Tuesday, 16 April 2013

WOW! - 96,000 Miles Without Safety Critical Intervention

I have just seen the ITIF Talk 10 April - At around 29 mins of the video, Chris Urmson of Google confirms that they have done in excess of 500,000 miles testing of the Google Self-Driving Car on the public roads, but the headline for me is that they have done 96,000 miles without safety critical human intervention.

Let's put that in perspective:
Extending Bryant Walker Smith's calculation from here:

The average person has a crash about once every 110,000 miles.
Previously Google told us 50,000 miles without intervention, which meant there was only a 27% probability that the car was as good at avoiding crashes as person.
But now, with 96,000 miles there is a 45% probability that the car is as good at avoiding crashes as a person. That is maybe 6 to 10 years for an average driver without a crash.....

Quite simply - Statistically the Google SDC in the 'ideal' driving conditions of the southern States is almost as good as the average person.

But I assume that we need at least 95% confidence that they will crash less than people - which we won't reach until 473,000 miles.

On DriverlessCarHQ we counted 32 unique license plates for Google SDCs - and we estimate that Google could possible be racking up somewhere around 1,000 miles/day of testing on public roads. We have no idea how much simulator testing they have achieved - but clearly any improvements seem to be feeding back well into the real world testing.

Whatever the metric that NHTSA require to certify the SDC safe for public use in the U.S., it is clear that Google are making very solid progress in their development program.

WOW! - Google are indicating that this technology will be in public hands in about 5 years time - so from when they first said this in Sept 2012 we can maybe expect 2017-2018.  That sounds very plausible given that their development program only started in 2009 and four years later they have a self-driving car that is already almost as good as a person statistically speaking.  

With another four or five years of development and the potential to rack up probably another 2 million miles of testing then this whole project looks very credible and on program based on my own interpretation of the data available.

As ever - what do you think?  Please add a comment and let me know.