Auto Pilot V8.x - When is database created?

[Three60guy]

I have a question.

Electrek provided a transcript of the press conference regarding Auto Pilot 8.X. In part Elon Musk stated:

"It’s a lot of software and quite a complicated job to fit that software on to the available computer hardware in the vehicle. It was a challenging software problem, but one that we were confident we could solve. And something that quite uniquely Tesla is able to solve because of the fact that we can use fleet learning to have all the Tesla cars out there effectively give us the geolocations of where all the false alarm occurs and what the shape of that object that causes the false alarm. That we know that at a particular position at a particular street or highway, that if you see a radar object of a following shape – don’t worry it’s just a road sign or a bridge or it could be a Christmas decoration that somebody put across the street.

As soon as a few cars pass that point, it will geo-coded into our system via fleet learning system that it is now being generalized to create a list of exceptions."

My question relates to when is this fleet learning taking place?

1. It has already learned this.
2. It will learn it in the first few weeks of use.
3. It will take months to learn.

I totally understand the advantages of using radar. But are we at higher risk during the phasing in of this technology or has this database of geo-coded info been created already.

If someone goes on a trip of a thousand miles right after the release of V8.X will the driver need to be extra careful because the database is still being created or will the driver be able to utilize the new technology right away as intended?

I can not find the answer to this in any of the transcripts so far.

What do you think?

 

[andrewket]

My guess is the database already exists and has for some time. The firmware updates we've been receiving for the past 9 months...

 

[Phoenixhawk101]

My guess is that the geo coded locations already exist to some degree and that they have been building up confirmation % for some time. It doesn't take an update to the car to feed this data to the mothership and then the mothership to run the data to see likelihood of working. The update would simply allow this data to be pushed back to the car.

Every time you run under a sign your car has been sending this gps tagged data to the central server. We now have locations where the "all clear" tag has been dropped 100,000 times. Now your car will be able to pull that data to say "I'm seeing something weird here, is this okay?" To which the server now can reply "Yep, we see that all the time. No different than every other car that went through there." That's really all the update will do.

 

[Saghost]

Elon mentioned that they needed new firmware for the radar to expose the raw data to see the objects that it (and all other car radars) normally discard.

So unless they've been pushing that updated firmware into the later 7.1 releases, they don't have the object list yet, except for areas the early release test cars have been driving. My expectation is that they will be building the list with 8.0 cars after the firmware release.

The heavily traveled routes will probably build reliable lists very quickly, but it will probably take months to get near 100% coverage of the road network.

I expect that Tesla won't wait for a specific time interval to start using the whitelist to do emergency braking - instead, I think it'll be based on the number of passes down that route they gotten - so it'll start using the data almost immediately on major freeways, but it may take a while for data from smaller roads to get integrated and developed to adequate reliability.

 

[DrGuest]

I think it'll be based on the number of passes down that route they gotten - so it'll start using the data almost immediately on major freeways, but it may take a while for data from smaller roads to get integrated and developed to adequate reliability.

So take care here in Montana, It will take a while for our roads to get covered and so keep both hands on the wheel and enjoy the views.

 

[Saghost]

So take care here in Montana, It will take a while for our roads to get covered and so keep both hands on the wheel and enjoy the views.

Assuming your lanes are well marked, AP should do just fine keeping us in them anyway. You just won't have confidence about stopped things on the road for a while - as long as the driver looks where the car is going, you'll still be fine.

I would also assume that a number of passes from a single car would be enough data, so anywhere you do as a daily or weekly commute should be covered fairly quickly, even if you're the only Tesla to use that road.

 

[Boatguy]

We can only hope that when 8.0 is actually released they will provide more complete information. As I read it, @Saghost is correct. The new database can't be built until the new firmware is released. The new database is to eliminate the false positives (i.e., firmware thinks collision is imminent when in reality it is not) that will come from the radar under various circumstances.

And as @DrGuest points out, more heavily traveled routes will populate the database first.

The confusing part is that as I interpret Elon the effect of the lack of a database entry will be acting on the false positive, e.g., slowing down or stopping when there is no need to do so. This could get kind of weird for awhile, but as @Saghost points out, this is going to be worse if you are touring as opposed to commuting.

 

[Saghost]

The confusing part is that as I interpret Elon the effect of the lack of a database entry will be acting on the false positive, e.g., slowing down or stopping when there is no need to do so. This could get kind of weird for awhile, but as @Saghost points out, this is going to be worse if you are touring as opposed to commuting.

I could be mistaken, but my expectation is that the initial data reading rollout period won't be a matter of time, but a matter of passes along a given road - so driving a new route it wouldn't react to radar any differently than it does today, it would just take notes.

I think the only place you'll get the false positive response is when a known/trusted route gets something new - a new road sign or the like. There's no way to weed that out unless you can use the camera to identify the source of the radar return as non-threatening...

 

[Boatguy]

I could be mistaken, but my expectation is that the initial data reading rollout period won't be a matter of time, but a matter of passes along a given road - so driving a new route it wouldn't react to radar any differently than it does today, it would just take notes.

If it's just "taking notes" when there is nothing in the database, then it will never respond to the semi crossing in front of you.

I think it has to operate in the opposite mode, it has to respond to the radar target (i.e., brake) because the target may be a semi, until enough cars pass that point and it learns that the "semi" is actually an overpass where the road dips below the overpass.

Tesla really needs to release a full description of the behavior that should be expected from 8.0 in various situations.

 

[DrGuest]

Assuming your lanes are well marked, AP should do just fine keeping us in them anyway. You just won't have confidence about stopped things on the road for a while - as long as the driver looks where the car is going, you'll still be fine.

I would also assume that a number of passes from a single car would be enough data, so anywhere you do as a daily or weekly commute should be covered fairly quickly, even if you're the only Tesla to use that road.

I just want to point out that Montana roads were never designed to go over 80 MPH so curves will often deactivate AP, we have guard rails that are very close to the road so I often deactivate AP when I move to where I am comfortable. We have many narrow bridges that often make me take control and we do have lots of construction with no lines or one being obscured. I drive with AP very often but I use it to reduce my fatigue by resting my arms on my thighs but mostly have both hands on the wheel as Tesla advises. I really think anyone multitasking with hands off the wheel in Montana are taking a risk that could reduce safety instead of enhance it. Keeping my hands on the wheel and arms relaxed does reduce my fatigue without increasing my stress if AP suddenly wants me to fully disengage and take over. So for long straight boring sections the AP is quite flawless but I still like staying in control, especially when taking these amazing curvy dream roads and really seeing what this car can do!

 

[Saghost]

If it's just "taking notes" when there is nothing in the database, then it will never respond to the semi crossing in front of you.

I think it has to operate in the opposite mode, it has to respond to the radar target (i.e., brake) because the target may be a semi, until enough cars pass that point and it learns that the "semi" is actually an overpass where the road dips below the overpass.

Tesla really needs to release a full description of the behavior that should be expected from 8.0 in various situations.

If they are doing it the way I'm suggesting, then it won't respond to the semi the first couple times the route is driven.

After that, when it reaches for the database, it'll be told that the route is established and there's nothing to see - after which when it sees the semi, it stops.

The key point being that it isn't just a white object list - it's also a list of routes that Tesla has confidence they've seen everything on, and the car's behaviour changes when on one of them from when not.

 

[Three60guy]

Well it looks like there is a variety of answers concerning this thread. I see no consensus. Maybe I need to call Tesla and see if I can find out some facts from them.

 

[Todd Burch]

8.0 has probably been running on all of our cars in passive only mode for awhile. So yes, I agree the database has already been getting developed for awhile.

Only thought that has me questioning that is that 8.0 seems to require a fundamental change to the radar driver which might restrict the ability to run 7.x and 8.0 simultaneously. (Plus, running 7.0 plus doing all the cloud processing might be too much for the processor).

 

[chillaban]

It's worth mentioning that the geocoding whitelist/blacklist is specific to the collision alert/avoidance mechanism. Hopefully that is not something that you are dying for (no pun intended) on day one of 8.0's release. The point cloud mapping, 2-car following, enhanced cut-in/cut-out response times, moving aside for adjacent cars, and almost every other point in the blog is immediately effective without any fleet-learning effects, and hopefully those are the features you enjoy most.

It's sorta like buying a car and being really excited about the crumple zones. I mean... yeah.... they're great to have and you hope you never need it.... but lamenting that a imminent collision mitigation system won't be 100% effective on day one is kind of morbid

 

[Boatguy]

If they are doing it the way I'm suggesting, then it won't respond to the semi the first couple times the route is driven.

After that, when it reaches for the database, it'll be told that the route is established and there's nothing to see - after which when it sees the semi, it stops.

The key point being that it isn't just a white object list - it's also a list of routes that Tesla has confidence they've seen everything on, and the car's behaviour changes when on one of them from when not.

OK, that makes sense, but it's sort of rolling the dice. You have to ask yourself "how many Tesla's have driven this route?".

Like many Tesla features (e.g., cell phone coverage for navigation maps) it's very biased to higher population density areas which are well traveled with good cell phone coverage.

 

[Saghost]

OK, that makes sense, but it's sort of rolling the dice. You have to ask yourself "how many Tesla's have driven this route?".

Like many Tesla features (e.g., cell phone coverage for navigation maps) it's very biased to higher population density areas which are well traveled with good cell phone coverage.

You may not know, but Tesla does. That could show up in a couple places.

The Navigation system could be given an option to prefer verified routes.

I also wouldn't be surprised to see a difference in the instrument panel display when the car has whitelist data and when it doesn't. In fact, I think Tesla probably needs to do that once they start using whitelist data.

I'm not sure the rolling dice analogy is completely fair - the worst case is that the system works the way it does now, without the extra assurance but certainly safer than most.

 

[daviddavidw]

Just my thought, and it might not be possible if there is too many code changes or driver update compatibility issues, but if I were implementing this, I'd add a fallback that when the point cloud is too sparse for a given route, i'd continue to collect the new radar data and build out the point cloud, but fallback to the current 7.1 algorithm with visual as the primary and radar as a backup until the car returns to a route that has sufficient density to prevent false positive breaking. That would virtually guarantee that the performance was almost alway equal or better than current. I think this is similar to what Saghost is saying.

 

[Saghost]

Just my thought, and it might not be possible if there is too many code changes or driver update compatibility issues, but if I were implementing this, I'd add a fallback that when the point cloud is too sparse for a given route, i'd continue to collect the new radar data and build out the point cloud, but fallback to the current 7.1 algorithm with visual as the primary and radar as a backup until the car returns to a route that has sufficient density to prevent false positive breaking. That would virtually guarantee that the performance was almost alway equal or better than current.

Maybe I'm confused, but I thought the blog post was pretty clear that this was exactly how it would work - the system operates like it does today (while also noting radar returns for the whitelist,) until they assemble the list, then it does all that plus braking/steering to avoid things that aren't on the list and aren't flagged by the camera.

 

[Todd Burch]

OK, I reviewed the transcripts from Elon's press conference.

It seems pretty clear to me that the non-beta-testing cars out there could not possibly be doing any of the point cloud database generation. It requires a change to the radar driver, which could only come as a software update (since it would require a firmware update to the radar). So only Tesla and the beta testers would seem to be building the database at the moment.

Moreover, based on how Elon said this processing pushes the CPU close to the limit, running 8.0 in the background while separate 7.1 processing is occurring seems almost impossible.

So I would expect that when 8.0 comes out, with the database virtually empty for many places in the world, the improvements stay marginal initially, but then improve dramatically over a relatively short period of time.