Google robocar breakthrough


Just released in a New York Times article and sidebar about highways and video, Google has unveiled an internal robot car project that has attained a remarkable level of robotic driving sooner than I and many others had predicted. The project combined the talents of Sebastian Thrun, leader of the Stanley/Junior team that won the Darpa Desert Grand Chellenge and came a close 2nd in the urban challenge, and Christopher Urmson from the CMU team that won the urban challenge and did second in the desert, along with 15 other engineers.

Their remarkable new Prius-based vehicles have completed over 140,000 miles of human-overseen driving on ordinary highways and city streets, including stretches of up to 1,000 miles without the human overseer feeling any need to apply a safety correction. By having a human in the car ready to grab the wheel, and a 2nd person also monitoring systems on a computer screen, the robotic operation on city streets is generally appraised to be legal.

As an example of the human intervention, during the test ride with reporter John Markoff, the human controller took the wheel when a cyclist ran a red light in a "just in case" intervention. Later examination of the sensors showed the car had indeed seen the bicycle and would have been expected to avoid it had the human not taken over.

This legal ability to have supervised driving should help build lots of great test data for robotic cars. Developers can build tools to try to judge whether, when a human intervened, the robot would have done anything particularly different, and look for those cases and judge and correct them. It also means, as I have described earlier, that we can start building the "trillion mile test suite" with all the data needed to do extensive virtual tests on new software revisions and prototype vehicles.

The new robotic Prius also looks a lot slicker than Junior, which very much has the experimental vehicle aesthetic. I expect the high resolution LIDARs to also get smaller and cheaper with time.

Later this week I should get a chance to see these vehicles up close and ride in one for more commentary. These results should make a stronger demonstration of how practical the technology is, to spur development and the legal steps necessary to move towards deployment when appropriate safety levels are reached.

Google of course is not a car company, but Sebastian Thrun has been involved there for some time as a creator of the street view camera car, and Larry Page has a longtime interest in transport innovation. Anthony Levandowski, creator of the Ghost Rider motorcycle entrant in the desert challenge and the PriBot (an earlier robotic Prius which he allowed to take him around the Bay Area while he supervised) is also a Google employee and on the team. Early research in robocars has come from academic labs and small teams, and it's good to see Google get into funding groundbreaking work in the area.

Google is not a car company -- but it has become one of the world's leading companies in mapping an navigation.

In the long term, robocars should have a positive effect on society that exceeds even that of the search engine; this could become the biggest thing that Google does.


Are already coming in greater numbers, and it's even more common in electric car design.

But yes, as I may mod the article to say, Google is no car company, but it has become one of the world's top companies in the navigation space.

I think the commercial opportunities in robocars are still years away. I get the strong impression that Google is doing this simply because enough people there think it is very cool.

"140,000 miles with only occasional human control." How many times would they have crashed if they had trusted the computer? That's the only question that matters.

Or rather, would they have ever crashed if they had trusted their computer. Google's not claiming a ready to deploy vehicle here by any stretch, but a new milestone using a new technique. They are not the first to do supervised driving with human interventions, that has taken place since the 90s. But they are getting the human intervention number down towards that eventual goal of zero.

Well, not strictly zero. I think a likely target might be a billion miles between fatalities and 10 million between accidents. Which is 10-20x better than humans.

The human supervision is what makes this into a rigged demo. The question is: How rigged?

In that they've been very open about the presence of the supervisory drivers. They report a 1,000 mile trip without the human wanting to take the wheel, which is pretty impressive, though they also report the human taking the wheel on the NYT test ride. In this prototype stage, you don't want the supervisor to hesitate in any safety situation. In the long term you would like to be able to figure out, when a human took the controls, if the robot would have done the same thing anyway.

Add new comment