For Google’s Waymo driverless cars, learning how to drive in Australia won’t simply mean remembering to stay on the left, it will also mean learning how to share the road with the worst Aussie drivers.
Waymo’s cars don’t just rely on a hard-coded list of road rules, they also take advantage of Google’s machine learning to hone their driving skills by paying attention to the world around them. The more time Waymo cars spend on the road, even with a person behind the wheel, the better they become at anticipating what’s about to happen next. Waymo’s driverless cars learn more about the environment and other drivers as they go.Credit:AP Waymo is yet to venture beyond the US, but even driving in different American cities presents challenges as the cars learn to cope with everything from the dust clouds to snow storms. Learning to drive somewhere new doesn’t just mean coming to terms with local weather conditions, it also means learning how to get along with local drivers, says Waymo chief technology officer Dmitri Dolgov.
“We’re not building a car, we’re really building a driver,” Dolgov says. “In any environment there are people behaving the way they shouldn’t — people run red lights, people drive the wrong way down the road — and you have to be ready for those reckless behaviours.
“As a first step we would deploy a fleet of vehicles with our drivers, just driving around a new city and observing how other people drive, so our cars can get better at predicting how people will behave in different situations.”
Watching how the locals drive allows Waymo’s cars to learn skills such as giving way to other cars, pushing into heavy traffic and coping when traffic lights are out.
Learning from experience even helps the cars cope with situations where drivers deliberately break the rules out of courtesy, such as the “Pittsburgh left” where — when the lights turn green — drivers choose to give way to cars turning left in front of them, to help ease traffic congestion. Waymo doesn’t need to teach its cars about such local habits, they figure it out for themselves. Waymo is trialling its robotaxi service in car-loving Phoenix, Arizona. Since 2012, Waymo’s testing program has extended to 25 US cities, with its fleet of cars clocking up more than 16 million kilometres on public roads. This is on top of the time the cars have spent on private test tracks and in simulators.
The autonomous car program shifted up a gear last year with the launch of the Waymo One ride-hailing trial in Phoenix, Arizona, picking up passengers with a human “safety driver” ready to take the wheel should the car encounter difficulties. This month, ride-sharing service Lyft is also deploying 10 Waymo cars in Phoenix, a popular city for autonomous car programs due to the generally clear weather.
Autonomous driving was one of the first practical applications to benefit from Google’s “deep learning”, using artificial intelligence to learn from its experiences and cope with new environments. AI assists with identifying objects — such as vehicles, pedestrians and animals — as well as predicting what they’ll likely do next.
Waymo cars rely on three types of sensors; LIDAR which relies on lasers for 3D vision, radar which relies on radio waves to cope with adverse weather conditions, and standard video cameras. Combined the three sensors produce a detailed model of the surrounding environment and even allow the cars to see in the dark, as well as cope with rain, snow, fog and glare.
Waymo is far from the only autonomous vehicle on the road, with several car manufacturers including Tesla in the race along with ride sharing giant Uber and technology companies such as Nvidia, Huawei and Baidu. Opinions vary when it comes to the technologies and onboard sensors required to ensure autonomous cars are safe on the road.
Most autonomous car platforms take advantage of LIDAR but Tesla chief Elon Musk recently declared the technology “expensive”, “unnecessary” and “a fool’s errand”. Tesla’s autopilot technology relies on cheaper ultrasonic sensors along with cameras and radar, as Musk aims to have a fleet of Tesla robotaxis on the road by next year.
“Having more sensors allows Waymo a lot of advantages, it’s easier to interact other vehicles and other objects in the environment,” says Waymo head of research Drago Anguelov.
“[Not using LIDAR] is a very big bet. It’s very, very risky and it’s not necessary.”