Buyer Beware: AP2 Extremely Dangerous

[Deans]

TACC slams on the brakes on its own just fine. No need for AutoSteer for that. What you get extra with AutoSteer, is the erratic turning left and right within the lane

TACC slams on the brakes without AutoSteer engaged. Happened 4 times on a divided non-highway and once on a 65 mph highway. The 65mph highway incident would likely have caused an accident if someone was following me even a little too closely, which is often the case.
I have a total of about 20 minutes of TACC usage. I'm pretty sure Tesla did zero testing before they put out the update. How could I and so many others find this problem so quickly but they couldn't?

 

[Lawsteve]

TACC slams on the brakes without AutoSteer engaged. Happened 4 times on a divided non-highway and once on a 65 mph highway. The 65mph highway incident would likely have caused an accident if someone was following me even a little too closely, which is often the case.
I have a total of about 20 minutes of TACC usage. I'm pretty sure Tesla did zero testing before they put out the update. How could I and so many others find this problem so quickly but they couldn't?

I have used TACC quite a bit since the update and have never experienced the brake slamming behavior you describe. Last weekend I took a 320 mile round trip from Tampa to Everglades City and used TACC for almost the entire trip with no issues (set at 75 mph while on the interstate). The only quirk I have experienced has been in changing lanes from behind a slow car to another lane with a faster car. It takes a second or two for the system to detect the car in the new lanes before it settles back in to the selected following distance. During that time it speeds up and slows down unexpectedly - sort of like it cannot make up its mind.

Hopefully all of the data being beamed up to SkyNet will allow our cars to become self-aware in the near future.

 

[Deans]

I sent tesla reports by email with the time on the car's clock, so at least they have the data.

 

[JRD1]

[snip]

But, a complication is that learning how to drive isn't all they're trying to do. If you look carefully at the most recent Tesla video, there are no lane lines! At least none that I can see. If the system is really doing DNN learning, there is no instruction to "find lane lines. stay in them." The car just stays in the lane because that's what human drivers do. When I pay no attention to the dancing lane lines and just look at the path my car is following, it is going pretty straight. Look at the lines and yikes! So, what I'm getting at is that Tesla Vision seems to aim to accomplish some algorithmic logic that requires the identification of things (lane lines, cars, people, street signs) that is very much separate from what is needed to just drive down the road. An example is setting speed based on a speed limit sign. I'm not sure a DNN could ever get enough data to generalize the presence of a 25 MPH sign to the fact that the driver is going 29 mph or whatever. Same for AEB or side collision avoidance.

[snip]

I agree with most of what you said re: the neural nets and learning.

I went back and looked at one of the videos again and you DO see the road lines but only on the dashboard. It looks like the steams on the side are the camera views with object recognition but DON'T have the road lines. With the sped-up video its hard to tell how much they dance but they do disappear from time to time.

 

[_jal_]

Hopefully all of the data being beamed up to SkyNet will allow our cars to become self-aware in the near future.

I hope not! My car would drive south from the salt-covered streets and cold, dreary skies of Chicago.

 

[_jal_]

I agree that TACC's braking is sometimes dangerous. It seems like it has mostly happened when I'm not following another car. I definitely keep an eye on the rearview mirror.

I went back and looked at one of the videos again and you DO see the road lines but only on the dashboard.

I too went back and looked and did see them after all. Maybe I should edit that. Still, there's a lot of time they're not there and the car is doing complicated stuff. Here's a capture of the movie when the car is turning and no lane lines.

In my mind, it seemed like the dancing lane lines betrayed a lot more about the inability of the car to navigate traffic than maybe it does. It is going to be interesting to see how things improve.

 

[NerdUno]

The slamming on brakes with TACC usually occurs when your car spots an overhead sign or overpass disguised as a car.

 

[NerdUno]

Today's Electrek article documents why Tesla's autonomous driving project isn't progressing quite as quickly as some of us would like. Seems they drove about 500 miles in October 2016, and most of that apparently was focused on producing Elon Musk's self-driving car video. Once that was out the door, the testing stopped almost completely. Stats like these fairly accurately document the vaporware that most AP2 buyers are experiencing.

 

[lunitiks]

Now that's a little bit disconcerning

 

[BestRadar]

Today's Electrek article documents why Tesla's autonomous driving project isn't progressing quite as quickly as some of us would like. Seems they drove about 500 miles in October 2016, and most of that apparently was focused on producing Elon Musk's self-driving car video. Once that was out the door, the testing stopped almost completely. Stats like these fairly accurately document the vaporware that most AP2 buyers are experiencing.

They got what they wanted... A Video that got suckers like me to fork over $3k for a function I might not ever see. The release we have now is no where close to where they were months ago, They seem to have gone backwards.

 

[MP3Mike]

Once that was out the door, the testing stopped almost completely.

No, they stopped testing on public roads in California. It doesn't mean that they aren't testing, or that they aren't working on fixing issues identified during that testing.

I suspect that currently most of their testing is virtual by using data captured by the AP2 HW.

 

[bushburner]

No, they stopped testing on public roads in California. It doesn't mean that they aren't testing, or that they aren't working on fixing issues identified during that testing.

I suspect that currently most of their testing is virtual by using data captured by the AP2 HW.

lets hope that is the case.

 

[gearchruncher]

Holy crap that report is depressing.

If you look at all the self driving, it happened between the 14th and 17th of October.

They were originally supposed to announce AP2 on Oct 17th, but they didn't until the 19th for "more refinement." Now we know why.

It appears it took them 550 miles and 168 disconnect events. That's a disconnect every 3.3 miles. Given they wanted a video without disconnects, it quite possibly took them 169 tries to get a video they could use.

 

[stopcrazypp]

Today's Electrek article documents why Tesla's autonomous driving project isn't progressing quite as quickly as some of us would like. Seems they drove about 500 miles in October 2016, and most of that apparently was focused on producing Elon Musk's self-driving car video. Once that was out the door, the testing stopped almost completely. Stats like these fairly accurately document the vaporware that most AP2 buyers are experiencing.

Note that EAP is semi-autonomous and would not be counted in that report. Right now Tesla is still working on EAP, so it makes sense that they don't have a lot of autonomous testing yet.

 

[BestRadar]

Holy crap that report is depressing.

If you look at all the self driving, it happened between the 14th and 17th of October.

They were originally supposed to announce AP2 on Oct 17th, but they didn't until the 19th for "more refinement." Now we know why.

It appears it took them 550 miles and 168 disconnect events. That's a disconnect every 3.3 miles. Given they wanted a video without disconnects, it quite possibly took them 169 tries to get a video they could use.

Seems like false advertising if that was true. The vehicle also stopped several times in that video when pedestrians were to the right of the vehicle.

 

[S4WRXTTCS]

Holy crap that report is depressing.

If you look at all the self driving, it happened between the 14th and 17th of October.

They were originally supposed to announce AP2 on Oct 17th, but they didn't until the 19th for "more refinement." Now we know why.

It appears it took them 550 miles and 168 disconnect events. That's a disconnect every 3.3 miles. Given they wanted a video without disconnects, it quite possibly took them 169 tries to get a video they could use.

All those numbers are potentially misleading unless you know how they're testing it.

The Google autonomous driving program has very few disengagements, but we don't know what roads they're testing it on. We also don't know if the majority of the testing is done at speeds faster than 25mph.

So we can't make heads or tails of Tesla's numbers. All we know is testing FSD is a little silly if you can't get EAP to work properly.

There were a bunch of events in 2016 that convinced me that the people in this world have lost their collective minds. Tesla pre-selling FSD ranked pretty high on these events.

 

[drshan]

I know this is old news but worth reading again. It is becoming more apparent to me that Tesla misrepresented what it knew was a long and arduous task of developing a algoritm to match and surpass Mobileye (who spent the last 18 years developing btw).

Mobileye Says It and Tesla Are Never Ever Getting Back Together

Its a phenomenal vehicle- no doubt about that. But to charge a less than informed customer 5k (2x the prior cost) for a system that is no where near the capability of the previous product (and may not be for years to come) is suspect in the very least. I think we can all agree on that.

 

[stopcrazypp]

I know this is old news but worth reading again. It is becoming more apparent to me that Tesla misrepresented what it knew was a long and arduous task of developing a algoritm to match and surpass Mobileye (who spent the last 18 years developing btw).

Mobileye Says It and Tesla Are Never Ever Getting Back Together

Its a phenomenal vehicle- no doubt about that. But to charge a less than informed customer 5k (2x the prior cost) for a system that is no where near the capability of the previous product (and may not be for years to come) is suspect in the very least. I think we can all agree on that.

A lot of the extra cost is to recoup the extra hardware. It has 4x the cameras (8 vs 2), the ultrasonic sensors have double the range (~30 ft vs ~16 ft), and the new PX 2 processor is drastically more powerful than the Mobileye one. Lest people make the point about how it is included in the base car, keep in mind in AP 1.0 the hardware was also included, but they recoup the cost when you order AP to activate it.

And to go down history lane: AP 1.0 was an option in October 2014 for $2500. Autosteer didn't get activated until October 2015, a year later.

AP 2.0 was announced October 2016, some limited Autosteer is activated starting January 2017. Just some perspective so people are aware of how Tesla does rollouts and that this is not something unique to AP 2.0.

 

[Saghost]

I know this is old news but worth reading again. It is becoming more apparent to me that Tesla misrepresented what it knew was a long and arduous task of developing a algoritm to match and surpass Mobileye (who spent the last 18 years developing btw).

Mobileye Says It and Tesla Are Never Ever Getting Back Together

Its a phenomenal vehicle- no doubt about that. But to charge a less than informed customer 5k (2x the prior cost) for a system that is no where near the capability of the previous product (and may not be for years to come) is suspect in the very least. I think we can all agree on that.

I don't see anything in that article which gives any insight into either what portions of Autopilot functions were being performed by Mobileye or how difficult it is to replicate - and certainly nothing to indicate that Tesla knew it would be very hard and deliberately lied to customers as you're suggesting.

 

[calisnow]

I know this is old news but worth reading again. It is becoming more apparent to me that Tesla misrepresented what it knew was a long and arduous task of developing a algoritm to match and surpass Mobileye (who spent the last 18 years developing btw)

For what it's worth, Mobileye and Nvidia claim to be taking very different approaches to learning. Mobileye did/does a lot of curated, human-annotated image training. Nvidia's approach is to brute force unsupervised learning with massive computational power running more complex neural networks. I guess time will tell if Nvidia's approach can equal and/or surpass Mobileye's.