Autonomous Vehicles Won’t Revolutionize Maps

Joe Morrison
7 min readApr 8, 2021

Originally published in February’s edition of my monthly newsletterAnything you read here gets sent there first. To get articles like this in your inbox once a month, sign up here.

A lot of folks don’t realize that Elon Musk got his first big break as an entrepreneur building mapping software. He and his brother Kimbal started Zip2 in 1995, which provided phonebook-style business listings along with embedded mapsfor a little wow factor. Part of me is envious of the days when your business plan could be as simple as, “we’re gonna put the phone book on the computer.”

After raising a few million dollars in 1996, Elon was removed as CEO and replaced by Rich Sorkin (who is now the CEO of yet another mapping startup, Jupiter Intelligence). In 1999, near the height of dot-com hysteria, Zip2 sold to Compaq for $307M.

The moral of the story is simple: if you wanna get rich, start a software company focused on maps.

I was perusing Elon’s tweets a while back when I stumbled across an interesting response from a guy named Robert Scoble. Since encountering him, I’ve learned that Robert is a professional internet provocateur. He stumbles into interesting tech or tech-adjacent topics, and like a drunk relative toasting the happy couple at a wedding reception, proceeds to embarrass himself with a confidence that is truly awe-inspiring.

We got into a fleeting disagreement about whether or not Tesla’s mapping work was soon going to eclipse Google or Apple. It all began when I glimpsed this exchange:

Robert acknowledges a good point by his interlocutor, Ben Dover.

“Tesla will soon crush Google Maps,” Robert proclaims.

On the rare occasion I read something so ponderously stupid that it hurts me physically, I try to use it as an opportunity for reflection. So, I got to thinking… why do I find this claim about the future of mapping supremacy so objectionable?

Below, I talk it out.

Imagining the Future of Autonomy

I’ll be honest. Elon Musk bothers me.

One of the reasons I’m uncomfortable with Elon as a role model is the way he has poisoned public discourse about autonomous navigation. He sells Tesla customers something called “full self driving capability” for $10,000 as an optional upgrade. You could be forgiven for thinking that would get you a car that can drive itself…fully. Autonomous navigation is traditionally rated on a scale of 0 to 5, and Tesla’s famed auto pilot ranks as a 2.

When it comes to claims about the impending self-driving car revolution, Elon simply cannot help himself. He has no idea how close or far away Tesla is from achieving fully autonomous navigation, but he is always certain it’s next year and has been since 2016.

“I really do not see any obstacles here,” Elon said of the path to full autonomy a couple weeks ago. Interpreting that statement in the best faith possible would require assuming that he meant to finish the sentence, “…because I’m not paying any attention at all.”

And yet…he’s a phenomenal recruiter. Tesla’s “Senior Director of AI” is a former Stanford researcher named Andrej Karpathy. His Software 2.0 talk is a glorious manifesto on the way to build organizations around modern best practices for machine learning. If it’s possible to achieve fully self-driving cars with the state of technology today, I bet Tesla will figure it out. But I’m really, really unconvinced it’s possible.

One of my favorite researchers on the topic, Dr. Missy Cummings of Duke University, had this to say about the idea circulating last year that Tesla “robo-taxis” would be pulling out of driveways everywhere by 2021:

Frankly, I don’t think we’re very close to self-driving cars based on my experiences trying to get deep learning algorithms to trace buildings and roads in satellite imagery. It’s not that the algorithms are weak; on the contrary, they’re exceptionally impressive (and tireless, to boot).

But! They are fragile. Specifically, they are vulnerable to “edge cases”-decisions that are rare, often unintuitive, and inherently unpredictable. Think about all the time you’ve spent riding around cars-there were probably one or two times when you really weren’t prepared for what happened and it scared the living daylights out of you. A family friend once had a deer jump off a bridge and kill itself on his windshield as he passed underneath; show me that in your training data.

What I mean is this: even if we can get 98% of the way to full autonomy this decade, the remaining 2% is so high-stakes that it doesn’t really matter. It’s an existential problem, in my opinion, for people looking to dedicate their career to *this* particular application of machine learning when there are so, so, so, so many other interesting applications for the technology that are better suited to occasionally fallible automation.

Assuming I’m Wrong (and I Often Am)…

Ok, so I just said we aren’t close to self-driving cars. I hope I’m wrong. And assuming I am, let’s think through how self-driving cars will function from first principles:

  • I don’t need a perfectly detailed 1:1 map of the world in order to drive. In fact, I don’t need a map at all. Just based on road signs and a basic understanding of geography, I can navigate pretty well even over long distances.
  • Autonomous vehicles, if they reach parity with humans, will at least have the same level of agency and competence when navigating. They may augment that ability with better, faster, further-reaching data to achieve better-than-human results. But if the 5G connection gets interrupted and they’re on their own…they should be able to keep functioning at a reasonable level.
  • Mapping the world in low definition is a hard enough task that no one has done it exhaustively yet. And for the groups like Google and the National Geospatial Intelligence Agency (NGA) that have tried, their annual budget is in the billions of dollars. It’s just not logistically feasible to get high definition data of everywhere on earth all the time updated regularly.
  • Therefore, as the technology behind autonomous navigation improves, I would expect car makers to focus their differentiation on directly-observed, in-car navigation systems supplemented by outsourced map data of varying levels of fidelity.
  • By Tesla’s very own admission, they’re focusing on a “vision-based” approach to navigation and counter-positioning themselves against Waymo, who they say is taking a “map-based” approach. Robert Scoble-if you’re reading this-watch the video linked in the prior sentence.

That’s a longwinded, terrible-use-of-bullet-points way of saying: Tesla isn’t going to “crush Google Maps,” and they aren’t even trying to. Even if they were trying to, there’s another big problem: most of the raw data directly observed by sensors on their cars is probably protected by personal privacy laws (even if all Tesla plans to do is use it for R&D purposes).

The California Consumer Privacy Act (CCPA) went into effect last year and is the American analog to GDPR for car manufacturers. Like it or not, California is a large enough “sub-market” within the United States that whatever regulations are enforced there tend to affect the way cars are manufactured everywhere else across the country. That law is not messing around: according to CCPA, you even have a right to contact Tesla and make them delete data they’ve collected about you.

I think the upshot of CCPA, and others like it to come, is that car makers are facing an enormous, uncertain liability if they collect troves of personal data (and I would imagine that video of your driving habits is very personal and sensitive). Therefore, they’re greatly incentivized to work with 3rd-party vendors who either simulate the data (“synthetic training data” as its sometimes called) or go to great lengths to secure direct ownership of the data and provide it to car makers cleanly.

The same thing has happened in every other facet of the mapping industry: eventually, the collection of basic, foundational data is competed down to a low-margin services business and the Big Money™ is made in applications that license that data.

The Finish Line

I believe the autonomous vehicle investment bubble is popping. There are already some signs…in December of last year, Uber divested of its self-driving unit for ~$4B (including forking over $400M in cash to help the new owner), which was down from its prior valuation of $7.5B just a year prior. If anyone has a strong incentive to figure out self-driving robo-taxi fleets, it’s probably Uber.

Autonomous vehicle folks are looking ahead to a long, slow slog of incremental progress under the weight of Elon’s hollow promises. But it’s not all bad-the insane investment in the space over the past five years is still likely to drive acceleration in technologies like edge computing, 3D reconstruction and depth inference from video, and LiDAR sensor manufacturing.

What’s not going to change is the fundamental way we map the world or the players who dominate that pursuit. Autonomous vehicle manufacturers may provide an economic engine that incentivizes broader and more frequent collection of foundational map data, but they don’t want to weigh down their absurdly optimistic valuations with scores of software engineers and field techs rebuilding mapping infrastructure that is already on offer for fairly affordable prices.

Tesla won’t crush Google Maps. Apple on the other hand

Originally published in the February edition of



Joe Morrison

Comedic relief at Umbra. Writing about maps and the people that make them. For inquiries: jrmorrison.jrm [at] gmail [dot] com