AutoPilot Crash today-Tesla response less than stellar?

[Tam]

"They're trying to sell auto-pilot..."

During a test drive on Model X, I asked my Tesla Rep whether I could turn on Autopilot now and she said "No! Not this type of road."

After reading this thread, I am now glad that she was stricter and adhered to the rules.

 

[reynirb]

Dude, what the car sees as stationary wouldn't be an object in front of it where the distance doesn't change, the car would know that means it is actually something that is moving at the same rate of speed as it is. An object that is getting pinged by the radar as approaching at close to the velocity of the car is what the car sees as a stationary object. How you don't get this when you specifically mention police radars being adjusted to take into account the police cruiser's own velocity is a little mind boggling. As others have mentioned, the car ignores stationary objects because they could be a variety of different objects (road sign, soda can, trash can, even a car in the next lane that is stopped on a curved road, or a car that is stopped in a lane over while you're changing lanes) and relies on driver input. At a determined distance where it's clear the object isn't any of these, that's when AEB kicks in for cars with only radar, which is why most (all?) manufacturers have language in their safety system descriptions that the AEB mitigates crashes and generally make no mention of it completely preventing them. Cars with cameras can use image processing to supplement the radar, but even then it isnt super reliable as of yet.

Like I said, a radar sees an approaching object approaching at the speed of the vehicle. If the vehicle in front of your is traveling the same speed as you are traveling, it is stationary.

If you are driving at 70MPH and you approach a car doing 50MPH, the radar sees that object approaching you at 20MPH.

Now, a radar can be programmed to ignore an object approaching you at 70MPH if you are driving 70MPH as that velocity would indicate a stationary object. But that is not how the radar sees the object. The radar sees the object approaching at your speed.

That you are unable to Google a simple search on Doppler Radar and read how it works is mind boggling.

That you completely don't understand how a radar works is understandable, but a vehicle radar used for Adaptive Cruise Control or AEB system will ignore a return from a soda can or a very small object like a waste basket. There is a threshold programmed into any radar, where a return scatter smaller than X is ignored. And for a good reason. If a radar responded to every soda can and road debris, the system would be effectively useless. They simply use a signal threshold filter where smaller objects are ignored.

However, most cars are pretty flat on the back, and provide and excellent strong return signal back which the system would use to respond to.

The AEB system calculates the velocity of the object and the estimated brake distance, and responds accordingly if the driver fails to react in time.

As noted however, a radar can't effectively respond to radar scatter that is above or below the object, or if they are using Lidar, it would rely on something reflective, as Matte objects, such as a dirty car, don't reflect those quite as well. Which is the same reason cops use your headlights or front license plate when using Lidar.

There can be multiple reasons why stationary objects present a problem for the system however, the #1 probably being an insufficient response time for the system to respond to something approaching you at that speed.

 

[bradhs]

Seems like the hill was too much angle for radar and camera to pick up the car in front. That's basically it. Salesperson probably didn't realize that and the driver should only do what they're comfortable with. I only allow Autopilot to do what I feel very comfortable with, anything less and I'm already in control.

 

[Mike K]

Is there really any debate that the failure here lies not with the car but with the employee? Enabling it on this type of street is the first warning sign followed by the fact that the TESLA employee told him not to brake for a red light... What in the everliving...

I can understand the OP's apprehension to follow through on his purchase after the accident but that too comes from a position of ignorance as this isn't an accident you should rely on this (or any) car to avoid.

 

[eloder]

I am not interested in getting anyone fired. I am interested in how Tesla responds to someone who really wants to drive a Model S but has had an experience like this. Thank you for the information on the hills. If I do end up getting one, I will be very prudent in its use.

If you use seat belts or air bags incorrectly, they can and will kill you. Even if they're used correctly they sometimes will still be fatal.

Autopilot, like a seat belt, is not an all-encompassing solution. It works great when used how it's supposed to be used, and not so much when it's used in a manner it's not meant to be.

The salesperson is 100% at fault here for advising you otherwise, for sure, but I don't think you should let this incident decide on whether to buy the car or not. Just like seat belts, be sure to review where AP is supposed to and not supposed to work (it's 100% in the manual) and go from there.

 

[croman]

Dude, your car was totaled and you're around discussing whether to buy it. Its clearly safe -- you crash tested it yourself!

 

[Boatguy]

The OP's comment was not about the AP technology, but about Tesla's lack of response to the incident.

In my experience, this is what is to be expected. Tesla has no customer support. They have service to fix your broken car, and they have a corporate PR organization to spin EM's latest story or tweet, but nothing in between; no organization or personnel that care about you as a customer. Sales people care about the sale, and after that nobody cares about your experience or situation, unless the car is "broken", which basically means they have a part or procedure they can apply to fix what is broken. If it doesn't fit into "broken", then you will literally find nobody at Tesla interested in talking to you; it's not anybody's job to talk to you.

It's sort of weird, but that's Tesla corporate culture: we have your money, now go away and don't bother us until you want to buy another one.

 

[⚡️ELECTROMAN⚡️]

Sounds like you won't be buying another one.

 

[Boatguy]

Sounds like you won't be buying another one.

That's entirely dependent on what my choices are in 2019. Tesla is at the moment the only long distance EV and only luxury EV. But it would be fair to say that Tesla is not doing anything to create customer loyalty.

 

[Stiction]

That's not necessarily true. Any car that relies on K-band or 77GHz doppler radar has this inherent problem that stationary traffic is not easy to detect because it looks like any stationary feature on the road, such as a speed bump or a piece of debris on the road or even an overhead sign / traffic light.

They all rely on cameras and image recognition to recognize a car, and certain things can prevent that from working correctly, ranging from an older system that was not trained against modern cars (my previous 3-year-old Audi stopped recognizing the Fiat 500 and Ford Transit Connect vans by the time I turned it in), weirdly painted vehicles or lack of brake lights, etc etc etc.

Inherently, only the LIDAR based systems can reliable measure and range large solid objects, but these systems also have the additional problem that from a long distance, it's difficult to target non-LIDAR-reflective materials like matte paint jobs and it can mis-read reflective objects as threatening (e.g. a garbage truck that has paper flying out of it looks the same to LIDAR as a truck throwing metal blocks at you)

THIS posting is the one to read to understand TACC and AEB limits (and Lidar limits). Lidar also doesn't work in snow and rain very well.

 

[reynirb]

The OP's comment was not about the AP technology, but about Tesla's lack of response to the incident.

Tesla has no customer support. They have service to fix your broken car, and they have a corporate PR organization to spin EM's latest story or tweet, but nothing in between; no organization or personnel that care about you as a customer. Sales people care about the sale, and after that nobody cares about your experience or situation, unless the car is "broken", which basically means they have a part or procedure they can apply to fix what is broken. If it doesn't fit into "broken", then you will literally find nobody at Tesla interested in talking to you; it's not anybody's job to talk to you.

It's sort of weird, but that's Tesla corporate culture: we have your money, now go away and don't bother us until you want to buy another one.

I'm not sure how that is any different from any other car company out there. When was the last time you called Nissan, Ford, GM, Toyota, Hyundai or anyone else and they cared about you as a customer, or your experience or situation in general?

Just curious on what your expectations are and why they would be different from any other car maker out there. I certainly haven't exactly found any warm and fuzzy car makers out there that generally care how my day to day experience has been.

 

[reynirb]

THIS posting is the one to read to understand TACC and AEB limits (and Lidar limits). Lidar also doesn't work in snow and rain very well.

Any sensor grid has an inherent limitation in certain conditions. Cameras struggle through dirty lenses. Lights get reflected off show, water and don't get reflected well of matte objects. Radar limitations are more dependent on beam shaping and detection thresholds, along with how the beam is aimed. I seriously doubt a radar is concerned about debris unless it is a couch or a end table, it is not going to be fooled by a Coke can bouncing on the road or a piece of paper.

I also don't understand how a radar would see a traffic light or an overhead sign, as it would be above its beam path.

Any system would be designed to cross check against different sensors to validate whether the object ahead was an issue. It would be the same reason that the system needs to learn over time what is a threat and what isn't, and learn the signature of various objects.

 

[chillaban]

Any sensor grid has an inherent limitation in certain conditions. Cameras struggle through dirty lenses. Lights get reflected off show, water and don't get reflected well of matte objects. Radar limitations are more dependent on beam shaping and detection thresholds, along with how the beam is aimed. I seriously doubt a radar is concerned about debris unless it is a couch or a end table, it is not going to be fooled by a Coke can bouncing on the road or a piece of paper.

I also don't understand how a radar would see a traffic light or an overhead sign, as it would be above its beam path.

Any system would be designed to cross check against different sensors to validate whether the object ahead was an issue. It would be the same reason that the system needs to learn over time what is a threat and what isn't, and learn the signature of various objects.

Radar is a very coarse beam and the only indication of object size is the signal strength of the received signal. Radar can resolve distances by both time-of-flight of a pulse train and by phase shift of a frequency-modulated signal.

Metal is dramatically better at reflecting radar than non-metal surfaces. So all of this is great for radar when you're measuring cars in your direct path. But half a mile away with 30-degree forward looking radar, the difference between an overhead traffic sign and a stopped SUV is extremely difficult to discriminate. They both look like distant stationary radar signatures approaching you. Except for a traffic sign, as it gets closer to you, it magically disappears because at some distance the radar beam no longer hits it, while the SUV would continue to get stronger and stronger. The question is, what is that "some distance"? If it's 200 feet and you're going 80mph, there won't be much time for a useful collision alarm. And of course, that distance changes depending on whether or not there's any hills that allow it to look like you're driving towards the sign for even longer, or if the road bends (does the car assume you always go straight, or is it detecting lanes and assuming you are following the roads?)

And for non-metal, it's even worse. A coke can near the road will have a radar reflection thousands of times stronger than a large moose which is made mostly of water (and radar reflects off trace metals/minerals in its bones and tissues). So at that point you'll really have to decide what things you brake or don't brake for.


Tesla did some clever things in 8.0, including attempting to build a 2D point cloud to make sense of radar readings. By watching how radar signatures evolve over time and cross-checking that against a fleet-learned database of consistent observations (which tend to correspond to nonthreatening stationary signatures), they have a much better shot of making radar useful as a primary sensor compared to anyone else.

 

[Boatguy]

I'm not sure how that is any different from any other car company out there. When was the last time you called Nissan, Ford, GM, Toyota, Hyundai or anyone else and they cared about you as a customer, or your experience or situation in general?

Just curious on what your expectations are and why they would be different from any other car maker out there. I certainly haven't exactly found any warm and fuzzy car makers out there that generally care how my day to day experience has been.

First of all, none of the companies you mention sell $100K - $150K (P100D) cars. So let's switch to Mercedes, BMW, Porsche, Lexus and Audi with whom Tesla is actually competing. Those are the companies who's cars I've owned and that have set my expectations. Additionally, Tesla asserted before my purchase that they were committed to "continual improvement and innovation" which I would experience through OTA updates. So Tesla set the bar higher for themselves in order to win my business.

From my ownership experience I rank them Lexus, Porsche, Mercedes, BMW, Audi. They all sell through dealers of course, and none were foolish enough to commit to improving my car after I purchased it, but any one of those is more responsive to customer issues than Tesla.

 

[reynirb]

First of all, none of the companies you mention sell $100K - $150K (P100D) cars. So let's switch to Mercedes, BMW, Porsche, Lexus and Audi with whom Tesla is actually competing. Those are the companies who's cars I've owned and that have set my expectations. Additionally, Tesla asserted before my purchase that they were committed to "continual improvement and innovation" which I would experience through OTA updates. So Tesla set the bar higher for themselves in order to win my business.

From my ownership experience I rank them Lexus, Porsche, Mercedes, BMW, Audi. They all sell through dealers of course, and none were foolish enough to commit to improving my car after I purchased it, but any one of those is more responsive to customer issues than Tesla.

My friends and family own Maserati's, Ferrari's, AMG Mercedes etc. and I haven't exactly seen those dealers been really interested in the ownership experience and responsive to customer issues. In fact, many of those dealers in my area are known for crappy customer service and incompetence of legendary proportions.

So, your post basically says, the other manufacturers weren't foolish enough to give a flying frog from day one, so they get a pass, but the Tesla does not?

Most high priced cars, at $100k plus, are limited production vehicles loaded with all kinds of technical gadgets which tend to have issues. Pretty much every exotic car out there is an expensive maintenance nightmare, so if you want to get into that area, the Tesla is a shining beacon of reliability in comparison.

My experience, and my friends and family's experience doesn't exactly reflect your ownership experience, but of course your experience may be different.

 

[reynirb]

Radar is a very coarse beam and the only indication of object size is the signal strength of the received signal. Radar can resolve distances by both time-of-flight of a pulse train and by phase shift of a frequency-modulated signal.

Metal is dramatically better at reflecting radar than non-metal surfaces. So all of this is great for radar when you're measuring cars in your direct path. But half a mile away with 30-degree forward looking radar, the difference between an overhead traffic sign and a stopped SUV is extremely difficult to discriminate. They both look like distant stationary radar signatures approaching you. Except for a traffic sign, as it gets closer to you, it magically disappears because at some distance the radar beam no longer hits it, while the SUV would continue to get stronger and stronger. The question is, what is that "some distance"? If it's 200 feet and you're going 80mph, there won't be much time for a useful collision alarm. And of course, that distance changes depending on whether or not there's any hills that allow it to look like you're driving towards the sign for even longer, or if the road bends (does the car assume you always go straight, or is it detecting lanes and assuming you are following the roads?)

And for non-metal, it's even worse. A coke can near the road will have a radar reflection thousands of times stronger than a large moose which is made mostly of water (and radar reflects off trace metals/minerals in its bones and tissues). So at that point you'll really have to decide what things you brake or don't brake for.


Tesla did some clever things in 8.0, including attempting to build a 2D point cloud to make sense of radar readings. By watching how radar signatures evolve over time and cross-checking that against a fleet-learned database of consistent observations (which tend to correspond to nonthreatening stationary signatures), they have a much better shot of making radar useful as a primary sensor compared to anyone else.

I think you may be pretty confused if you think a Coke can will have a reflection 1000x stronger than a large Moose. Radar reflects of surfaces, the density of the surface is a factor along with size, but water is pretty dense. Living creatures, including a Moose, are mostly water, and a Moose is also a horizontal object where the reflection is directly towards the source, vs a Coke can is a round object, and will scatter the reflection in all directions.

This is also the reason why radar range dramatically decreases when it rains or snows, as it causes scatter and attenuation of the signal.

The software algorithm will be the issue more than the radar itself, if it is designed to ignore stationary objects. That however isn't a radar limitation, it is a software limitation of the system that can't distinguish objects in your path from objects that are out of your path.

Also, I think your distance specs are nowhere close to reality. Typical radar systems use a dual or single radar sources, one sees typically 100 feet and the long range radar sees typically 600 feet. There is no car based radar system in use that has a 30 degree vertical beam width and sees 2500 feet ahead of the car.

Radar, like any transmitted energy. dissipates square of the distance, so a radar designed to see 600 feet would never see a signal from 2500 feet away.