Brad Templeton Home
In October of 2010, Google announced they had an internal robotic car project underway. While still at the very early stage, the cars have driven over 140,000 miles on ordinary streets, moving along with other cars, cyclists, pedestrians and everything else.
This included a remarkable 1,000 mile stretch where the car never encountered a situation where there was a desire for its human supervisors to take over.
Human supervisors? For many reasons, including assuring the project operation on regular streets is legal, the Google cars always have a team of operators inside. One is sitting in the driver's seat, watching and ready to grab the wheel or hit the brakes at any time. Another is monitoring the computer systems on a laptop screen, assuring all is going well. The "driver" is told to take the wheel if they feel there is any safety risk at all. They are told not to hesitate, and so they do take over from time to time, even in situations where the car would have handled the situation just fine.
California officials agreed, when asked by the New York Times, that this style of operation was legal. As long as a human is responsible, and watching, this is more like an extreme version of the automatic systems already found in cars, such as following cruise-control and lane departure avoidance.
How did Google get to such a milestone? The main technique was to build upon the company's existing expertise in mapping. Team leader Sebastian Thrun, aside from having led the Stanford team in the Darpa Grand Challenges, had also led Google's Street View project which already had a fleet of cars scanning and photographing every street they could. They took it a step further, and had human driven cars with robocar sensors drive a wide variety of roads and get detailed maps and models of them, crunched in Google's large data centers.
The result is a centimetre accurate map of every street before the robocar ever drives it. These maps not only allow the car to know where every lane maker, curb, sign, tree and street light is, they allow the car to figure out very precisely where it is even without relying on a GPS, though the GPS makes it all happen faster.
Among the tough roads driven by the cars are Lombard Street in San Francisco, which is super-windy and constantly full of tourist cars and pedestrians, and the rugged and windy Highway 1 coastal road in California.
Google recruited a great team to build their cars. In addition to Thrun, Anthonly Levandowski, who first came to public attention when he tried to enter a motorcycle in the Darpa Grand Challenge, had already built a self-driving Prius and allowed it to drive him around San Francisco and Oakland, even before getting rulings of legality for the process. Christopher Urmson was a leader on the other top robocar team in the world, the CMU team that won the Urban Challenge and did very well in the desert challenges. All of this was backed by Google's hefty financial resources and a strong announced interest in autonomous vehicles and advanced transportation by Google's top management, including recent talks by CEO Eric Schmidt saying that it was a shame that "the car had been invented before the computer."
Google attained this milestone earlier than even the optimists were predicting, but it is not yet a car that would you would put on the road without supervision. The requirement that a human driven car pre-drive each road is not that large a burden, but it also means the system can't perfectly handle changes to the road situation, as might be caused by construction, detours, new traffic signs or even a police officer redirecting traffic. In such situations the occupant would have to be called upon to figure out what to do -- at least for the first car to encounter the change. The car would not run into things, but it might not know where to go, or far more seriously, where not to go. (The car does have cameras and LIDAR to handle certain dynamic situations, including changing traffic signals, jaywalkers, and even other vehicles running red lights.)
So there's much, much more to do. But it is quite impressive what has been accomplished so quickly, and with only modest amounts of money.
Disclaimer: In January of 2011, after this article was written, I began consulting for the Google Robocar team. Prior to that I had known members of the team as well as Google management for many years.
Google released a video of its car performing door-to-door operations, including entering and leaving driveways, going through a fast-food drive-through, and stopping at a strip mall, all with a highly blind operator in the driver's seat.
Videos can also be seen of the car giving high speed "race course" style rides in a closed parking lot.
Another sub-project disclosed in 2011 was an automated golf-cart system named Caddy. In this system, one could summon a car from a garage using a computer or phone, and it would come to you, and then drive you or let you drive it. These videos also give many other technical details of how the car operates.