Tesla is crazy not to use maps


Every time I compare Tesla with other contenders, people will say things to the effect of "they don't count, they use maps so they can only drive in tiny regions" because they think that's a bug rather than a feature.

Maps are super useful, and a car that can try to drive without a map is a car that can make a map, and so maps scale just fine and help you drive more roads, rather than fewer.

So I made a video outlining the virtues of maps, why they are cheap, and why it's OK if construction changes the road after you map it.

This is also accompanied by a text version of the transcript which can be found at Tesla is crazy not to use maps


But Tesla does have maps!

Yes, the video states that Tesla makes use of lane-level maps in many (though not all) places and even uses more detailed maps in some places. They deny they use highly detailed maps anywhere though there is evidence they do use them in some places.

The key is to use maps with good detail, and use them everywhere they are needed. They definitely don't do that yet, even in Silicon Valley, where every road has Teslas going down it constantly and where the developers live.

Humans drive without a map

This is covered in the video. Humans can drive without maps but prefer to have them and do better when they remember the road. They are more likely to make mistakes on roads they have never seen compared to roads they are very familiar with. But most of all, robots are not humans, and there is no sign they will be as smart as humans any time soon, or that the best way to solve the problem of understanding the road is making them think like humans. Any more than the best way to fly is to flap your wings like birds, though that works if you're a bird or fictional plane in Dune.

How doe the maps get to the car?

99.99% of map data does not need to be sent real time, and is updated when connectivity is available (in particular if the car parks somewhere with wifi as they often do.)

The only updates that need be sent in real time are

  1. Data from the car if it's the very first to encounter a change to the road, as soon as it gets a connection
  2. To the car if it has plotted a route that eventually goes along a road that has an update

In the worst case if connectivity is very spotty, you can send the location of a map change to a car that is planning to drive that street so that it knows to avoid it or at least is not surprised. In addition, changes located near the car which it might decide to drive down will come as data is available. Needed data is not very large and low speed (3G) connectivity would be far more than is needed.


deepmap was very expensive several years ago. 5000 dollars per km turn key. easy to search.

you never provide an estimate of the density distribution of vehicles per paved lane mile to support your assistive map long term is never provided. We all do not live in cali.

Here TomTom Mobileye Atlatec Carmera Momenta Ushr DeepMap Zenrin Mapbox DynamicMap NavInfo

way too fragmented

lol - Infosysbpm presentation $32 billion for US

talk about meaningless research with no smidgen of reality.

*Source: (DEEPMAP data) 2020

duplicative costs of every oem building their own maps is insanity.

just for one country, $32 billion to build 3d point maps with lidar for localization? yes it is probably less expensive today, but by how much.

and all that $ for just one country.

even if other lidar 3d point maps are 50% less expensive than DeepMap fees, the 10 largest oems duplication of what the other oems do is beyond a substantial spend if using lidar maps.

and the observation that lidar currently is mostly used for localization and not object detection (james douma) is numbing when realizing most automotive lidar is above $500 per unit.

carmera, toyota, ford, tesla, nissan and mobileye should get together and build a global map for pennies.
put a camera and a chip in the windscreen at $150 per vehicle and build the maps for mere pennies.

source: Mapping an autonomous driving future - Infosys BPM

While that is an overestimate, and the cost keeps dropping as tech improves and gets cheaper, it actually would be a quite affordable cost if it had to be paid. $100 per American? Under $200 per driving American? Who drives 12,000 miles/year on average and spends $8,000/year doing it? Also it's not a cost per year, it's the one time cost of the initial map with a lower cost to update it.

Even if you thought it would cost this much, it's a tiny drop in the cost of car travel if it makes it safer and more reliable. But it doesn't cost this much.

Admin note: I do prefer if people post more than a few times here and want to be Anonymous that they make up a pseudonym and use it to tell one anonymous poster from another. Regular users are also welcome to create an account which allows many useful features, like editing your comments, putting in HTML if needed, and authenticated userid.

please help with EFF's pursuit for anonymity if possible

I don't deny people the ability to be anonymous here (though EFF feels that the owner of an online space is free to deny that, and even can be wise to deny it, but defends the existence of anonymous communication where it is wanted or needed, and does not want governments to forbid it.)

However, I have a user here who has an account but I don't know the name of. I could make effort to track his or her IP addresses but I have not done so, and it might not be fruitful anyway if the person uses TOR or a VPN. Usually most people are not that worried.

I request a pseudonym just because it makes it easier for people to understand and track threads. As the other participant said, there is somebody here posting anonymously who is fairly confusing and mildly chaotic. I welcome almost all views here, as long as they are useful and interesting, but I will shut down flamewars or ad hominem.

You can certainly be anonymous if you'd like. But what Brad is suggesting is that you pick a consistent pseudonym and use it for all of your postings here. That way, Brad, and his readers, can better understand if a single person is making all of these short comments dumping incomplete thoughts without any conversational niceties, or if a whole bunch of different people are just filling up the threads with hard-to-understand snippets.

If you are trying to engage in a conversation, it will be a lot more productive if you explain something about the points you are trying to make, and perhaps even ask questions. Capitalization, punctuation, and complete sentences would help, too.

capable of generating High Definition Maps ?.

So it would start from maps, though it could have a map-making module to make the maps the simulator needs.

I commented, tongue in cheek, on the YouTube video that Tesla was deliberately making their data more interesting by refusing to learn from what they had seen before, therefore increasing the number of difficult scenarios their autopilot has to deal with.
Running red lights and getting stuck in tramways or dead ends gives their learning algorithm so many more "interesting" scenarios to work compare to boring self driving stacks that used high definition maps.

Everybody certainly tries to do this in sim, to find the craziest and most interesting scenarios.

One of the controversies today is "how much testing should you do on public roads." One answer is "no so much that you're finding bugs faster than you can fix them." Tesla clearly is way beyond that. Aurora feels they can find bugs in sim so fast that they are not on the roads much yet.

E.W. Niedermeyer

Aug 22
Developing a "general solution to autonomy" by validating a single intersection so an influencer's videos look better, you literally couldn't make this shit up

Warren Craddock never cites any body of scientific research in the sensor choice and sensor fusion discussion.

Deep Learning For Autonomous Driving Workshops on youtube 3D-DLAD-v4

Dr. Eric Richter from BASELABS as an example.

previous years also available
IEEE sensor fest

"This year, we'll see approximately 1 million V2X cars in Europe. The expected penetration will be in the 10 million-ish range by around 2024–2025"

What does it mean to be a v2x car? Is this a car regularly transmitting BSMs? A car that uses them? Some other type of V2X?

10 million of the 250M cars in Europe means that the chance that any two in an interaction could use V2X is 0.0016 (about 1/6th of one percent.)

Ars Technica yesterday in NTSB report.

"30 MHz is enough for basic safety features," said Balázs Tóth-Pintér, a communications specialist with V2X company Commsignia, who pointed out that the EU has only allocated a total of 40 MHz for V2X. And unlike in the US, deployment is happening in Europe. "This year, we'll see approximately 1 million V2X cars in Europe. The expected penetration will be in the 10 million-ish range by around 2024–2025," he told Ars.

Just what are those million cars actually doing with their V2X?

NXP blog explains much.
Some insight as well at Conti and Siemens.

In 2022, the company plans to map the entirety of Europe, which will require the system to scale up to 200,000 concurrent vCPUs for 20 days—96 million vCPU hours in total.

The REM team updates the map in near real time: accessing, changing, rebuilding, and stitching together more than 2 million kilometers of drivable paths with detail down to the level of a single stop sign.

aws dot amazon dot com/solutions/case-studies/mobileye-video-case-study/

electrek dot com Oct 24 2022

XNGP (navigation-guided pilot) was announced as XPeng’s last step before achieving fully autonomous driving. XNGP combines all scenarios of the automaker’s existing ADAS capabilities (highway, city, parking) into one holistic system that will soon no longer require high-precision maps to function – essentially opening up its availability to any and all areas.

ISA (intelligent speed assist) content for Europe from HERE manages the non-posted speed limit challenge by delivering speed limit information based on a combination of data sources. Reliance on vehicle camera data alone, for example, is prone to quality challenges without a base map and validated mapping system.

So if China pursued ISA, what does XPeng do?

Add new comment