This is the second in a series of articles about Waymo. The first can be found here.
There is a great term in software engineering; “Code smell”. 
This concept could be equally applied more broadly even beyond technology; cracks forming in walls, pets behaving strangely or the tide going out unexpectedly – signs that all is not good.
But how could Waymo (from Alphabet (from Google)) be possibly on the nose?
The zeitgeist suggests that Google is successful at everything it turns its attention to.
The reality is very, very different. So we need to ask ourselves is Waymo going to be the next Google Search or the Google Knol.
There are four areas that are troubling to me. This includes their:
Let’s first consider their testing strategy and subsequent posts will focus on the others.
Since 2012, Waymo has a “secret test centre” outside of Atwater in California. This 60-acre site is surrounded by Chainwire fence covered in black plastic to stop people peering in. It has been opened to the media only once or twice for carefully choreographed demonstrations. In October 2017, Waymo sent 50 reporters on a three-hour journey to the site and they all duly reported essentially the same story here, here, here and here. There is one notable exception; a remarkable piece of work from The Atlantic by Alexis C. Madrigal stands out from the crowd.
Despite the secrecy surrounding this installation, we will use Google’s own technology to peek through the fence. Let’s do a satellite-based photo-essay to map out Waymo’s journey, their challenges and dig for clues about what is going on behind closed doors (or in this case plastic covered Chainwire fence).
In 2013 a site was acquired near Atwater California near an airfield. (Interestingly, the Castle Air Museum is next door that features an SR-71 Blackbird.)
If you were tasked with creating an autonomous vehicle then it would seem reasonable that you should get a test track to safely train your software.
So Waymo creates a series of roads and intersections including a “freeway”.
Waymo learns about the shortcomings of their software stack so they make the track more complex. Using the data from Atwater, they can feed it into simulation software called “Carcraft” and run different variations (good article here). The have run 20,000 different situations that they call “structured tests”.
Cul-de-sacs, suburban streets and three different signalised intersections are also added. These problems are referenced in the Atlantic article. The two-lane roundabout freaked out a Waymo in Texas to such an extent that they had to build one specifically. How - will - it - handle - the - roundabouts - of- Brisbane?
Based on the testing at Atwater and supervised on-road trials in Chandler, Waymo is now confident to let some cars drive by themselves in the Waymo One trial. Almost no new changes are made to the Atwater facility with the exception of some strange lines on a circular track that look to be calibration markings or similar.
Presumably, feedback from on-road trials reveals that driveways are clearly a problem. Fourteen driveways are created to test any number of everyday situations one encounters driving around suburban roads. These were used as part of the demonstrations to journalists who were invited to the Open Day in October 2017.
There is also more evidence of ‘Waymos doing donuts’. This is very strange.
We have three ideas;
The on-road testing reveals that even more driveways are needed – 22 in total. Now this is really strange. If driveways are a problem, then will the strategy of “just add more driveways” really get you closer to a solution? I don’t think so.
Waymo has now recorded 20 million miles of autonomous on-road driving, and 15 billion miles in simulation supported by sophisticated modelling back at HQ via their Carcraft simulation. The car-cloud-model-car feedback loop is now closed and any further benefits from Atwater should be entirely inconsequential. At the beginning of 2020 it appeared that Atwater had been mothballed – along with the early drafts of this post.
But then in July 2019 (and this image was only recently released) the facility was teeming with Waymos. There are at least 28 of them in this image. Atwater is ALIVE!
This imagery is now incredibly high resolution. Well worthwhile having a poke around. Click on the image above. It is hyperlinked the test track.
Why are they doing even more testing on this track? What benefit could it be providing over and above what they are learning in the field?
The cost per day of having 28 cars driving around this track would be astronomical.
I think I know… Anyone got any theories? Add them to the LinkedIn comments.
But before we jump to a theory, more clues lie 676 miles to the south east around Phoenix AZ, where the Waymo trials have been since early 2017.
As an interesting aside, despite Waymo clocking up 20 million autonomous miles on public roads, guess how many they are clocking up per day now? Exactly 0. Coronavirus has closed them down.
Coming up next post; a satellite-based photo essay of the on-road Waymo testing area and what it reveals about their challenges.