Fatal Flaw in AP2 Design

[VikH]

@VikH Better brush up on your reading comprehension. Inability to detect lane lines is not the same as "cresting a hill." In our example, there was no hill.

Thank you for being courteous Your first post is not clear. You state there is a dip in the road and that AP disengages, but then also say the car swerves. The car is swerving with AP off?

Well, yes, it does require more mental effort because it requires you to be able to take over and correct in a fraction of a second, which is insufficient to gain complete situational awareness, therefore it follows you must always be as aware of your surroundings as if you were driving yourself or more. Think driving instructor, does it take more or less mental effort to drive with a student driver? After all, most students can keep the car in the lane probably better than Tesla AP (1 or 2).

Now, just because the AP fine print requires it, doesn't mean you can't get away with less. Lots if people do it and some eventually end up in accidents. Statistically speaking there will be people who will not get into accidents while completely trusting AP too. Just because someone survived Russian Roulette doesn't make it a safe game to play.

To be clear, I am talking about usage of AP on long distance highway travel (not local driving). Being aware of your surroundings is not the same as driving (watching the road+modulating throttle+steering input). Also, I never said I completely trust AP. There are several scenarios where AP does not do well, as mentioned previously.

I never said driving with AP is like playing Russian Roulette. I was using the game to illustrate the fallacy of "lots of people do something and they are fine, so must be safe " used by a lot of people in this thread

Who said AP is 100% safe? You are assuming that everyone using (or misusing) AP has no clue of the risks or issues with the system. I think a lot of people do understand the risk and they manage that risk accordingly.

True, unfortunately I don't have stats (Tesla does not release those). Insurance companies do, but it takes time before they are public. We do know a number of insurance companies have been hiking the Tesla rates recently, read into this what you will.

That is a pretty big leap and first I have heard that insurance rates are increasing because of AP caused accidents. How do you know all insurance companies are hiking rates for Teslas across the country? If true, there are several factors (unknown to us) that may be causing high rates - maybe its related to the expensive aluminum body repairs and limited number of Tesla body shops.

 

[whitex]

To be clear, I am talking about usage of AP on long distance highway travel (not local driving). Being aware of your surroundings is not the same as driving (watching the road+modulating throttle+steering input). Also, I never said I completely trust AP. There are several scenarios where AP does not do well, as mentioned previously.

AP on local roads requires extra vigilance and almost certain interventions, but that is not what I was talking about. I was talking about highway driving. Now as you point out, even on the highway AP doesn't do well in all scenarios. The problems with that are:
1. Those scenarios can change from version to version
2. New scenarios where it doesn't do well may be added. If you are the first to encounter it (rare scenario) or missed that thread on TMC, you risk getting into an accident unless you're paying extra attention.
3. A number of those scenarios are hard to understand and therefore reliably identify by non-technical people (the need to understand how radar works, how cameras work, what they see, etc). This further limits applicability of AP to the general public.

In my mind the status of AP is "experimental" at best. Just like an experimental plane, it requires a qualified pilot to operate safely because a regular pilot may get in a lot trouble with it. Experimental planes do fly well, even beat records, but should not be rushed into regular service, so why should AP?

 

[AmpedRealtor]

NO the car WILL NOT KILL YOU. It may FEEL that way but it actually won't. I can't tell you how many times mine has "swerved" (in reality probably moved inches) toward a K rail for whatever reason. Finally I left it the hell alone and guess what? It didn't hit. OF COURSE IT DIDN'T because if it's flaw was so obvious, so common and so easily repeatable, people would be crashing left and right.

I certainly hope the boldface part (emphasis mine) doesn't get you into trouble someday. If Autopilot was steering me into a guard rail, I would certainly intervene. At such a moment, the human will always knows better than an immature self-steering system.

 

[AmpedRealtor]

My biggest disappointment in this car is the TACC Cruise control - how the hell does the car slow down over a freeway underpass (elevated freeway, road under) - (not an overbridge above the freeway or an overhead sign even? but yeah I have had lots of those). Other's have figured this tech out - 2007 infiniti, 2014 BMW 2015 Lexus - I've driven all these TACC cars with never an unexpected slowdown! Even a one on a trip is one too many. It's not what you'd expect on a 90K+ car.

This is still "experimental software" and I'm suprised they made us pay $3k for it. There ought to be a disclaimer on every time the car is started.

The genius of Tesla is getting customers to pay thousands of dollars out of their own pockets to do testing that Tesla should be doing on its own. You don't see GM, BMW, Mercedes, and others using their customers as beta testers like this. Safety features, at that.

 

[heffneil]

Mercedes drives great on every road without incident with less cameras.

 

[stopcrazypp]

Mercedes drives great on every road without incident with less cameras.

Not according to reviews. Mercedes had more interventions required than AP1 (old version before 8.0).
Semi-Autonomous Cars Compared! Tesla Model S vs. BMW 750i, Infiniti Q50S, and Mercedes-Benz S65 AMG - Feature
Testing (Semi) Autonomous Cars With Tesla, Cadillac, Hyundai, and Mercedes - Motor Trend

Also, rather than simply crossing the line a bit, the Mercedes may lock on to the opposite lane and start driving the wrong way, as in this review:

"We come in fast. In this turn we let the Mercedes do what it wanted. It switched lanes and stayed in the lane for oncoming traffic. The Tesla is a little surprised, but pulls in and reduces speed. We do not need to take a grip at the steering wheel. (Foto:Jamieson Pothecary / Autofil)"

Hands off

There's a ton of talk here about how dangerous Tesla's AP1/AP2 is, but so far the independent reviews done under the same conditions don't show other systems being any better (despite being "fully tested," whatever that means).

 

[heffneil]

My S550 drive me quite a far distance on highways. I also don't have fear driving that car on the roads especially with large bends. Much more reliable than my AP 2.0 Model X

 

[hiroshiy]

You don't have to do anything, if enough people pass by an area with a false detection it'll add to the list. It knows it is a false detection since the car is able to pass right through the detected object.

In addition to the sudden swerving getting into tunnels and small bumps/up-and-downs, I also have this sudden braking issue at some of the overhead signs or objects. My commute on highway is about 6km (3.7 miles) and I counted 30 overhead objects, including signs, displays , bridges and tunnels. At least 3 or 4 out of 30 objects cause sudden braking every day. I think if Tesla relies on whitelist approach, it may not be enough.

 

[stopcrazypp]

In addition to the sudden swerving getting into tunnels and small bumps/up-and-downs, I also have this sudden braking issue at some of the overhead signs or objects. My commute on highway is about 6km (3.7 miles) and I counted 30 overhead objects, including signs, displays , bridges and tunnels. At least 3 or 4 out of 30 objects cause sudden braking every day. I think if Tesla relies on whitelist approach, it may not be enough.

I don't know how popular the Model S is in Japan, but Tesla's approach likely requires enough different vehicles passing by. If you are pretty much the only Model S passing by on that route, then it might never whitelist it.

 

[NerdUno]

I don't know how popular the Model S is in Japan, but Tesla's approach likely requires enough different vehicles passing by. If you are pretty much the only Model S passing by on that route, then it might never whitelist it.

If the object of the exercise is to identify objects that aren't dangerous to drivers (particularly those 20 feet above the road), then why would it matter if the same car identifies the object as safe a dozen times in a row versus 12 different cars identifying the object as safe. That sort of logic has kinda gotten us to where we are, wouldn't you say??

 

[stopcrazypp]

If the object of the exercise is to identify objects that aren't dangerous to drivers (particularly those 20 feet above the road), then why would it matter if the same car identifies the object as safe a dozen times in a row versus 12 different cars identifying the object as safe. That sort of logic has kinda gotten us to where we are, wouldn't you say??

Because that 1 car might be slightly miscalibrated and be giving misleading data. Having multiple cars drastically reduces that probability. This is critical if they aren't manually pruning the list (which they likely aren't).

 

[hiroshiy]

Because that 1 car might be slightly miscalibrated and be giving misleading data. Having multiple cars drastically reduces that probability. This is critical if they aren't manually pruning the list (which they likely aren't).

That's not what I wanted to mean. We have very small number of Tesla cars in Tokyo, but even if there were a lot of cars here, it doesn't make sense to whitelist objects coming every 650 feet in every road in Tokyo?? I would think Tesla need to adjust the software not to detect high-up overhead objects. I've attached one of the worst spot on the expressway, where AP2 always hard brakes (not a panic-like brake but your passengers will surprise) and then cross the lane marking to the right (sometimes to the left).

 

[jamiejag]

I had a similar experience as the OP a few days ago as I crested a slight rise in the road on an outer lane. AP2 decided to veer off the road at nearly 60mph, and had I not been alert, we may well have wound up in an A & E ward.

 

[22522]

I had a similar experience as the OP a few days ago as I crested a slight rise in the road on an outer lane. AP2 decided to veer off the road at nearly 60mph, and had I not been alert, we may well have wound up in an A & E ward.

They should integrate the accelerometer data steam into the reacquisition of roadway lines coming over crests. Idea being to steer to hold lateral acceleration constant whenever gravity derivitive is neagative and lines disappear.

 

[NerdUno]

That's not what I wanted to mean. We have very small number of Tesla cars in Tokyo, but even if there were a lot of cars here, it doesn't make sense to whitelist objects coming every 650 feet in every road in Tokyo?? I would think Tesla need to adjust the software not to detect high-up overhead objects. I've attached one of the worst spot on the expressway, where AP2 always hard brakes (not a panic-like brake but your passengers will surprise) and then cross the lane marking to the right (sometimes to the left).
View attachment 228336

Very strange. Whether a car is miscalibrated or not, why wouldn't that overpass be whitelisted once you've driven under it several times? In Atlanta, there are a dozen similar overpasses near downtown, and I've never experienced any braking going under any of them. The only difference is the number of Teslas. I simply don't buy the need for multiple Teslas to encounter a perceived obstacle before it's whitelisted. If the object is up a foot, down a foot, or over a foot, it's still a stationary object that is safe. If it means whitelisting it twice at slightly different locations based upon different Tesla camera angles, who cares? The GPS coordinates still will be the same.

 

[stopcrazypp]

That's not what I wanted to mean. We have very small number of Tesla cars in Tokyo, but even if there were a lot of cars here, it doesn't make sense to whitelist objects coming every 650 feet in every road in Tokyo?? I would think Tesla need to adjust the software not to detect high-up overhead objects. I've attached one of the worst spot on the expressway, where AP2 always hard brakes (not a panic-like brake but your passengers will surprise) and then cross the lane marking to the right (sometimes to the left).
View attachment 228336

For something like that where it is tunnel-like the height isn't the main problem, it's the depth. When the radar is shooting slightly upwards, it will hit the underside of the overpass and that may make it similar to an object of lower height at a farther distance. The ideal way is to do it like most ACC systems out there and ignore radar signatures from overhead objects, but again, Tesla decided not to do that after the Florida accident (they rather have false braking than the Florida accident again). I'm sure the algorithm will improve as they update the software, but it is what it is on this version.

 

[22522]

I think a second high radar like on the MB and Waymo self drivers solves this. I suspect they block all the returns from ground clutter.

The pair can triangulate.

 

[Mark Z]

My latest problem with AP2 are some new lane markers on the 91 freeway westbound between the 57 freeway and the 5 freeway. Caltrans decided to use a wider dashed stripe between the HOV lane and the remaining lanes. Attempting to exit the HOV lane was not successful on a morning drive about 10 AM. Twice the 2017 MS 90D would cross the lane halfway and then return to the HOV lane. Manually turning the steering wheel was the only way to make the lane change. The standard lane paint lines were no problem for all remaining lane changes with AP2.