Skip to main content
No. of Recommendations: 2
This 3 month old video is Mobileye biased.

https://youtu.be/1Ew5OtibrXE

Many EV companies (and some ICE Vs?) are anticipating Autonomous Driving. Some (TSLA) are advertising it as a way to rent your car to clients while you work, sleep, vacation, etc.

The linked video describes the concepts behind the sensor technologies, the limitations and the potentials.
These are my notes from the linked video, and are a bit erratic.

FROM A MOBILEYE POV.

All the sensor technologies utilize AI to interpret the sensor data.

TSLA promotes itself as being the only company that gathers billions of data bits from its cars... FREE. The owner pays TSLA for the car, and real world driving data is reported back to TSLA for AI processing and then pushing to each TSLA car, rinse, repeat. SEE #5, below.

But wait!
This video, Mobileye brags about its millions of km of data, more data better AI, same virtuous cycle, demanded its data suppliers to abandon their own self driving programs. In 2020 Mobileye Technology is in 60M cars, 6M km/day data, expects 14M in 2022. Video says this is more than any other company.

Camera only (Camera Centric)- Mobileye (Intel). Uses cameras to build 3D representation of the world. Users: China NIO, Paris RATP(?). Cameras are much cheaper than radar/lidar. But, how to get a 2D image to 3D?
Decided, in 2004, to build its own chip called EyeQ for 2D to 3D image processing. (AAPL and TSLA followed suit). 2014

Lidar. GOOGL Waymo. Was the putative self driving king in 2014. Cost was $30k-85k, but now just a few thousand, expected to eventually drop to hundreds (Innoviz). Mobileye founder said Lidar would not scale economically "If it doesn't scale, it's going to be a science project." And scorned Lidar as the "lazy approach". Lidar gives a 3D image, without growing the AI ability.
Better than cameras for small objects, fake objects hacked into car controls, low light and poor weather conditions.

Radar has its uses. But also detriments - difficulty identifying stationary objects. Strength is to detect speed of nearby, adjacent objects.

Camera, radar. and ultrasonic TSLA. NPU (Neural Processing Unit) chipset developed in house by TSLA. See #5 below.

High definition GPS maps. Mobileye placed the car location via HD GPS Maps, within centimeters, which gives a redundancy to what the camera is seeing. Confirms accuracy of cameras. While Musk initially scorned HD MAPS, TSLA has recently adopted the same philosophy** and now incorporates MAPS.

NVDA (at 21 minute mark) AI processing (recent ARM Holding acquisition?) may become Mobileye biggest competitor by mid 2020s. But NVDA is late to the party.

Waymo. Relies on expensive Lidar. Not much real world data, so it seems to be stalling?

AI, Deep Learning, Machine Learning. TSLA relies on this. Mobileye, not so much.

TSLA today, is copying concepts that Mobileye had in 2016, 3D bounding boxes, tri-focal focus fields (at 25:00 minute), HD MAPS, Mobileye presentations, and more.


Random info:
Mobileeye launched 1999.
https://www.mobileye.com/
https://www.mobileye.com/about/


RATP Group - government entity responsible for Paris public transportation.
https://en.m.wikipedia.org/wiki/RATP_Group
While RATP's Paris-related activities are still a major part of its business, its operations have extended since 2002 to include business around the globe in various modes of urban and regional transportation. RATP Dev, the Group's international operations and maintenance subsidiairy, is present in 13 countries in Africa, Asia, Europe and North America.

LIDAR LIght Detection And Ranging.
Radar RAdio wave Detection And Ranging.

** Musk sure seems to have a lotta petty Urination contests with others.

5. Benzinga, ARK invest, says TSLA AI NPU chip, data collection, etc is 4 years ahead of all competition (Mobileye and NVDA).
https://m.benzinga.com/article/17872810

😷
ralph
Print the post Back To Top
No. of Recommendations: 4
Very interesting!

Two comments:

The market, not technology, determines the winner. Tesla's vertical integration gives it more control but Mobileye might wind up in more cars. Time will tell.

Sailors don't use maps, we use charts which are the same thing. ;)

Over reliance on maps is problematic because the terrain changes all the time. Mariners are requested to report any changes they observe. In the old days this took a lot of time which the Internet solves but not entirely because it takes time to retrain AI. On board AI has to have the last word just like the captain of a ship.

Since Google had maps they decided to leverage them but this might turn out to be a costly mistake for them.

Thanks!

Denny Schlesinger
Print the post Back To Top
No. of Recommendations: 0
There seems to be recognition that a universal system will be challenging for a while. But one that uses the same terrain again and again will likely be first. Delivery of pizzas on a college campus. Unmanned trucks on a stretch of interstate highway.

With experience confidence rises and coverage area expands.

And of course several accidents in the news raise concerns. Not ready for prime time??
Print the post Back To Top
No. of Recommendations: 0
Over reliance on maps is problematic because the terrain changes all the time. Mariners are requested to report any changes they observe. In the old days this took a lot of time which the Internet solves but not entirely because it takes time to retrain AI.

Why should it take a long time to retrain AI? If a route suddenly changes - say, a detour is put up around a construction area - an AI car should notice that, after say 10 out of 10 sister cars in its AI network have changed their route, it should do the same. That ought to take maybe a whole half hour (if ten sisters passed by in that time frame).

Simple: share information.
Print the post Back To Top
No. of Recommendations: 3
But wait!
This video, Mobileye brags about its millions of km of data, more data better AI, same virtuous cycle,


Not disagreeing with the technical details about Mobileye.

But...one thing to consider is this. Intel bought Mobileye a few years ago.
Intel management has been pre-occupied lately. Assaulted by AMD on the CPU side and Nvidia and AMD on GPUs where they are trying to be relevant. Apple is dropping them as a supplier for Mac computers as they are ramping using their own CPU. At the same time they are floundering at what they dominated the industry in for the last 30 years...they "were" always the best in FAB process technology, now TSMC is the undisputed leader as Intel has had repeated delays at 10nm and smaller processes.

Bottom line...I don't know. Either Mobileye is flying under the radar within Intel or they might be needing to justify every dollar they spend. Mobileye is probably the tenth most important thing at Intel right now. Compare to AI within Tesla.

Mike
Print the post Back To Top
No. of Recommendations: 0
Why should it take a long time to retrain AI? If a route suddenly changes - say, a detour is put up around a construction area - an AI car should notice that, after say 10 out of 10 sister cars in its AI network have changed their route, it should do the same. That ought to take maybe a whole half hour (if ten sisters passed by in that time frame).


Unfortunately this is not at all how AI works.
But before I explain, consider this. Has there EVER been a documented case where something like this has been demonstrated or even mentioned that it does work like this?
Second, in my Tesla I get a software update about once a month, maybe twice. It takes ~25 minutes. I don't get an update 30 minutes after someone puts up a construction sign.

Sure, if the cloud map data has has a road closure noted due to construction, then if I request a route to someplace it should warn me and/or route me on a different road. But this has nothing to do with "training the AI" in the car. Waze (Google) maps has been doing this for more than a decade and it really works better being not specific to a car vendor, IMO.

Convolutional neural nets work by training a model with a large dataset so that it can "generalize" how it operates -- i.e. not having to have exactly seen the same situation before but still be able to make the correct choices. Training the AI model controlling an individual car is a different thing than some cloud code or possibly some cloud AI looking at what current traffic is doing and recommending different routes.

Mike
Print the post Back To Top
No. of Recommendations: 0
if I request a route to someplace it should warn me and/or route me on a different road... Waze (Google) maps has been doing this for more than a decade

But this has nothing to do with "training the AI" in the car... in my Tesla I get a software update about once a month, maybe twice. It takes ~25 minutes. I don't get an update 30 minutes after someone puts up a construction sign.


Then it sounds like Waze is a fundamentally better-designed product, at least as far as map updates are concerned.
Print the post Back To Top
No. of Recommendations: 1
Why should it take a long time to retrain AI? If a route suddenly changes - say, a detour is put up around a construction area - an AI car should notice that, after say 10 out of 10 sister cars in its AI network have changed their route, it should do the same. That ought to take maybe a whole half hour (if ten sisters passed by in that time frame).

You just proved my point, during that whole half hour self driving cannot rely on maps, it must rely on local input.

Denny Schlesinger
Print the post Back To Top
No. of Recommendations: 1
Then it sounds like Waze is a fundamentally better-designed product, at least as far as map updates are concerned.

Not better or worse, Waze is map centric, Tesla is camera centric.

Denny Schlesinger
Print the post Back To Top
No. of Recommendations: 1
Not better or worse, Waze is map centric, Tesla is camera centric.

All I want is to get to my destination as safely and time-efficiently as possible. Sounds like Waze does the time-efficient part waaaze-better.
Print the post Back To Top
No. of Recommendations: 2
Updating a map has nothing to do with training the AI. It is just changing the data on which the AI operates.
Print the post Back To Top
No. of Recommendations: 0
Updating a map has nothing to do with training the AI. It is just changing the data on which the AI operates.

And you know what they say about data in computing: garbage in, garbage out.

Up to date data whips AI done on garbage data.
Print the post Back To Top
No. of Recommendations: 0
Sounds like Waze does the time-efficient part waaaze-better.

Aside from any self-driving stuff, what Waze is used for today is much like what any GPS & maps guide program does, but with better current local information. Navigating with Google Maps shows me the extent of the slowdown on the highway. Waze tells me there is a accident on the right shoulder, because Waze users have reported that.

Any self-driving software needs the information in the mapping system to be able to navigate. It also must handle what is actually there, regardless of how different it is from what the map tells it. If it can take advantage of up-to-the-minute information from Waze users, or other vehicles, or traffic helicopters for that matter, so much the better. But that is mostly just doing navigation better.
Print the post Back To Top
No. of Recommendations: 0
If it can take advantage of up-to-the-minute information from Waze users, or other vehicles, or traffic helicopters for that matter, so much the better. But that is mostly just doing navigation better.

"Just"?

Let’s go back to the example of a detour around a construction area thrown up this morning. How is an AI car that lacks timely data input going to know this? Only way I can think of, is it reads the "Detour" sign and follows it, and hope each turn on the detour is clearly signed (they often aren’t). But after four years of intensive AI development, Tesla cars still can’t read signs. Good luck "just" getting to work and back in that situation.

I like Tesla, I own an M3. But it still doesn’t know the speed limit in my neighborhood, which has been clearly signed and unchanged for at least five years that I can recall. Count me as not impressed with the rate of AI (non-)learning.
Print the post Back To Top
No. of Recommendations: 3
Then it sounds like Waze is a fundamentally better-designed product, at least as far as map updates are concerned.

I think your idea of "better designed" doesn't match with what is possible or needed.

An individual car needs to be able to operate "autonomously"
That means by itself with no external input from anything, since at any time that data connection could be limited or not available at all. If (a big if) it does get external input, such as from the cloud, from another vehicle or from the local infrastructure (such as a smart city traffic light), then great.
A system like Waze is different. It is not driving the car or responsible for its safety. It doesn't know about the light that just turned red. It doesn't know about the car that just caught fire right before your eyes (and camera). It is a crowd sourced database that allows generation of GPS navigation routing that is probably better than the year old or month old maps and could be based on just a few minutes delayed data.
There is a proper place for both types of systems and products.

Mike
Print the post Back To Top
No. of Recommendations: 0
Waze is different... It is a crowd sourced database that allows generation of GPS navigation routing... based on just a few minutes delayed data... There is a proper place for both types of systems and products.

Tesla navigation I think uses Google Maps for its current data. Google Maps is pretty good on spotting traffic jams. I don’t know how they do it, perhaps satellite images?

What puzzles me then is how Tesla fails to have an accurate speed limit database. Unless Google Maps doesn’t do speed limits and only reports traffic jams. Still a big mystery to me. Do you know?
Print the post Back To Top
No. of Recommendations: 2
Google Maps is pretty good on spotting traffic jams. I don’t know how they do it, perhaps satellite images?

Real-time satellite images of such detail are limited to the military, despite what really stupid movies show.

No, I think there is a simpler answer for how Google Maps deals with traffic jams. Google owns Waze!
Print the post Back To Top
No. of Recommendations: 0
Real-time satellite images of such detail are limited to the military, despite what really stupid movies show.


Sort of laughable that there would be real time images of all freeways and streets, yet when you look at your house on Google maps the picture is 2 or 3 years old

Mike
Print the post Back To Top
No. of Recommendations: 0
Sort of laughable that there would be real time images of all freeways and streets, yet when you look at your house on Google maps the picture is 2 or 3 years old

Not laughable in the camera and drone era. What is laughable is that Tesla's vaunted AI can’t even figure out speed limits or read traffic signs (though recent beta is finally said to read stop signs and traffic lights, reliability unknown).

Around here we have giant banner-message boards spanning across the autobahn that change the speed limit frequently depending on traffic and weather. I don’t know how the data is sourced and analyzed that is instructing these LED overlords, but they are highly accurate. Cameras measuring traffic congestion? Cops reporting?
Print the post Back To Top
No. of Recommendations: 0
Cameras installed by traffic engineers in critical places makes a lot more sense than satellite images.

Mike
Print the post Back To Top
No. of Recommendations: 1
(Hit send to quickly)

But I think the primary way that Waze and other mapping/traffic software works is much simpler.
They get the GPS data from your phone when you have the app running and notice how long you spend traveling short segments of roads. If it is a relatively long time, then that becomes traffic congestion in that location. If your position is moving along nicely then it is not congestion. If you are ~stopped then it severe congestion.
Of course, one car alone doesn't decide this, but rather multiple cars all confirming the same thing.
This doesn't mean that they don't also get data from highway patrol and other official sources, such as CalTrans in CA.

Mike
Print the post Back To Top
No. of Recommendations: 0
Waze gets the GPS data from your phone when you have the app running and notices how long you spend traveling short segments of roads. [thus infers your speed]. Of course, one car alone doesn't decide this, but rather multiple cars all confirming the same thing.

Sounds perfectly logical.

So why doesn’t Tesla do the same, to figure out correct speed limits (since its speed limit database is riddled with errors)?
Print the post Back To Top
No. of Recommendations: 1
So why doesn’t Tesla do the same, to figure out correct speed limits (since its speed limit database is riddled with errors)?

Probably because the number of Tesla's on the road is no where near large enough to be statistically good enough compared to Google Waze/Maps. It also makes much more sense to get data from all drivers not just Tesla drivers. There is nothing unique about a Tesla stuck in traffic compared to anybody else stuck in traffic. Tesla should spend their resources where it makes the most sense.

Mike
Print the post Back To Top
No. of Recommendations: 1
So why doesn’t Tesla do the same, to figure out correct speed limits (since its speed limit database is riddled with errors)?

My Model 3 detects speed limit signs and updates the maximum speed for autopilot accordingly and detects traffic lights usually before I can. It doesn’t always recognize directional traffic lights (left turn green arrow). But this is on regular autopilot, not the full self driving beta, which is apparently much better.

This is in the US, so maybe it’s different in other countries. I’ve also never noticed that the speed limits are “riddled with errors” though.

Tesla’s mapping appears to be as good as Waze in figuring out the fastest route based on current traffic conditions. Maybe they use the same data. I’m sure if they wanted, they could leverage cameras to more accurately measure traffic, but that might consume far more cellular data and computing capital to make it worth the effort.
Print the post Back To Top
No. of Recommendations: 0
My RAV4 detects speed limit signs and displays it* (POORLY!) for my use. Here is Toyota's ad:
https://www.toyotavacaville.com/blog/how-the-toyota-road-sig...

Mobileye Technology also detects speed limits.
https://www.mobileye.com/us/fleets/technology/speed-limit-in...
Mobileye’s speed limit indicator scans for speed limit signs along the road. When a driver exceeds the speed limit, a visual indicator alerts them - a helpful reminder for the driver to maintain a safe speed..

Mobileye licenses its technology to many companies.

😷
ralph

*Toyota's display is immature at best. Truly terrible. I cannot understand why it is so bad. It's almost as if they INTEND to obfuscate the info.
Print the post Back To Top
No. of Recommendations: 0
*Toyota's display is immature at best. Truly terrible. I cannot understand why it is so bad. It's almost as if they INTEND to obfuscate the info.


The display on the Tesla certainly makes its autopilot features much more useful. When the car shows you what it is seeing and thinking, it's much easier to trust that it's going to do the right thing.

Our Mercedes has adaptive cruise control and lane keeping assist, but it doesn't show you where it thinks the other cars are or where the lanes are. So you have to trust that it's not going to crash into the median or merge into another car. There's a little icon that suggests lane keeping is on, but I've had it seemingly ignore lanes. Apart from trusting it to speed up and slow down with traffic and blind spot alert, I can't really trust it to do much.
Print the post Back To Top
No. of Recommendations: 0
why doesn’t Tesla do the same, to figure out correct speed limits (since its speed limit database is riddled with errors)?

Probably because the number of Tesla's on the road is no where near large enough to be statistically good enough


No statistics needed. Speed limit signs rarely change. They just sit there.

Tesla drivers are variable, yes but just ten of them driving by the same speed limit sign should correct for that.

Tesla should spend their resources where it makes the most sense.

It would make quite a bit of sense to know what the speed limit is. As they say, Speed Kills.
Print the post Back To Top
No. of Recommendations: 1
My Model 3 detects speed limit signs

Apparently Tesla did finally catch up to 2010 (a decade later) by adding the capability to read speed limit signs in their recent September update in the US (only):
https://www.carscoops.com/2020/09/tesla-software-update-lets...

That it took so long to achieve what competitors have long had, does not inspire confidence in its AI approach.
Print the post Back To Top
No. of Recommendations: 1
Apparently Tesla did finally catch up to 2010 (a decade later) by adding the capability to read speed limit signs in their recent September update in the US (only):
https://www.carscoops.com/2020/09/tesla-software-update-lets......

That it took so long to achieve what competitors have long had, does not inspire confidence in its AI approach.


I am not sure where the issue was. Their autopilot has been available only on freeways until recently, and I've never seen any issues with wrong speed limits. I understand there may be certain areas that have issues but it's not widespread. Again, that might be because it's in the US where speed limits do not change. There's no reason to know speed limits on surface streets unless autopilot is enabled.

I don't believe the issue has anything to do with their ability to achieve self-driving. Perseverating on this detail reminds me of the guy who wrote over and over on Seeking Alpha to short TSLA because of pictures of misaligned Model X door handles.
Print the post Back To Top
No. of Recommendations: 0
Probably because the number of Tesla's on the road is no where near large enough to be statistically good enough

No statistics needed. Speed limit signs rarely change. They just sit there.


Sorry. I was still thinking of the speed of traffic not the speed limit signs.
I don't know why any car would need crowd sourced speed limit sign reading.
The database should have the speed limit as a first piece of info.
If the car sees a sign or road marking then that number replaces it.
This could be a normal speed limit sign or a temporary one such as during construction.
The car should also take into account other factors to self impose a lower speed based on ability to "see" properly, such as in rain, snow, fog.

None of this involves any "learning" on the fly. Just executing the already learned ability to sense what it sees (i.e. inference).

Mike
Print the post Back To Top
No. of Recommendations: 0
I don't know why any car would need crowd sourced speed limit sign reading.

Not “crowd” - as I said, just ten Teslas passing by should do.

They need it since the database is riddled with errors. Also, public GPS is not precise enough. Speed limits begin where the sign is. GPS is not that precise. In addition, the driver may need to begin slowing down well before the sign point.

The database should have the speed limit as a first piece of info.

But it doesn’t. And even if it did, the second point above (GPS imprecision) would still be a problem.

If the car sees a sign or road marking then that number replaces it.

I agree. Especially with the "if".

The car should also take into account other factors to self impose a lower speed based on ability to "see" properly, such as in rain, snow, fog.

Yes that would be a good idea. If it existed, which it doesn’t.

None of this involves any "learning" on the fly.

? The fleet can’t learn while it’s sitting in the garage. It has to be on the fly.
Print the post Back To Top
No. of Recommendations: 0

Not “crowd” - as I said, just ten Teslas passing by should do.


Crowd sourcing is the term used for a multiple other inputs via the cloud. Doesn't matter if it is thousands or 3. The mechanism is the same.

Also, public GPS is not precise enough.

Good enough for speed limit sign alignment. You wouldn't be off by more than a fraction of a second (60 mph is ~88 ft per sec. GPS is close enough for this)

The fleet can’t learn while it’s sitting in the garage. It has to be on the fly.

What I'm saying is that new features have to be learned by the car maker, tested and certified to OK. Your individual car isn't going to learn something new that hasn't been downloaded from the home base. Complying with the required ISO 26262 spec for safety critical systems pretty much has to be like this, IMO.
Sure, your car can learn where your home is or your favorite destinations are. Those details aren't safety critical. But your car isn't going to modify the neural net training that takes place on the car maker's servers.

Mike
Print the post Back To Top