Accelerating the pace of learning

Dmitri Dolgov
Published in
3 min readFeb 1, 2017


Despite today’s cars getting smarter and more advanced, road fatalities in the U.S. are on the rise. We know that human error is involved in 94 percent of all crashes. That’s why we’re working harder than ever to bring self-driving cars, that don’t get tired or distracted, to our roads.

In the last few years we’ve amassed over 2.5 million miles of autonomous driving on public roads. We’ve taught our cars to handle some of the most complex driving tasks (e.g., navigating construction zones), and how to interact with other drivers (e.g., merging into a lane during rush hour). We can handle unexpected situations, from cars driving the wrong-way down the road to horses crossing our path. It was this level of sophistication that enabled us to complete the world’s first truly self-driving ride on public roads at the end of 2015.

Now, as we get ready to bring fully self-driving cars to more people, we’ve accelerated the pace of testing, on our private test track, on public roads and in simulation. And on all these fronts we’re making progress.

One indication can be seen in this year’s California disengagement report, which was posted online by the CA DMV today. Disengages are a natural part of the testing process and occur when a driver takes manual control of a vehicle while it is in autonomous mode. Testing, including disengages, allows our engineers to safely add to our software’s driving skills, expand hardware capabilities, and identify areas of improvement.

Our report shows a marked improvement in our fully self-driving technology. Since 2015, our rate of safety-related disengages has fallen from 0.8 disengages per thousand miles to 0.2 per thousand miles in 2016. (Even as we increased our driving by 50% in the state — racking up a total of 635,868 miles — our total number of reportable disengages fell from 341 in 2015 to 124 for this reporting period).

This four-fold improvement reflects the significant work we’ve been doing to make our software and hardware more capable and mature. And because we’re creating a self-driving car that can take you from door to door, almost all our time has been spent on complex urban or suburban streets. This has given us valuable experience sharing the road safely with pedestrians and cyclists, and practicing advanced maneuvers such as making unprotected left turns and traversing multi-lane intersections.

We’ve been able to make dramatic improvements to our technology because we use each of these disengages to teach and refine our car (that’s why we set our thresholds for disengages conservatively). For each event we can create hundreds — and sometimes thousands — of related scenarios in simulation, varying the parameters such as the position and speed of other road users in the area. This allows us to do a more thorough job identifying the root cause of any disengage and resolving any problems in a robust way. Last year alone, we drove over a billion miles in simulation, with a focus on tackling some of the toughest situations people could encounter on the road.

By any measure we’re happy with our progress (full report here), but anxious to work even harder. With a hundred tragic road deaths in this country every single day, we’re motivated to work with governments and policymakers to deploy our technology safely and quickly. With eight years of careful development and testing, we’re optimistic that we’re closer to the day where fully self-driving technology can begin to make a difference.