Waymo Safety Report — Confirms what I have been saying about Public Shadow Driving etc

Waymo released their report to start putting L4 vehicles in one geofenced area- Phoenix. The report confirms everything I have said all along. Something Waymo wasn’t even saying 6 months ago.

The report link — https://waymo.com/safetyreport/

Article on the report — https://arstechnica.com/cars/2017/10/5-things-we-learned-from-waymos-big-self-driving-car-report/

First — Props to Waymo for coming around and doing this right

They did not exaggerate their capabilities like almost everyone else

They told the public they have to geofence and after 8 years are starting with one city

They don’t use Public Shadow Driving for AI or L3 because they know it will harm thousands of people needlessly and never lead to an autonomous vehicle

· Geofence — Phoenix — when it doesn’t rain hard — Which means they don’t have sensor coverage or redundancy in bad weather. Which also means they don’t have vehicle scenarios driven in all those conditions

· News articles say they have a lot of redundancy. Yes — but not in bad weather. I think they are finding out LiDAR and the Cameras can’t get it done in bad weather. And the radar they use is not enough. I have been saying for a while 3D radar is needed.

· Phoenix has much simple street layouts than most places. Not a lot of road pattern variation. It’s mostly a grid that is very well lined with great signage etc.

Realizing public shadow driving and L3 (they call if handover) is unsafe and can never lead to an autonomous vehicle due to the miles that need to be driven and the expense. (One trillion miles and over $300B)

· When the vehicles are in trouble they have a “Fallback Mode” which has the vehicle driving itself to a safe place and help is automatically called. They appear to not want and handover. (I see no mention of that remote capability which was very concerning. Maybe they dropped it?)

Since Waymo is only this far after 8 years this means they didn’t realize they need to switch to simulation for a long time. They only have 3.5M miles driven and 2.5B simulation miles. They would have far more simulation miles if they started earlier. (They also have said they created their own simulation which also wasted time. I would also like to understand why so many simulation miles with so much work to do? Is it being calculated as a driving equivalent? I say this because most scenarios take place well within a mile. In simulation you need on drive the scenario length no more. Even with running them over and over for various reasons. If you assume just 10% is unique runs that would be 250M miles. at 1/4 mile per scenario we are at 1B scenario runs. If Waymo were just 10% done there would be 10B scenarios? Primary and their threads of variations. That seems extremely high.)

If this is where Waymo is, given their 8 year run, most of the others are well behind this. The only ones who are not would be the one who understood the need for simulation earlier. (I know this includes at least Toyota)

If Waymo is using actual advanced simulation, is working the hardest scenarios in parallel and is working the scenario matrix from the top down with other data sources in addition to data from driving and in an object oriented manner tied to the simulation they could be first to broader L4.

Miscellaneous

· They do not mention their black box can handle very bad accidents or fire. Or if their data transfers are frequent enough to make sure accident data is sent before the black box might be damaged. I do not see any discussion on creating a scenario matrix from sources other than driving or simulation. Isn’t really required so they probably don’t want to give that away. Sure hope they are augmenting that list top down from other data sources. Will save years of time.

A couple of my articles on the subject

Letter to Congress — Handling of minimum standards for Autonomous industry — This includes reasons to not use Public Shadow Driving or L3 and other significant issues in the industry

Who will get to Autonomous Level 5 First and Why

· https://www.linkedin.com/pulse/who-get-autonomous-level-5-first-why-michael-dekort