A cautionary autopilot video

[StealthP3D]

What does this have to do with AEB?

AEB is used by Extended AutoPilot in such situations. AEB comes on all Model 3's and it is also incorporated into EAP.

I'm being critical of it not because we really expect it to detect a pedestrian in the dark at this point, but the fact that it doesn't illustrates that we're pretty far away from FSD.

I'm not sure it didn't detect the pedestrian in the dark. If it did, it knew it could bring the car to a complete stop before impact and didn't need to react until it had more certainty.

You see, these driver aids get lambasted if they react when nothing is there and they get lambasted if the human picks up on it first (even if there was still plenty of time for it to handle the situation). You may have forgotten this but the human application of the brake cancels any action of the computerized driver aids.

So, no, I don't think this video illustrates how close or far we are from FSD.

 

[StealthP3D]

99% of FSD is moron detection/protection.

True, and it might be the moron in the driver's seat.

 

[S4WRXTTCS]

AEB is used by Extended AutoPilot in such situations. AEB comes on all Model 3's and it is also incorporated into EAP.

I'm not sure it didn't detect the pedestrian in the dark. If it did, it knew it could bring the car to a complete stop before impact and didn't need to react until it had more certainty.

You see, these driver aids get lambasted if they react when nothing is there and they get lambasted if the human picks up on it first (even if there was still plenty of time for it to handle the situation). You may have forgotten this but the human application of the brake cancels any action of the computerized driver aids.

So, no, I don't think this video illustrates how close or far we are from FSD.

I disagree on both accounts.

AEB is a crash mitigation system designed to cut the speed by 25mph in an event of a unavoidable accident. It's not a crash prevention system like other manufactures have. It is a safety component that comes on all Tesla cars that have been sold since late 2014 or so. The fact that it comes standard on all vehicles is one of the cool things about it. AEB really shouldn't be coming into play unless it's a poop your pants sort of moment. AEB shouldn't have played any role in this as it wasn't an especially difficult detection scenario for a machine.

So it's strictly an Autopilot thing that relies on TACC for braking in case of a detection event. As a TACC component it should always be a predictable response where the driver has developed a trust relationship with the car. In all my experience with autopilot it's way more of a cautious driver than I am where its often alerting me or slowing down before I really need to.

I'm well aware of the fact that human application of the brake cancels any computerized driver aids, but that isn't applicable as the OP expected a detection response before they took over. From the video it also appeared to be a last second take over event, but it's hard to tell distance/speed from a 2D video. So I'm mostly giving the OP the benefit of the doubt on this one.

What I'd love to see is detection overlays on the dashcam video. That way when reviewing dash cam videos we can see what the system detected. It could be like you said in that it saw the pedestrian, but was waiting for additional confirmation (maybe additional detection's to make sure it wasn't a fluke). But, I'm skeptical as AP2 is overly sensitive, and is known in the TMC community as being prone to false braking.

 

[StealthP3D]

I disagree on both accounts.

AEB is a crash mitigation system designed to cut the speed by 25mph in an event of a unavoidable accident. It's not a crash prevention system like other manufactures have. It is a safety component that comes on all Tesla cars that have been sold since late 2014 or so. The fact that it comes standard on all vehicles is one of the cool things about it. AEB really shouldn't be coming into play unless it's a poop your pants sort of moment. AEB shouldn't have played any role in this as it wasn't an especially difficult detection scenario for a machine.

So it's strictly an Autopilot thing that relies on TACC for braking in case of a detection event. As a TACC component it should always be a predictable response where the driver has developed a trust relationship with the car. In all my experience with autopilot it's way more of a cautious driver than I am where its often alerting me or slowing down before I really need to.

I'm well aware of the fact that human application of the brake cancels any computerized driver aids, but that isn't applicable as the OP expected a detection response before they took over. From the video it also appeared to be a last second take over event, but it's hard to tell distance/speed from a 2D video. So I'm mostly giving the OP the benefit of the doubt on this one.

What I'd love to see is detection overlays on the dashcam video. That way when reviewing dash cam videos we can see what the system detected. It could be like you said in that it saw the pedestrian, but was waiting for additional confirmation (maybe additional detection's to make sure it wasn't a fluke). But, I'm skeptical as AP2 is overly sensitive, and is known in the TMC community as being prone to false braking.

OK, I defer to your superior expertise. Because you are obviously an expert on all this. I would like to meet you someday and pick your brains.

 

[AlanSubie4Life]

What I'd love to see is detection overlays on the dashcam video. That way when reviewing dash cam videos we can see what the system detected.

Came back here because I wanted to say exactly this today. It would be great for the system to save two copies, one without overlay and one with, just in case the overlay obscured anything important. It would be good to have the vehicle detection overlay as you suggest, the driving mode (EAP or not, etc.) annotated, current speed, what pedals are being pressed, steering wheel position, g-forces (lateral and forward/backward), turn signal state, etc. Would be even more helpful for accident claims I would think. (Though it could go either way!)

From the video it also appeared to be a last second take over event, but it's hard to tell distance/speed from a 2D video.

Yes, it looked last second, but it's really hard to tell how fast cars are traveling in any of these TeslaCam videos.

So it's strictly an Autopilot thing that relies on TACC for braking in case of a detection event.

First, I don't know exactly what you mean by this (I'm not trying to be snarky, just actually not sure I understand all of what you are saying with this statement - and you may well be correct). That being said:

Regarding the capabilities of TACC & AEB, I am curious about this...Tesla has claimed that safety features will always be included (even without EAP), so I wonder whether TACC & Autosteer actually provide any additional benefit in a case like this. AEB will override the accelerator position and brake the car, unless you press the accelerator hard during an AEB event. The OP was using EAP. But even a driver not using TACC with their foot on an accelerator to maintain speed on the street in the video should be no worse off in a situation like this than a driver using EAP. That is my understanding at least. I'm not saying that is the case (I don't know), but I guess I'm saying that my belief is that for a suddenly/unexpected detected obstacle, an EAP/TACC-equipped and non-EAP/TACC-equipped vehicle should behave the same (unless steering is required?), assuming the car actually successfully detects the obstacle. Of course the behavior would be different when the object in question is "previously known/expected" - i.e. a "followed" vehicle (can't track without TACC) - only the TACC-equipped vehicle will slow and prevent a collision with a tracked vehicle in front that slows down (most of the time). Whereas the Non-TACC will presumably just mitigate (and maybe prevent at lower speeds) the collision, via AEB. But in the OP's case we are not dealing with an identified "followed" object. So I don't see that the car would behave any differently in this case regardless of whether or not is was equipped with EAP.

Getting away from speculation, though:
1) TACC cannot detect all objects and may not brake for stationary vehicles or objects, especially in situations when you are driving over 50 mph and in situations where a vehicle you are following moves out of your driving path and a stationary vehicle or object is in front of you. Depending on TACC to avoid a collision can result in death. In addition, TACC may react to vehicles or objects that do not exist.

2) Regarding Autosteer, it should never be used on city streets, in construction zones, or in areas where bicyclists or pedestrians may be present.

3) AEB is designed to reduce the severity of an impact. It is not designed to avoid a collision. 
 It cannot detect all objects, vehicles, bikes, or pedestrians. When driving 35mph or more, it will release your brakes after it has reduced your speed by 30mph. (So I guess this means it is intended to reduce the speed by 30mph, not 25mph, but whatever, close enough.)

 

[S4WRXTTCS]

Came back here because I wanted to say exactly this today. It would be great for the system to save two copies, one without overlay and one with, just in case the overlay obscured anything important. It would be good to have the vehicle detection overlay as you suggest, the driving mode (EAP or not, etc.) annotated, current speed, what pedals are being pressed, steering wheel position, g-forces (lateral and forward/backward), turn signal state, etc. Would be even more helpful for accident claims I would think. (Though it could go either way!)



Yes, it looked last second, but it's really hard to tell how fast cars are traveling in any of these TeslaCam videos.



First, I don't know exactly what you mean by this (I'm not trying to be snarky, just actually not sure I understand all of what you are saying with this statement - and you may well be correct). That being said:

Regarding the capabilities of TACC & AEB, I am curious about this...Tesla has claimed that safety features will always be included (even without EAP), so I wonder whether TACC & Autosteer actually provide any additional benefit in a case like this. AEB will override the accelerator position and brake the car, unless you press the accelerator hard during an AEB event. The OP was using EAP. But even a driver not using TACC with their foot on an accelerator to maintain speed on the street in the video should be no worse off in a situation like this than a driver using EAP. That is my understanding at least. I'm not saying that is the case (I don't know), but I guess I'm saying that my belief is that for a suddenly/unexpected detected obstacle, an EAP/TACC-equipped and non-EAP/TACC-equipped vehicle should behave the same (unless steering is required?), assuming the car actually successfully detects the obstacle. Of course the behavior would be different when the object in question is "previously known/expected" - i.e. a "followed" vehicle (can't track without TACC) - only the TACC-equipped vehicle will slow and prevent a collision with a tracked vehicle in front that slows down (most of the time). Whereas the Non-TACC will presumably just mitigate (and maybe prevent at lower speeds) the collision, via AEB. But in the OP's case we are not dealing with an identified "followed" object. So I don't see that the car would behave any differently in this case regardless of whether or not is was equipped with EAP.

Getting away from speculation, though:
1) TACC cannot detect all objects and may not brake for stationary vehicles or objects, especially in situations when you are driving over 50 mph and in situations where a vehicle you are following moves out of your driving path and a stationary vehicle or object is in front of you. Depending on TACC to avoid a collision can result in death. In addition, TACC may react to vehicles or objects that do not exist.

2) Regarding Autosteer, it should never be used on city streets, in construction zones, or in areas where bicyclists or pedestrians may be present.

3) AEB is designed to reduce the severity of an impact. It is not designed to avoid a collision. 
 It cannot detect all objects, vehicles, bikes, or pedestrians. When driving 35mph or more, it will release your brakes after it has reduced your speed by 30mph. (So I guess this means it is intended to reduce the speed by 30mph, not 25mph, but whatever, close enough.)

For the overlay it can actually be kept in a separate file, and maybe Tesla can have their own viewer. That way you could see the video, overlay data, and any logging info (like everything you mentioned). So you could the entire event as the car saw it. That's probably wishful thinking, but it would be cool. One of the things that bugs me about the Model 3 is it's hard to see what messages pop up. I don't want to take my eyes off the road when something is happening, but I'm still curious about what came up. I really wish their was a notification history at the very least.

What I meant about "it's strictly an autopilot thing" is autopilot is compromised of two parts. The TACC part that's responsible for braking/accelerating and the Autosteer. TACC has freedom in that it's allowed to false brake, and in fact TACC on AP2.5 false brakes quite a bit (the insurance institute did a study on this). AEB is really tuned not to false brake as a sudden 30 drop in speed would be incredibly dangerous/jarring. Speaking of AEB I didn't realize they upped the AEB braking from 25mph reduction to 30mph, and clarified the language a bit from the last time I looked at it. That 5mph might not seem like a lot, but from an injury protection it can be the difference between life and death. So I'm glad you pointed out my mistake in assuming it was still 25mph.

I'm of the firm belief that TACC when used correctly can be safer than driving without it. It's a second set of eyes that can watch out, and it doesn't suffer from contextual problems. That's probably not the best way to put it, but my brain has stopped working at this late hour. What I mean by that is humans watch out for things they expect. No one expects to see some dude standing in the middle of a busy road at night. I honestly may have not seen him myself.

I don't use TACC or AP myself in situations like this. I don't because TACC isn't situationally aware. It ignores the context of the situation and simply follows the set speed unless something is in the way. It could care less if it zoomed past a car at a 30+ mph differential. Plus there is no stop light detection, and all the limitations the radar has like how it doesn't allows see stopped objects for example.

So while it has the potential to be safer with a knowledgeable driver it tends to be more dangerous in this situation due to all the drawbacks that an average Tesla driver might not know about. Autopilot (or TACC only) in this situation will probably be safer in a few years. Where it can react to a situation faster than FCW+driver reaction time, and won't suffer from the limitations it currently has.

For now a person driving without EAP/TACC should be able to rely on FCW as their second set of eyes.

When it's all said and done I think it's irrelevant if it's AP/TACC or manual driving with FCW. In all those cases we're relying on the HW2.5 computer running a neural network that's supposed to detect pedestrians. So even if an idiot is standing in the road we have some warning.

The video posted lacks any of the overlay/logging data that we both wished was there so we don't know if FCW sounded.

We only know the OP said they didn't hear it. I certainly wouldn't expect them to look at the screen to see what it said while they were braking.

 

[StealthP3D]

The video posted lacks any of the overlay/logging data that we both wished was there so we don't know if FCW sounded.

We only know the OP said they didn't hear it. I certainly wouldn't expect them to look at the screen to see what it said while they were braking.

Again, once the driver applies the brakes, there will be no Forward Collision Warning. When AEB kicked in during my previously detailed AEB experience, FCW kicked in simultaneously with the Automatic Emergency Braking. Had I stepped on the brake 1/10 second before the AEB did, there would be no audible warning at all.

 

[S4WRXTTCS]

Again, once the driver applies the brakes, there will be no Forward Collision Warning. When AEB kicked in during my previously detailed AEB experience, FCW kicked in simultaneously with the Automatic Emergency Braking. Had I stepped on the brake 1/10 second before the AEB did, there would be no audible warning at all.

In the OP's scenario my expectation would be that FCW would sound ahead of AEB because the car was driving towards a person standing in the lane. So it's not a sudden event like someone cutting in front of me which would cause simultaneous FCW/AEB.

If it didn't go off before then I'd set FCW to early.

But, all the discussion is getting away from the main point. The main point the OP was making is we can't rely on TACC/AP to handle this kind of situation in an acceptable manner.

Now I don't think they should have been using AP on this kind of road (like they already acknowledged), but that's because Autosteer can cause them to veer into the oncoming lane. My expectations for TACC would be to exceed a humans ability to see a pedestrian at night in the rain. We know that didn't happen as the driver took over, and braked much harder than TACC does. The OP posted it because they felt like it was an last second oh, crap moment.

Imagine the kind of human/machine bonding that would have occurred had it braked smoothly for this guy the OP wouldn't have even seen until they wondered why the car was slowing down. I don't see any technical reason why that shouldn't have happened. The car is billed as having the sensor suite required for full self driving, and has a neural network that includes pedestrian detection.

At the end of the day it simply didn't work. Sure I understand that my expectation might exceed what's currently there, and my entire position might be premature.

There isn't anything about AEB or FCW within that. The lack of FCW is a bit concerning, but we don't know what the driver had that set to. The AEB not braking isn't surprising as the driver likely took over before that comes into play.

 

[Russell]

I disagree on both accounts.

AEB is a crash mitigation system designed to cut the speed by 25mph in an event of a unavoidable accident. It's not a crash prevention system like other manufactures have.

What manufacturers have AEB crash prevention systems?

 

[StealthP3D]

At the end of the day it simply didn't work. Sure I understand that my expectation might exceed what's currently there, and my entire position might be premature.

I've pointed out multiple times that none of those features will kick in if the driver is already braking. In other words, there is no evidence that it was not going to work - we simply don't know.

So stop acting like this is evidence of a failure of the system. The only way you can get that kind of evidence is if the pedestrian was actually hit without the driver having taken action and without the system kicking in.

 

[Vegan Pleather]

I am alarmed that 66% of poll takers are saying that they use autopilot under a condition the manual says it should not be used.

When will I see a cautionary tale about a Tesla used on secondary roads failing to recognize a light has turned red?

 

[StealthP3D]

When will I see a cautionary tale about a Tesla used on secondary roads failing to recognize a light has turned red?

Failing to notice a light turning red is independent of whether or not AP is engaged.

Arguably, the driver has more attention to devote to such things when the car is maintaining proper following distance and lane-keeping.

 

[Vegan Pleather]

That's not be point. Maybe I should have clarified by saying a "Tesla on Autopilot", but I digress.

 

[S4WRXTTCS]

I've pointed out multiple times that none of those features will kick in if the driver is already braking. In other words, there is no evidence that it was not going to work - we simply don't know.

So stop acting like this is evidence of a failure of the system. The only way you can get that kind of evidence is if the pedestrian was actually hit without the driver having taken action and without the system kicking in.

So you're classifying the driver having to brake really hard, and the pedestrian stepping out of the way at the last second as not being a failure of AP?

I completely agree with that it's not evidence of FCW or AEB failing as the driver took over before they kick in, but those features aren't being discussed. What's being discussed is AP.

I call it a failure because AP didn't detect the pedestrian early enough to slowdown/brake without any dramatics.

How do I know it didn't? I'm giving the OP the benefit of the doubt, and from what the video shows me.

 

[StealthP3D]

I call it a failure because AP didn't detect the pedestrian early enough to slowdown/brake without any dramatics.

So, you're saying the human driver (the OP) failed also? In my mind, considering the circumstances, I would say anything that avoided impact was 100% successful.

 

[S4WRXTTCS]

What manufacturers have AEB crash prevention systems?

Lots of cars have AEB systems designed to stop at up to a certain closing speed. For some reason Tesla doesn't discuss closing speed, and simply mentions how much it will decrease the speed by.

In actual tests of these systems other manufactures do better in coming to a complete stop when AEB engages. In this article they refer to AEB as autobrake for some reason.

Tests uncover issues for advanced features

Now I wouldn't necessarily say these other cars have better AEB systems. It's a balance between false positive, and false negatives. I'd rather have a false negatives (for AEB) than false positives as I pay attention when I drive. I don't have issues rear ending people so I'd rather my car didn't false brake while being manually driven. I've never experienced a false AEB braking event my years of driving Tesla vehicles, and I'm happy with that.

 

[S4WRXTTCS]

So, you're saying the human driver (the OP) failed also? In my mind, considering the circumstances, I would say anything that avoided impact was 100% successful.

The human driver met or exceeded expectations.

But, expectations are low for typical squishy water filled human meatbags with sub-optimal sensing capability.

 

[StealthP3D]

The human driver met or exceeded expectations.

But, expectations are low for typical squishy water filled human meatbags with sub-optimal sensing capability.

Whatever...yada, yada, yada. I guess the AP failed then when I let it save me from hitting a car pulling out in front of me because I could have stopped sooner? So it's all about failure with you.

 

[hill]

Irony of ironies. The video, if anything, exemplifies how primitive our Tesla driving assist systems are, in relation to full autonomy. Especially in Cali, where dumb clueless pedestrians have the right away? Forgetit - and why not. If you're on a back road daytime or night & a cow or deer trots out in front of you, much less a human, you're not expected to run over a pile of stupidity. To blame it on that stupid person for doing what stupid people (or enen a blind person might) naturelly do it is just completely
irresponsible. End of story. On top of that, consider how anyone who previously & regularly drove AP1. Most of them will confirm that the ap2 system is not much - if at all better for those accident detections - after driving both systems. Why else do you think, that many outraged & frustrated ap2 system buyers were willing to accept "class-action" settlements/payments for their purchasing an accident detection sytem that was subpar to their prior ap1 system, yet causing them (on false promises) to get anotherTesla car for the better system? Including yours truly?

Do you think Tesla owner/fans wanted to go this route? Hell no! Even so, those of us who picked up a paltry settlement for the ongoing autopilot failure hope to send a message to Tesla to stop with the over-promising that was just intended to bring in more customers. Still despite all of the over-promising, Tesla is wonderful, a great car, a leader, light years ahead of the competition. But when owners rely on over promises that were simply stated so buyers would bring in more business, then that's just an overpromise. Yet despite the over-promises, many of us owners of both AP1 & ap2 still say, "go Tesla".
One more comment on stupidity. In another post I commented on the anguish of taking part in a Tesla class action. A detractor mockingly laughed about a settlement that was only a couple hundred dollars for one person ..... & maybe a few hundred dollars for another, maybe only $100 for another. Maybe it would have made them happier if 100's of $1,000 of dollars were granted to each? So yeah, how stupid would that be. That was just to say .... how stupid some of us over the top Tesla fans can be .... of which - ironically I still consider myself. Why else would we have purchased a model X AP2 - after purchasing our earlier Model S.
.

 

[Russell]

Lots of cars have AEB systems designed to stop at up to a certain closing speed. For some reason Tesla doesn't discuss closing speed, and simply mentions how much it will decrease the speed by.

In actual tests of these systems other manufactures do better in coming to a complete stop when AEB engages. In this article they refer to AEB as autobrake for some reason.

Tests uncover issues for advanced features

Now I wouldn't necessarily say these other cars have better AEB systems. It's a balance between false positive, and false negatives. I'd rather have a false negatives (for AEB) than false positives as I pay attention when I drive. I don't have issues rear ending people so I'd rather my car didn't false brake while being manually driven. I've never experienced a false AEB braking event my years of driving Tesla vehicles, and I'm happy with that.

Again, what manufacturers have AEB crash prevention systems? None. They all say they only reduce impact.