Waymo Simulations Let Self-Driving Cars Rack Up 8 Million Virtual Miles Every Day

The simulator allows self-driving cars to practice situations they may not regularly encounter in the real world, Waymo explained in a Medium post. A major challenge for self-driving cars is ensuring that they know what to do in any given situation, but engineers can't guarantee the vehicles will be exposed to everything on real-world drives. That's where the simulator comes in.

Using sensors, Waymo constructs a detailed virtual world to test self-driving cars, and fine tune it to specific problems or scenarios based on real-world conditions. The software that controls Waymo's self-driving cars can then be tested in those scenarios thousands of times. Every software tweak can be tested right away, before being applied to a real-world vehicle.

Engineers can also create endless variations of a scenario through a process called "fuzzing." This allows them to tweak different variables, like the number of vehicles on a simulated street, or the speed those vehicles are moving. It can even simulate a motorcycle splitting lanes, or joggers veering into the street.

Once the system learns how to deal with a simulated scenario, this new "skill" is shared with all of the cars in Waymo's fleet. The update is then validated in real-world testing, both on public roads and closed test tracks. Waymo claims the majority of improvements to its autonomous-driving software now come from simulations, with cars running 2.5 billion simulated miles in 2016.

Self-driving cars need to be able to adapt to every real-world driving situation, and the only real way to do that is to write software telling them what to do in each of those situations. Waymo's simulator could speed up that process dramatically.