training and testing self-driving cars. Instead of deploying 100 real-life cars to collect data, thousands of simulations can be run simultaneously, generating new worlds and complicated scenarios for the cars to learn from. Even better, these simulations can be designed to look almost-identical to real life.

Finally, one of the biggest advantages to using simulation for self-driving cars is in the realm of public safety. As evidenced by the 2018 Uber accident in Tempe, Arizona, bystanders and the general public can accidentally become involved in dangerous situations that can lead to casualty. When a company wants to test their car algorithm’s ability to prevent collisions with pedestrians, testing in real-life New York City may not be the best idea — instead, modeling a high-traffic urban area in real-life can prevent accidents. Are the pedestrians in the test scenario staged to be there, and knowledgeable of the test? Still, it is better to play it safe. 

You get the point. From dodging pedestrians to swerving to avoid a giant spool rolling down the highway, simulations are the easy, efficient, and safe way to go.

[1] https://www.businessinsider.com/waymo-engineer-explains-why-testing-self-driving-cars-virtually-is-critical-2018-8 [2] https://www.forbes.com/sites/davidsilver/2018/07/26/waymo-has-the-most-autonomous-miles-by-a-lot/#ee40e9d7ee53