60 mph; excellent road markings and AP2 tried to throw the car at the median

[AnxietyRanger]

Yeah, it is the fear of sudden brakings and now sudden wrong turns that is disconcerting about AP2.

It is not AP2's inability to handle unexpected events, no, it is easy to be vigilant for that. Just pay attention to the road and more so if unexpected road conditions are emerging.

It is not about AP2's inability or ability to steer steady, that has actually gotten better - but even the ping-pongs were not that bad, just a little unpleasant...

It is AP2's ability to cause unexpected events, that makes it so scary these days. That requires more than paying attention, it requires basically applying a "protective" force to the steering wheel all the time just in case...

Frankly, it is much easier to just drive yourself. It was not this bad earlier, though the ghost brakings have always been disconcerting in their own right. Somewhere along the way the steering went sour too.

 

[AnxietyRanger]

But AFAIK EAP doesn't even use all cameras, so the 3 would probably be developed parallel. At least the changes necessary because of camera placement.

EAP only uses two cameras at this time, as best we know: front main and narrow (not wide). In the future it is expected to use the side repeater cameras as well, but no more. Full eight is for FSD, if Tesla's EAP/FSD marketing is to be believed (which it is not, but that's a different story).

 

[u00mem9]

Is the camera setup in Model 3 the same as Model S? Model 3 has cameras high up in the B pillar that I don't recall seeing on any Model S, for instance. Could it be that Model 3 AP will be superior to Model S?

Sure buddy, we're all screwed but you should be fine with that model 3 reservation coupon you have there.

/s

 

[thegruf]

drove an AP1 loaner car today - used AP1 most everywhere. So cool - loved it all over again.

Note to Tesla - please don't put out AP1 loaners to AP2 owners.
The wounds are still fresh and don't need salt rubbing in ...

Now where's my f/w update -seen nothign since I started this thread not even the auto headlights one.
Definitely feeling left out.

 

[JohnnyG]

Is the camera setup in Model 3 the same as Model S? Model 3 has cameras high up in the B pillar that I don't recall seeing on any Model S, for instance. Could it be that Model 3 AP will be superior to Model S?

The camera setup is the same. Every Model S built since late October 2016 has the side B Pillar cameras.

 

[JohnnyG]

Hmm--I am pretty certain that you are describing the AP2 system. The MobileEye, AP1 system as I understand it is camera based only. However, I stand to be corrected if I am wrong. ...

AP1 uses a centered forward facing radar, in addition to the camera. As of f/w v8 it has become a shared primary system. Prior to v8 it was a secondary system.

 

[5YJSA]

Search "Truck lust" or just ""lust" on topic title and you will see this has been extensively discussed.

Weird. I did that before I posted. Saw nothing.

 

[Mark Z]

I think it happens if you cancel the indicator before the car is fully in the lane.

On September 7 the exact opposite happened. I made sure the left turn indicator wasn't touched after moving it to change lanes. The Model S 90D moved halfway over, then quickly returned to the original lane. I didn't touch the turn indicator and it eventually made the lane change.

It's almost as if someone added an "abort lane change" routine in the computer code and when a false event occurs, it causes the vehicle to abort the lane change and try again ASAP. This only started in the last month and occasionally occurs when you least expect it. A detailed message on the touch screen of what the vehicle discovered would help, as we could request a "bug report" when false events occur.

 

[AnxietyRanger]

Something new - for me - happened today.(Sorry, didn't make a note of the version, but it updated itself a few days ago.) I was once again briefly using AP2 since I was driving alone and wanted to test it.

Clear conditions, divided motorway, right-hand lane, left-lane empty, clear white lane markings, lead car ahead, slight curvature towards the left, nothing major, this is a motorway after all. Then the lead car moves over to the lane on the left. When the lead car changes lanes, AP2 actually starts to follow it to the other lane. Never mind these clear white markings on the road and the fact that I'm not touching the blinker stalk...

There was nobody around, so I let it do its thing. Halfway through following the lead car (somewhere on top of the lane markings), AP2 kind of fidgets and doesn't seem to think this is a good idea after all (really!?!) and turns back a bit, then back left a bit, and back right one more time, fidgeting smack in the middle of two lanes, before calling it quits and disabling itself with a bong.

That was interesting. I guess the algorithm really, really likes its lead car. The road marking were clear, the weather and light was good, there was no sun facing us, it was suitably overcast. AP2 just decided to abandon the gently curving lane markings and follow the sideways moving lead car instead.

While new in detail, this, generally speaking, is really typical of AP2. It just does not default to following the logical road markings and makes no "safe bet" assumptions about what's likely to come next, it seems. It seems it has no "smarts" about what the road likely might contain.

Someone noted that the older AP1 will happily follow road markings to a concrete divider if they lead there and judging by the infamous video that is true. AP1, while this is counter-intuitive in a way, is actually very logical and smart here. If the system is limited (as both AP1 and AP2 are) and does not know better, it is better to follow a lane to a divider, if the lane actually leads to a divider, than to make something up. Of course disabling the system and making an evasive maneuver while/before doing so would be even better, but what would not be good is deciding to "invent" virtual lanes (like a human can e.g. in a non-painted roadworks area), unless the system can really understand how to invent them...

Neither AP1 or AP2 can understand the road beyond road markings or sides, no that kind of NN magic is not there yet. AP1 assumes road markings more often than not make sense, so it defaults to following them in a way that makes sense, instead of picking up on some stray tar-line and twisting towards that like AP2 does at times... Unlike AP1, AP2 will happily not follow any such logic, but do something quite different instead - nobody just knows what and when. I'd rather take the tame AP1 logic at this stage of development.

It is better to have a predictable system that is limited, than an unpredictable system that is just as limited.

 

[rgreen]

AP1 uses a centered forward facing radar, in addition to the camera. As of f/w v8 it has become a shared primary system. Prior to v8 it was a secondary system.

Thank you very much for clarifying my misunderstanding. I was thinking of the system prior to v8 for sure.

 

[SlicedBr3ad]

Something new - for me - happened today.(Sorry, didn't make a note of the version, but it updated itself a few days ago.) I was once again briefly using AP2 since I was driving alone and wanted to test it.

Clear conditions, divided motorway, right-hand lane, left-lane empty, clear white lane markings, lead car ahead, slight curvature towards the left, nothing major, this is a motorway after all. Then the lead car moves over to the lane on the left. When the lead car changes lanes, AP2 actually starts to follow it to the other lane. Never mind these clear white markings on the road and the fact that I'm not touching the blinker stalk...

There was nobody around, so I let it do its thing. Halfway through following the lead car (somewhere on top of the lane markings), AP2 kind of fidgets and doesn't seem to think this is a good idea after all (really!?!) and turns back a bit, then back left a bit, and back right one more time, fidgeting smack in the middle of two lanes, before calling it quits and disabling itself with a bong.

That was interesting. I guess the algorithm really, really likes its lead car. The road marking were clear, the weather and light was good, there was no sun facing us, it was suitably overcast. AP2 just decided to abandon the gently curving lane markings and follow the sideways moving lead car instead.

While new in detail, this, generally speaking, is really typical of AP2. It just does not default to following the logical road markings and makes no "safe bet" assumptions about what's likely to come next, it seems. It seems it has no "smarts" about what the road likely might contain.

Someone noted that the older AP1 will happily follow road markings to a concrete divider if they lead there and judging by the infamous video that is true. AP1, while this is counter-intuitive in a way, is actually very logical and smart here. If the system is limited (as both AP1 and AP2 are) and does not know better, it is better to follow a lane to a divider, if the lane actually leads to a divider, than to make something up. Of course disabling the system and making an evasive maneuver while/before doing so would be even better, but what would not be good is deciding to "invent" virtual lanes (like a human can e.g. in a non-painted roadworks area), unless the system can really understand how to invent them...

Neither AP1 or AP2 can understand the road beyond road markings or sides, no that kind of NN magic is not there yet. AP1 assumes road markings more often than not make sense, so it defaults to following them in a way that makes sense, instead of picking up on some stray tar-line and twisting towards that like AP2 does at times... Unlike AP1, AP2 will happily not follow any such logic, but do something quite different instead - nobody just knows what and when. I'd rather take the tame AP1 logic at this stage of development.

It is better to have a predictable system that is limited, than an unpredictable system that is just as limited.

Thanks again for another vivid recollection of eap going on killing spree. I can definitively say that @AnxietyRanger is giving me anxiety eap. Well played

 

[S4WRXTTCS]

Thank you very much for clarifying my misunderstanding. I was thinking of the system prior to v8 for sure.

The AP1 TACC has always used the radar as the primary sensor.

This is absolutely critical to this entire thread because the events that happened as a result led to the hasty roll out of AP2.

Back before V8 the radar was tuned to only seeing moving objects (like every other radar based TACC system). As a result there were a few cases of people running into stopped cars (in either TACC or AP). It did use the camera for object recognition, and to verify what the radar was seeing. But, there was only so much the single camera could offer or the mobile eye neural net that it went to. For example it wasn't trained to see the side of a truck trailer.

With firmware V8 (or around that time) they brought some enhancements to it. One change was the ability to see two cars ahead, and this led it to being even more smooth.

They also tweaked the radar firmware to get a way more detailed point cloud from the radar. There was even talk about allowing the radar to be used to stop the car without confirmation from the camera.

Upgrading Autopilot: Seeing the World in Radar

This blog posts really confused things as it doesn't say anything about when the car will start acting on all this new data.

I'm led to believe that AP2 does because of all the reported false braking events due to overhead road signs. etc with that system.

One thing Subaru does WAY better than Tesla is they tell you exactly how the systems work. For example you know exactly how your AEB works. The manual spells it out in detail. You know exactly the speed differential the system can handle.

Does the manual in the Tesla spell it out? Barely. Even if they did it would change next week.

Instead people have to infer things based on what Elon says, or from an electrek.co article.

You had the car for 3500 miles, and had no idea how it worked.

I've had the car for over 25,000, and I'm not even sure about how involved the camera is. I know for absolutely sure it uses the radar as a primary source.

All I know is there is no WAY I would ever get a car without TACC. I think GM really screwed up in not offering TACC with the Bolt from the start.

 

[rgreen]

The AP1 TACC has always used the radar as the primary sensor.

This is absolutely critical to this entire thread because the events that happened as a result led to the hasty roll out of AP2.

Back before V8 the radar was tuned to only seeing moving objects (like every other radar based TACC system). As a result there were a few cases of people running into stopped cars (in either TACC or AP). It did use the camera for object recognition, and to verify what the radar was seeing. But, there was only so much the single camera could offer or the mobile eye neural net that it went to. For example it wasn't trained to see the side of a truck trailer.

With firmware V8 (or around that time) they brought some enhancements to it. One change was the ability to see two cars ahead, and this led it to being even more smooth.

They also tweaked the radar firmware to get a way more detailed point cloud from the radar. There was even talk about allowing the radar to be used to stop the car without confirmation from the camera.

Upgrading Autopilot: Seeing the World in Radar

This blog posts really confused things as it doesn't say anything about when the car will start acting on all this new data.

I'm led to believe that AP2 does because of all the reported false braking events due to overhead road signs. etc with that system.

One thing Subaru does WAY better than Tesla is they tell you exactly how the systems work. For example you know exactly how your AEB works. The manual spells it out in detail. You know exactly the speed differential the system can handle.

Does the manual in the Tesla spell it out? Barely. Even if they did it would change next week.

Instead people have to infer things based on what Elon says, or from an electrek.co article.

You had the car for 3500 miles, and had no idea how it worked.

I've had the car for over 25,000, and I'm not even sure about how involved the camera is. I know for absolutely sure it uses the radar as a primary source.

All I know is there is no WAY I would ever get a car without TACC. I think GM really screwed up in not offering TACC with the Bolt from the start.

Correct--I did not know that AP1 used radar in the way that you have described and there was nothing I saw in Tesla documentation that described it this way, and the sales team was of no help in this regard. Thank you for clarifying. I agree with your comment on TACC and for me as well I would not buy a car that was intended for multi-hour drives that did not have it. Colorado to Montana in a single day in my Tesla would have been very difficult without it; and 10 hours roundtrip time to view the total eclipse in Wyoming would have been impossible without the TACC in my Subaru. With my AP1 Tesla, I tried to subjectively measure the benefit to me of TACC with and without auto steer. I think I came down to the benefit being 70-80% from TACC and 20-30% from auto steer. I agree with the GM-Bolt error in not including TACC. For my use case, the Bolt is generally not driven more than 100 miles at a time so the lack of TACC is acceptable. The bigger issue is that there is no viable CCS DC charging network in the western states so this very much limits where I can use the car. The Bolt is really good otherwise, a great "driver's car", a good maneuvarable urban vehicle, and is generally under-appreciated. That said--I am looking forward to the day when I can go back to Tesla.

 

[xav-]

Doesn't this whole thread basically boils down to Radar vs Lidar?

Radar will detect any metallic object with their speed and a rough estimate of where they are heading. The main drawback though is that it cannot really tell the difference from a small metallic object to a large metallic object.

In the air that is not really a problem. If I detect a piece of metal moving at 600 mph at 30,000 feet, it's not like you need a mobileye camera to figure out what it is. It's not like there are going to be pedestrians at 30,000 feet, and there are no stationary metallic objects like speed signs.

That's why radar has been used since the 70s successfully as autopilot in the airline industry.

But how can we get something like this to work in a road?! Can we really believe that image recognition will address all the shortcomings?!

Lidar on the other hand would do a full 3D picture of the road and cars and would likely be much safer than any humans. We heard Musk criticizing Lidar in many occasions but we have many instances of self driving technology used successfully with Lidar.

That's my understanding on the topic please apologize and correct me if I am mistaken

 

[Electroman]

I think you are spot on..

 

[conman]

But how can we get something like this to work in a road?! Can we really believe that image recognition will address all the shortcomings?!

Well this is what the debate boils down to - whether you think software can be made smart enough to use all the camera input in addition to the radar to accurately identify everything with the same (or greater) level of reliability that a human does with not dissimilar amounts of information. The offloading of the workload almost entirely to software is the key here that sets AP2 apart from AP1 which had significant recognition work done in hardware. There is no doubt one day software and computing power will be smart enough to do all of this, but the question is whether it can be done now with neural networks and existing algorithms. It's clear from the opening post (what this thread is about) that it didn't remotely work in his case and others have also described similar scenarios with EAP and the current state of software. So the debate really is when will then be now? Hopefully soon.

 

[AnxietyRanger]

Doesn't this whole thread basically boils down to Radar vs Lidar?

Radar will detect any metallic object with their speed and a rough estimate of where they are heading. The main drawback though is that it cannot really tell the difference from a small metallic object to a large metallic object.

In the air that is not really a problem. If I detect a piece of metal moving at 600 mph at 30,000 feet, it's not like you need a mobileye camera to figure out what it is. It's not like there are going to be pedestrians at 30,000 feet, and there are no stationary metallic objects like speed signs.

That's why radar has been used since the 70s successfully as autopilot in the airline industry.

But how can we get something like this to work in a road?! Can we really believe that image recognition will address all the shortcomings?!

Lidar on the other hand would do a full 3D picture of the road and cars and would likely be much safer than any humans. We heard Musk criticizing Lidar in many occasions but we have many instances of self driving technology used successfully with Lidar.

That's my understanding on the topic please apologize and correct me if I am mistaken

Nobody aims for just lidar AFAIK. Radar is there too.

I think what is deep down boils to, actually, is this:

Is vision sufficient for self-driving? Is vision-only good enough or better than other methods?

Tesla is basically at the moment aiming for a vision only system, complemented by narrow front radar for highway driving and slow ultrasonics for low-speed maneuveuring. But basically it is a vision system, given the inherent limitations of the latter two.

Meanwhile, the competition seems to be aiming for a 360 degree vision, lidar and radar combination as all three technologies have certain benefits and only by a 360 view for all can all scenarios be covered.

I am sure there will be other setups between these extremes, but IMO the argument Elon and Tesla are making is that vision is enough, when the rest of the major players believe otherwise.

 

[Tangible]

I received a software update a few days ago. The release notes promised resolution of some unspecified bugs, so I tried a few things.

Summon: I have a two car garage with separate doors, so there's a center divider between them. There's plenty of clearance on both sides. I told the system to back out into the driveway. With the car halfway out of the garage it inexplicably turned the wheels and aimed the front fender at the center divider. I managed to abort the process about two inches from a collision.

AP2: out on the highway I had several instances of hard braking for no apparent reason. Then, driving in the left lane at about 65 mph, it took a sudden lurch toward the metal barrier. Because I had my hands on the wheel I managed to avoid the collision, but overcoming the steering wheel resistance caused a rapid overcorrection to the right and then a correction back to the left.

My frightened passenger, who long ago promised to love and honor - but not to obey - me, begged me to never again turn on the AP system with or without her in the car. I'm a compulsive early adopter of new technology. She tends to embrace what works and improves our lives, and reject gadgetry that just adds complexity. So when she says NFW, I listen.

The novlety of AP was exciting. Now I'm at the point where all it does is cause me extra stress and anxiety while Decreasing safety. I'm a safe and confident driver. I would love the convenience of a competent Adaptive Cruise Control, like other cars have. AP2's other features should be recalled, voluntarily or not, until it can be made safe.

Note to those who attack anyone expressing concerns about their Teslas, and are about to reply with criticism of my reading comprehension or general intelligence, or to demand statistical evidence, or tell me to just stop using it if I can't handle it: My concern isn't just for my own car. I think Teslas operating under AP2 are a hazard to other vehicles as well as their own, even with a vigilant driver. The sudden braking for no reason could easily cause a rear-end collision. Attempts to correct sudden lurching into barriers or other lanes can affect other cars. In the face of Tesla's stonewalling and continued failure it's time to blow the whistle.

 

[Mark Z]

One more to report. At night the 91 westbound freeway through Corona has been flawless with excellent reflectors. Not on September 8th. Twice the MS 90D AP2 changed lanes without a turn indicator. After the first partial one where the vehicle returned to the original lane, I called Tesla and got through. While on the phone with Tesla, my vehicle started to weave a bit, crossed over the left lane line, drove on both lanes for a second and then continued to center itself in the lane to the left of where I had been driving. I was able to explain this to Tesla in real time! No turn indicator was touched. This occurred on pavement with clear lane markers and no other lines on the road. I suggested that they download the logs immediately to capture what occurred. Thankfully, no passengers were aboard who may have been upset over the anomaly.

 

[AnxietyRanger]

One more to report. At night the 91 westbound freeway through Corona has been flawless with excellent reflectors. Not on September 8th. Twice the MS 90D AP2 changed lanes without a turn indicator. After the first partial one where the vehicle returned to the original lane, I called Tesla and got through. While on the phone with Tesla, my vehicle started to weave a bit, crossed over the left lane line, drove on both lanes for a second and then continued to center itself in the lane to the left of where I had been driving. I was able to explain this to Tesla in real time! No turn indicator was touched. This occurred on pavement with clear lane markers and no other lines on the road. I suggested that they download the logs immediately to capture what occurred. Thankfully, no passengers were aboard who may have been upset over the anomaly.

This sounds similar to my experience in #189. Was there a lead car in either of your cases?