I’ll be sure to post a reply here when the article goes live so you can be sure to see it. :)

Yes there are several different flavors of autonomous, and I’ll talk about that in the article. Most of this stuff is discussed in more detail in the article but I didn’t want to fail to respond to your comment. I think this is an extremely important aspect of autonomous that doesn’t get much press and deserves to be part of a larger conversation.

The flavors of autonomy fall broadly into the “Google” model of detailed pre-mapping for complex urban environments, and the “Tesla” model of no-pre-mapping for simplified highway / interstate environments.

Both are children of the Carnegie Mellon model created by Red Whitaker that basically assumes no network will exist to provide services to support autonomy. In my opinion, these methods are a kludge designed to work-around the lack of a network. It’s like Qualcomm’s mesh-repeater walkie-talkies that they built in the late ’90s — similar to a cell phone, but limited, because of the lack of a network to support the device.

These two methods are robust to get us to level 3 autonomy, that is, highly autonomous with limited driver input, but these methods will fail to reach level 4, which is 5–9’s reliability for pure end-to-end autonomy w/o driver input at any point including first mile / last mile.

Pilots for highly autonomous vehicles are emerging currently, and the rate of learning matches Moore’s Law expectations, but vehicles are not just a silicon endeavor. Human-coded software doesn’t match Moore’s Law, nor do mechanical systems (vehicles) or structural systems (roads). So only one aspect of the environment can progress at an exponential rate, which is a gating factor to widespread deployment. We need innovations that resolve the mechanical and structural limitations to remove that gating factor and nobody is really working on those currently. (Other than us, but I digress.)

However these level 3 autonomy pilots do not demonstrate level 4 capabilities, as level 4 capabilities require support services that do not currently exist. This is similar to how the landline telephone network could give us cordless phones, but not cellular phones, as cellular phones required a new kind of network to support them. Currently level 4 development is highly constrained to very small geophysical areas that supply these additional network services.

Now, I’m no Google insider, I’m not even from the Bay area, and Google has been less than eager to talk about the architectural limitations of their programs, but I have contacts and pay close attention to press releases. This is a preface to explain that if I make an error in my analysis of Google it’s from a lack of information, not a lack of effort. I’m happy to be corrected if I make an error.

I understand Google has three cohorts of “autonomous” vehicles — the initial modified-VW fleet which has been deprecated, the modified-Lexus SUV fleet that is used for simplified highway / interstate testing, and the self-built two-seat pod-car fleet that is used for complex urban environment testing.

My comment about the 25mph limitations relates to the pod cars Google has been using to demonstrate the complex urban testing. Sure, their Lexus fleet can go highway speeds in a very controlled environment, but their pod cars can only go 25 mph in the complex urban environment that requires (in Google’s model) extensive pre-mapping. As I understand it that isn’t because the propulsion systems are incapable of going faster, but because the autonomous systems are incapable of supporting a faster rate of travel.

And then of course there’s the problem of scaling a highly manual process for converting Streetview point-clouds into a machine-readable database of localized navigation data. That’s another headache entirely. So we have both a front-end limitation for the complex urban navigation model (limited network support), as well as a back-end limitation (manual data manipulation). These are very difficult problems but we believe we have a method to solve them both at the same time.

Thank you for hearing me out, Robin, it’s been enjoyable conversing with you about this. We read about Zipcar in some of my business courses as an undergrad and I have long admired your work in that area, so I’m glad I have had a chance to interact with you personally.