DMV report shows self-driving cars are learning – fast

DMV report shows self-driving cars are learning – fast
Waymo logged more than 600,000 miles of self-driving tests in 2016
Waymo logged more than 600,000 miles of self-driving tests in 2016
View 1 Image
Waymo logged more than 600,000 miles of self-driving tests in 2016
Waymo logged more than 600,000 miles of self-driving tests in 2016

Among all the questions about self-driving cars, one of the most serious relates to safety. No one wants to put their lives into the hands of a driverless car with anything less than bulletproof reliability, and a high-profile autonomous accident could set the breed back by years, forcing the likes of Google to subject their cars to millions of miles of testing. Those miles are starting to make a difference, too, with new figures showing the number of self-driving disengagements dropping from 2015 to 2016.

Disengagements, or the driver taking over from self-driving software, are one of the most important metrics when it comes to the testing of self-driving cars. Google (and its Waymo self-driving project) describes them as a "natural part of the testing process," because they help identify situations where the self-driving software didn't know what to do, forcing the driver to take control.

Based on the latest disengagement reports submitted to the Californian DMV, the number of situations confusing autonomous systems is dropping at a rapid rate. Last year, Waymo cars covered 635,868 miles (1,023,330 km) and reported just 124 disengages, compared to 341 disengages in just 424,331 mi (682,895 km) during testing in 2015.

That's a significant improvement, an improvement which Google says has come despite the cars being tested in tricky urban situations, from sharing the road with cyclists to handling multi-lane intersections. When the car does hand off to the driver, the team says it can extrapolate that data to fit hundreds of different scenarios in simulation, adding billions of miles worth of valuable knowledge to the program.

Google isn't the only company making big strides in its testing program. Ford suffered just three disengagements in 590 miles (950 km), two of which were caused by cars overtaking at high speed during a lane change. BMW only suffered one disengagement on Californian roads in 2016, despite covering a not-insignificant 638 miles (1,027 km).

What does this prove? Well, self-driving cars still aren't ready for the big time. If there's no steering wheel in the car, and that's what Google is planning, there needs to be no chance of confusion or disengagement. Still, the progress being made is impressive, and shows the future of self-driving cars might not be all that far away.

Source: Waymo, DMV

Kevon Lindenberg
The number of disengagements is slightly silly when we look at how many they've had vs the amount of miles they've driven. What's really impressive is the (M)iles/(D)isengagement for 2016. The # of M/D & the lack of miles driven for Ford and BMW don't really sell me on their tech. Waymo 2016 | 5128 M / D Waymo 2015 | 1244 M / D BMW | 638 M / D Ford | 197 M / D
Another factor is where. Driving miles in urban environments is VASTLY different from driving along a nice freeway. But I cannot wait for the day a car does the driving, unless it's a nice road, and then only takes over if it sees an impending crash.
Bob Flint
Any "Disengagement" is not Fully Autonomous, even for sunny California. Any real statistics driving in the great white north, or any where else there is real weather, traffic, anything else that may cross ones path during driving in real conditions???
They should NEVER remove the manual controls from a car, I don't care if they get it down to 99.9999999999% reliability. Chit happens, and if the computer suffers a failure or slight lag (you know, like every other computer on the planet), what can you do? Just sit there and wait for death? I think people should be made to always control the car, and then have a computer intervene when they make a mistake. Current collision avoidance systems are good, and I think improving those should be the main focus.
Compared to the fatigued overworked Taxi driver swerving on the road with me in the back,... might have to flip a coin on autonomous or not...
Captain Danger
One thing I have never seen mention with autonomous cars is what happens when sensors start to go. My cars are all more than 10 years old and if a sensor goes south it is usually not a big deal. I can keep going. I would think that an autonomous car would have at least a magnitude of sensors and actuators more than the current generation of cars. for Liability purposes they would all have to be working perfectly. As soon as something goes wrong your stranded , esp if they are no manual controls. The repair bills on these things most likely will be ridiculous.