I grew up in a world of stick maps — street centerlines and roughly digitized curves. All topologically correct, but crude. This was the absolute minimum required to power the pioneering Etak Navigator back in 1985. 

At Etak we took extremely simplistic digital map data from the US Census Bureau, called GBF/DIME files¹, and using the information from the paper maps published by the US Geological Survey added shape, topology and any other missing data we could find. This was an extremely labor intensive process. At its peak we had about 36 workstations and ran 24×7 shifts. It took us years to get there, but eventually we digitized the whole of the US and much of Europe. 

Fast forward to today’s world and organizations that want to make a map have it much easier, but it’s still really hard. If you want to own the map you can’t just copy Open Street Map. Just like Etak did in 1985 you have to start from scratch. But thanks to Gordon Moore and his law you now have a night-day-day technology advantage. And in the US at least you can start with the US Census Bureau’s TIGER files² which are a tad more shapely than their GBF/DIME file predecessors. 

You can lease a large fleet of vehicles, equip them with expensive cameras and LiDARS, drive all the roads and vacuum up all the data. But while this will get you a lot, it won’t get you everything. You’ll get lanes, street signs, traffic lights, speed limits and maybe if you’re lucky some addresses or businesses. But you won’t get post codes or administrative areas. Or rivers. Or golf courses. Or indoor maps. Or 3D building models. 

And of course all this won’t come cheap. Plan on a budget that starts with a number greater than one and ends with a ‘B’. 

At the end of it all you’ll have a beautiful map. But that construction your vehicle passed when it was collecting data? Well that was changing the intersection from a four way stop sign to one controlled by traffic lights. Your beautiful map is now out-of-date. Sucker!

Maintaining a general purpose map that is used for finding locations and turn-by-turn navigation is hard. Really hard. Believe me — I lived through it at Etak, at MapQuest and at Apple Maps. Even supposedly simple things like keeping speed limits up-to-date is horribly hard. There are four million miles of drivable road in the US alone. Your private fleet is not going to be able to drive the entire network every day, no matter what the market cap of your company is. It’s just not tenable. 

So now let’s switch gears. Let’s up the ante. Let’s talk about creating a map not just for finding stuff and getting there. Instead let’s talk about a map to support autonomous vehicles. Now you really have to be on top of your game. Just about every company in the autonomous vehicle business will tell you that you need something called an ‘HD Map’ or high-definition map. It’s like the general purpose map from Google Maps or Apple Maps but with excruciatingly more detail and centimeter perfect accuracy. 

There’s a ton of money pouring into the HD Map business. According to a report from MarketsAndMarkets, it’s projected to reach US$16.9B by 2030. The theory is that you absolutely need an HD Map to support Level 3+ autonomous driving system. The difficulty of producing an HD Map is illustrated by the fact that autonomous vehicles are commonly limited to specific geographic areas. This is partly due to climate — operators want to reduce risks of failure due to sensor obfuscation from road grime, but it is also due to the fact that the vehicles need an HD Map and those HD Maps are expensive to produce and so have very limited coverage. 

I’m not an expert in autonomous systems, but I suspect many of them rely on a method of differentiation to operate. By that I mean the autonomous systems take what the vehicle sees and dynamically compares that to the HD Map as a reference. They use this comparison to deduce (1) where the vehicle is, (2) where it can go and (3) what’s around the vehicle that is not part of the map, for example other vehicles and objects. 

Back in 1985 when I was at Etak we used to joke that the Etak Navigator would be like the introduction of the electronic calculator. Just like calculators eliminated humanity’s ability to perform arithmetic in their head, the Etak Navigator would eliminate humanity’s ability to remember how to get from A-to-B. Sure enough this prophesy has turned out to be completely true. My wife reminds me of it constantly — “Why do you need directions home? Don’t you know where to go?” 

Etak used a system of cassette tapes to store the map data. We imagined cars roaming around aimlessly at the edge of our map coverage — their owners completely lost due to having no EtakMap. The brains inside many autonomous vehicle systems are like these poor owners of the Etak Navigator — they’d be completely lost without an HD Map. 

So the big question is this: if it takes billions of dollars to maintain a plain Jane general purpose map, how can organizations possibly build and maintain an HD Map? 

The theory is that eventually your personal vehicle will collect data as well as navigate. So if you’ve bought that snazzy new Waymo Bubble Car it will vacuum up data while it’s driving you around — and that data will help keep the HD Map current.³

Clearly the issue is scale. Today no member of the public owns a Waymo. Waymo operates 25,000 vehicles, but how often do you see one drive down your cul-de-sac? Not as often as an Amazon van I suspect. There are simply not enough vehicles in these dedicated fleets to collect the necessary data for an HD Map and, more importantly, keep it current. 

But there is one way out of this conundrum. 

What if your system doesn’t rely on an HD Map? 

As I said in the title of this post — this is very much a religious question — and it’s the folks at Tesla who have a completely different religion. They don’t use an HD Map. This means that, in theory at least, their vehicles aren’t limited to drive in a particular area.

If you haven’t already watched the recent video from Tesla’s AI Day 2022, I strongly recommend you do so. It’s quite technical, but it will give you a very clear idea of how Tesla thinks — or to use Steve Jobs’ phrase: how Tesla “thinks different”. 

If you distill everything Tesla talked about in their presentation — humanoid robots, full self driving methodologies and home-grown supercomputers — I think you will come away with the takeaway that Tesla is fundamentally about two things: 

• Tesla is about scale 
• Tesla is about efficiency

This is demonstrated by their effort to produce a humanoid robot called Optimus, which I predict is destined to become the Model ’T’ of robots. Yes, as predicted, the technical media immediately scoffed at Tesla’s efforts, saying it was nothing like what you see from Boston Dynamics. But Boston Dynamics started their dream 30+ years ago! ⁴ Tesla has been working on Optimus for just 13 months. 

So the reaction is a little like the initial reaction to iPhone. I would be somewhat cautious about immediately dismissing their efforts. 

Tesla is focused on building Optimus out of cheap, readily available materials — no carbon fibre for example — and they plan to manufacture it using techniques they’ve learned from making Tesla vehicles. Elon Musk predicts they could ultimately produce millions of units with each one costing less than a car. This is an example of how Tesla focuses on scale. 

Tesla is also leveraging everything they’ve learned from their other work to speed development. This will help them leapfrog everyone else in the industry. For example, they’re using their full self driving software to give Optimus the brains it needs to navigate indoor spaces. This is an example of how Tesla focuses on efficiency. 

If you look at Boston Dynamics by comparison, they’ve done some very impressive work, but now they have significant challenge ahead of them. They don’t have Tesla’s high volume manufacturing prowess, nor do they have Tesla’s autonomous navigation expertise, nor do they have a high volume factory floor that they can use to test and refine their robots at scale.

But let’s get back to the question of HD Maps:

For full self driving, Tesla uses a map, but to use their words, it’s a “coarse road-level map … this map is not an HD Map”. This ‘coarse’ map is used in combination with vision components built from vehicle camera data to dynamically derive lane connectivity in real time.

By choosing not to rely on HD Maps, Tesla has undoubtably chosen a much harder problem to solve as their vehicles have no theoretical ‘ground truth’ to compare to.  But assuming their approach is successful it should result in a much more capable, intelligent and independent system that doesn’t have the extreme cost burden of building and, more importantly, maintaining an HD Map. 

Tesla still spends an enormous amount of time, energy and money to process information used to train their neural networks, particularly the neural networks used for what they call “auto labeling” ⁵. Where does all this training data come from? Well of course a lot of it comes from all those Teslas driving around — there are now about 2 million Teslas on the road today. Given Tesla’s fleet is orders of magnitude greater than anybody else’s autonomous fleet they can further accelerate away from their competition. (And now they’re going to do the same in robotics.)

The cost of Tesla’s differing approach is still significant. For example, it includes the development of Tesla’s Dojo supercomputer to solve the massive parallel computing problems of processing petabytes of data. Their supercomputer architecture is significantly different from conventional approaches that traditionally amass banks of GPUs. As a result it provides significant efficiency gains. I don’t see GM, Toyota or VW developing a brand new supercomputer architecture like this anytime soon. Perhaps NVIDIA? We shall see. 

I suspect there will be other benefits to Tesla’s approach, some of which even Tesla has yet to anticipate or realize. 

It’s ironic, but if anyone could build and maintain an HD Map that covers a large geographic area, like the USA or Europe — and keep it super current — it’s Tesla. They’re the only ones that have a large enough fleet to do it.

I pity other manufacturers. It’s going to be insanely hard to keep up. 

In the meantime we’ll see who ultimately wins this religious battle.

Stay tuned.

---

Footnotes:

¹ GBF = Geographic Base File; DIME = Devilishly Insidious Map Encoding. Credit: Marv White

² TIGER = Topological Illusion Generating Extensive Rework. Credit: Marv White

³ In the interim traditional OEMs hope that third parties like Intel’s Mobileye, Toyota’s Camera and Nvidia’s DeepMap will help them collect data from the cars they already sell. However, the issue is most of their vehicles just don’t have the cameras or sensors to collect the required data at the quality that is needed. 

⁴ See https://www.bostondynamics.com/about: “We began the pursuit of this dream over 30 years ago, first in academia and then as part of Boston Dynamics”

⁵ This is the process of automatically identifying and categorizing features and objects that the cameras in Tesla vehicles see as they drive.  

Acknowledgments:

• Marv White, Chief Technology Officer at Sportvision. Marv is an amazing mathematician with a wicked sense of humor and was my mentor at Etak
• Tesla and everything they presented at AI Day 2022.
• MarketsandMarkets for their report on the HD Map market