Original article was published by Dave Gershgorn on Artificial Intelligence on Medium
Actual Self-Driving Taxis Are Hitting City Streets
Waymo is officially launching a fleet of self-driving cars in Phoenix
OneZero’s General Intelligence puts the week’s biggest A.I. news into context.
Waymo, the driverless car company spun out of Google in 2016, is finally fulfilling its promise to bring truly autonomous cars onto city streets.
The company announced Thursday, October 8, that anyone in a 50-mile swath of Phoenix, Arizona, would be able to hail a fully driverless car in the “near term.”
This is undeniably a big step. Only 5% to 10% of Waymo’s rides so far this year have been fully driverless, according to Business Insider, for a select group of passengers who have signed NDAs. The other 90% to 95% of rides have been completed with a safety driver in the front seat who can stop the car from making a mistake that could kill someone.
The big step also comes with a big caveat: Waymo’s fully driverless feature is limited to a 50-mile area that the company has surely mapped to the millimeter. To give some idea of the amount of data each Waymo car collects, last year it released a dataset of the equivalent of what one car driving for 5.5 hours would collect, and it totaled nearly 2 terabytes of data. The company has been testing in Phoenix for two years with between 300 and 400 cars, according to TechCrunch, meaning it has potentially logged petabytes of information about the city’s landscape.
“Just like a human driver who has driven the same road hundreds of times mostly needs to focus only on the parts of the environment that change, such as other vehicles or pedestrians, the Waymo Driver knows permanent features of the road from our highly detailed maps and then uses its onboard systems to accurately perceive the world around it, focusing more on moving objects,” Waymo wrote in a blog post last September.
Going fully driverless, in addition to bringing Waymo closer to its goal of building an all-purpose driving algorithm, could be a way to reduce the risk for safety drivers, who have been forced to work in potentially unsafe conditions through the pandemic and wildfires, according to The Verge. Though the effect will likely be minimal, as safety drivers will also still be used for the foreseeable future wherever else Waymo is testing its cars, with or without passengers.
Waymo surely doesn’t publish all of its A.I. research, and it doesn’t collect every paper that its researchers publish in one place. But Waymo’s research lab heads, like Dragomir Anguelov, are typically added as authors to the lab’s most impactful work. So by tracking Anguelov’s publications this year, we can construct a picture of what’s happening in Waymo’s labs. What seems apparent is that the simplest and hardest problems of building a self-driving car have not yet been solved. Researchers are still tackling how to better track objects around the car — like pedestrians — how to predict where those objects will go, and how to get even more training data that explains the world to its cars.
Autonomous cars that are ready to travel more broadly won’t be achieved by one particular breakthrough. Rather, the slow improvement of basic, key systems are each small steps toward the incredibly large and complex goal of an all-purpose driving machine.
Here’s a more detailed look at what Waymo has been up to:
Predicting the trajectory of moving objects
Driverless cars need to be able to predict where other cars, bikes, scooters, or pedestrians are heading. In this paper, Waymo discusses how it can generate potential trajectories for objects around a car, then rank how likely they are to happen, so the car can plan its own route that doesn’t hit the object. It also takes the scene into context, meaning it analyzes signs or lights that might influence where the pedestrian or other car is going to move.
Tracking multiple objects
In order to predict where a pedestrian or car is going to go, first, the car needs to be able to keep track of all the other surrounding vehicles and pedestrians. Researchers say this problem is so hard because cars and pedestrians flit in and out of vision, and interact with each other in complex ways. This Waymo research not only focuses on identifying pedestrians and cars in cluttered scenes, but on recognizing them again when they leave and come back into view.
Synthesizing even more training data
It seems that petabytes of driving data simply aren’t enough. Waymo researchers are exploring ways to increase the effectiveness of data they already have, by using it to generate new, synthetic data. They do this by using the data from cars to build a 3D model of a scenario, and then looking at that scene from other viewpoints, as if the car was in different locations on the road. This new data can then be fed into the algorithms meant to detect objects, predict behavior, and plan the car’s own motion. Squeezing blood out of the proverbial stone could also allow Waymo to scale up its operations more quickly when it starts expanding to a new location by quickly repurposing the new mapping data into even more driving experience.