Phantom Braking - follow distance

[paulvinh]

I have experienced autopilot phantom braking at the same underpass about 3 times now. Today I realized that all 3 times that it happened, I was on follow distance 3. All times no cars were in front of me and the car behind was about 5 car distance.

Ive been on the same route and same lane for about 1.5 months and am usually on 2 car distance because of traffic and it has been fine. Which led me to believe it phantom braking was fixed but it is not.

I hope software 9 fixes this issue.

Anyone else taking notice of what distance the car is on when phantom braking happens?

 

[Tam]

...phantom braking...

I am no expert, but as long as the system relies on radar as a primary sensor, I expect this problem will continue for a foreseeable future.

I expect the problem will be fixed when switching to LIDAR but Tesla doesn't have it so I'll take TeslaVision.

 

[S4WRXTTCS]

Have you done a TACC versus AP phantom braking test?

To determine if AP is more susceptible to phantom braking?

There are arguments both ways.

One argument is TACC is the accel/brake of AP so it's going to be the same.

The other argument is they don't try to do as much (as in collision prevention) with TACC as they do when AP is engaged.

 

[S4WRXTTCS]

I am no expert, but as long as the system relies on radar as a primary sensor, I expect this problem will continue for a foreseeable future.

I expect the problem will be fixed when switching to LIDAR but Tesla doesn't have it so I'll take TeslaVision.

If you read the report from the insurance institute (that was recently published on Electrek) it goes into detail that it was detecting shadows.

So that would be the vision system.

I have a feeling (which needs verification) that TACC is mostly if not all radar, and when AP is engaged it uses more vision. The primary reason I believe they do it that way is because that's how I would do it. Where I would feel like there was little need for TACC to do any more than Radar (as it's intended for freeway use with a driver who is engaged in the driving task, and doesn't want the car to be all that intelligent), and AP is for people who want the car to do more.

I should be able to test this a bit with an 3PD if I do take delivery of it in a couple weeks.

 

[Knightshade]

Every description I've seen of this so far says it's neither AP or TACC... it's AEB.

Because everyone describes it as heavy braking that slows them down around 25mph and then stops braking. Which is what AEB does.

TACC/AP will slow you to a dead stop if it thinks there's something solid ahead of you not moving. AEB slows you by 25 mph and expects the driver to have taken over by then or you to have already hit the thing.

 

[RedModel3]

I've had it happen in different places, with my following distance set to max. I can't think of a common link between the two times, but once it was approaching an underpass with a shadow. I'm pretty sure that the first time I was not using TACC or AP. It was certainly disconcerting, and I wonder what might have happened if a car had been tailing me, as so often happens here.

 

[Knightshade]

I've had it happen in different places, with my following distance set to max. I can't think of a common link between the two times, but once it was approaching an underpass with a shadow. I'm pretty sure that the first time I was not using TACC or AP.

That seems to confirm it's AEB. Which unlike TACC/EAP is always on.

 

[paulvinh]

Every description I've seen of this so far says it's neither AP or TACC... it's AEB.

Because everyone describes it as heavy braking that slows them down around 25mph and then stops braking. Which is what AEB does.

TACC/AP will slow you to a dead stop if it thinks there's something solid ahead of you not moving. AEB slows you by 25 mph and expects the driver to have taken over by then or you to have already hit the thing.

This might be it!

Ive also read that it doesnt happen when there are cars close by.

 

[Mobsidian]

That seems to confirm it's AEB. Which unlike TACC/EAP is always on.

If it was AEB wouldnt someone have encountered the braking while they were driving manually? I havent experienced that yet.

 

[Knightshade]

If it was AEB wouldnt someone have encountered the braking while they were driving manually? I havent experienced that yet.

The guy I quoted did experience it manually (without EAP or TACC on)

 

[TesSpartan]

I don't have auto pilot. But I have set my AEB on my M3 to it's most sensitive setting and cannot get it to work! I try to drive quickly up to a vehicle at highway speeds in the slow lane and nothing happens. Should something happen?

 

[Mobsidian]

The guy I quoted did experience it manually (without EAP or TACC on)

Sorry didnt read that part. But still it would be happening just as often on manual driving but it isnt. It would help if the screen showed what safety feature was used when a braking event occurs.

 

[xav-]

If you read the report from the insurance institute (that was recently published on Electrek) it goes into detail that it was detecting shadows.

So that would be the vision system.

I have a feeling (which needs verification) that TACC is mostly if not all radar, and when AP is engaged it uses more vision. The primary reason I believe they do it that way is because that's how I would do it. Where I would feel like there was little need for TACC to do any more than Radar (as it's intended for freeway use with a driver who is engaged in the driving task, and doesn't want the car to be all that intelligent), and AP is for people who want the car to do more.

I should be able to test this a bit with an 3PD if I do take delivery of it in a couple weeks.

AP seems to really be pretty much all radar to me.. (other than lane recognition which is obviously done by the cameras).. cameras also do a descent job at detecting parked cars unlike with AP10... but where are parked cars when you use AP as designed?

Pedestrian detection seems non existent right now

 

[Runt8]

Every description I've seen of this so far says it's neither AP or TACC... it's AEB.

Because everyone describes it as heavy braking that slows them down around 25mph and then stops braking. Which is what AEB does.

TACC/AP will slow you to a dead stop if it thinks there's something solid ahead of you not moving. AEB slows you by 25 mph and expects the driver to have taken over by then or you to have already hit the thing.

If it's really the camera detecting shadows like the Electrek article states, then it could be that by the time the vehicle has slowed roughly 25mph it has travelled far enough for the camera to detect that the shadow isn't something that should be slowed for. This would account for variations in the speed delta as well as the reason why the vast majority of cases seem to involve AP.

 

[Tam]

...vision...

After the fatal autopilot accident in Florida when the camera interpreted the bright color of the tractor-trailer as the bright skylight in the background, Tesla published its blog officially assigning radar as a primary sensor:

Upgrading Autopilot: Seeing the World in Radar

Again, I am no expert but I understand that today's car radar technology uses doppler radar to measure speed so the car detects moving obstacles and brake for them very well while filtering out stationary obstacles with the speed of zero. It needs a genius to write a program to distinguish which stationary obstacles are harmless like an overpass bridge and which ones are fatal as in Mountainview fatal autopilot accident when the stationary obstacle of the concrete median was in right front of the car and the car didn' brake because of the nature of how doppler radar works: the concrete was stationary speed of zero.

When TeslaVision will be activated, it would be competent in labeling all objects: which one it can run over such as shadows and which ones it cannot such as a stationary vehicle in front.

 

[S4WRXTTCS]

After the fatal autopilot accident in Florida when the camera interpreted the bright color of the tractor-trailer as the bright skylight in the background, Tesla published its blog officially assigning radar as a primary sensor:

Upgrading Autopilot: Seeing the World in Radar

Again, I am no expert but I understand that today's car radar technology uses doppler radar to measure speed so the car detects moving obstacles and brake for them very well while filtering out stationary obstacles with the speed of zero. It needs a genius to write a program to distinguish which stationary obstacles are harmless like an overpass bridge and which ones are fatal as in Mountainview fatal autopilot accident when the stationary obstacle of the concrete median was in right front of the car and the car didn' brake because of the nature of how doppler radar works: the concrete was stationary speed of zero.

When TeslaVision will be activated, it would be competent in labeling all objects: which one it can run over such as shadows and which ones it cannot such as a stationary vehicle in front.

The problem with that blog post is NOTHING happened. There was no follow through or updates on the progress being made.It seems like that whole thing was premature before the Engineers really had a chance to evaluate whether the white listing would work well.

It makes WAY MORE sense to have it in conjunction with Tesla vision like you described.

The insurance institute study was rather bizarre in that they didn't bother upgrading the firmware of both vehicles. They had an AP1 vehicle which they left all the way back at firmware 7.1 but the AP2.5 based Model 3 car had firmware 8.1. So it's really hard to tell if the phantom braking was a result of the difference in how the radar was being utilized between 7 and 8 or if it wasn't a Tesla Vision issue (changing from MobileEye which was a proven system to an in-house solution).

They also don't say whether the slow downs occurred with normal driving, TACC, or AP.

It simply says

"Unnecessary or overly cautious braking is an issue IIHS noted in the Model 3. In 180 miles, the car unexpectedly slowed down 12 times, seven of which coincided with tree shadows on the road. The others were for oncoming vehicles in another lane or vehicles crossing the road far ahead."

With AEB it's supposed to show an alarm in the vehicle as to why it's stopping. With my AP1 car I've never had this alarm, but I've had FCW alarms (more so after firmware 8). I've also had phantom braking with TACC, but only a few times in tens of thousands of miles.

 

[S4WRXTTCS]

AP seems to really be pretty much all radar to me.. (other than lane recognition which is obviously done by the cameras).. cameras also do a descent job at detecting parked cars unlike with AP10... but where are parked cars when you use AP as designed?

Pedestrian detection seems non existent right now

I'm not sure where it's at with AP2, but AP1 has it.

Tesla Autopilot significantly improved pedestrian detection in v8 update tests show, now renders humans

 

[xav-]

I'm not sure where it's at with AP2, but AP1 has it.

Tesla Autopilot significantly improved pedestrian detection in v8 update tests show, now renders humans

AP1 in this regard is still far superior to AP2

It’s not too surprising considering Mobileye has been trained to recognize hundreds of things over many years

I do believe though that AP20 does a better job reading lanes and detecting stationary cars so overall I’m quite happy with it (especially the lane part.. autopilot can be stressful when lane reading is imperfect).. so overall I prefer ap20.

 

[VValleyEV]

I had a phantom braking incident happen yesterday on my first long trip in a Tesla. My car is a Model S built in Feb with the upgraded autopilot and FSD option, running the new version 9 software, 2018.39.6. This was scary and could have caused an accident if someone was anywhere close behind me and not using a similar automated system.

I was on I10 east between Phoenix and Tucson, about mid-day, Autopilot engaged, leftmost lane about 80mph. Just before an overpass the car braked hard and then continued. No other warning that I saw or heard, but I could have missed it in the surprise and my reactions to look out ahead for some reason rather than the controls.

There was a very well-defined shadow strip under the overpass, so my only theory is that this was detected as a stationary object. Your experience sounds very similar. I am not sure what my follow distance was, but I don’t think I was following anyone at the time and so probably had it on a high number.

 

[Mobsidian]

I had a phantom braking incident happen yesterday on my first long trip in a Tesla. My car is a Model S built in Feb with the upgraded autopilot and FSD option, running the new version 9 software, 2018.39.6. This was scary and could have caused an accident if someone was anywhere close behind me and not using a similar automated system.

I was on I10 east between Phoenix and Tucson, about mid-day, Autopilot engaged, leftmost lane about 80mph. Just before an overpass the car braked hard and then continued. No other warning that I saw or heard, but I could have missed it in the surprise and my reactions to look out ahead for some reason rather than the controls.

There was a very well-defined shadow strip under the overpass, so my only theory is that this was detected as a stationary object. Your experience sounds very similar. I am not sure what my follow distance was, but I don’t think I was following anyone at the time and so probably had it on a high number.

I've noticed that it seems to be picking up the speed limit of roads near the overpass, like the offramp or the road above itself. Since those aren't highways it forces the car to a speed not more than 10kph over that non-highway speed limit. I've seen the blue cruise control speed change at the same time phantom braking occurs. But not always. Keep an eye out for that. Also, theres a pattern here where it only does this if theres no car behind you, everyone who reports these cases always says "luckily there was no one behind me". Tesla really needs to sort this out, its quite jarring and scary, especially for passengers.