Poor Model Y Tesla Vison Experience

[Lordswine]

Hello Everyone and thank you for all of your responses ahead of time.

I recently bought a 2023 Model Y (owned it only two weeks at the time of this post and never owned a Tesla before) with "Tesla Vision" and I am very disappointed in the Autopilot (not even talking about FSD). For example today the car today was so close to rear-ending the car in front of me that my new seats have pucker marks! it lunges and breaks and stops before signs and lights. It is stressful to even drive the car with Autopilot. I bought the car to remove some of the stress of driving and now I feel cheated. I am messing around with FSD but I expect it to be buggy because it is in Beata and I knew that going in. But for the car to just drive this crappy with Autopilot is very disappointing, to say the least.

My point here is not to rant but to ask around if anyone else suffering from this Tesla Vison crappiness. I honestly would bring the car back and wait for it to work better at a later date if I could. Is there any hope or news or was Tesla Vison one BIG FAT mistake from Tesla?

Seriously Disappointed Tesla customer.

 

[BitJam]

Good observations and I too am learning as I get to using more of the TACC an AP. The one region where TACC/AP definitely loses references is when there are huge ups and downs and you are going up the mound. The vision cannot see beyond the top of the hill/mound and it starts slowing down/braking.

I agree. Although this has gotten significantly better for me. There's a small hill on the two lane road to my house that I mentioned before. My car used to slow down there like clockwork. But in the past year or so it has only slowed down there once or twice. A funny coincidence: my old, under-powered Volvo 240 wagon would slow down at the same place when on cruise control because CC limited how hard it would press the gas pedal. I was astounded the first time the Tesla slowed in the same place.

But I wouldn't call this phantom braking. More like being overly cautious which it also is with cars crossing my path or cars slowing down to turn ahead of me. There are levels to this:

1. Slowing down due to something I can see
2. Slowing down for no apparent reason
3. Hard braking due to something I can see
4. Hard braking for no apparent reason

I would only classify (4) as phantom braking. It's been a long time since I've experienced either (3) or (4). The worst case was over a year ago in the summer when I was going down a straight, mostly flat 70 mph two lane highway (one lane in each direction). It would hard brake frequently due to mirages in the road ahead. I solved this by pushing the accelerator gently when I saw a mirage ahead. It was annoying at first but no big deal.

The second worst problem I can remember was braking when it saw an oncoming truck in the opposite lane on a two-lane road or highway. This was a hot topic here at TMC when it was happening. Again, this was easily solved by putting my foot on the accelerator when I saw I truck approaching in the opposite lane. Sure, this was surprising and annoying at first but soon it became second nature to compensate. The problem went away six months or a year ago.

 

[PVia]

The forward camera has a range of 820 feet. It sees the stopped cars at that range and takes action. 75mph is 110 feet per second. That means it has to stop in about 7 seconds.

Do people think it's the software or a bug causing this behavior? It's performing exactly how it's designed to. It's simple math.

I don't think it's a bug...
I just know that when I see red lights piling up in front of me at 65-70mph, I start letting go of the throttle and use my judgement. AP doesn't seem to do that and I get a bit antsy. It's never NOT done what it's supposed to do, but for peace of mind I'd like a little more forgiveness in that respect. I didn't know about the 820 ft range; sometimes it seems/feels less than that.

 

[enemji]

I agree. Although this has gotten significantly better for me. There's a small hill on the two lane road to my house that I mentioned before. My car used to slow down there like clockwork. But in the past year or so it has only slowed down there once or twice. A funny coincidence: my old, under-powered Volvo 240 wagon would slow down at the same place when on cruise control because CC limited how hard it would press the gas pedal. I was astounded the first time the Tesla slowed in the same place.

But I wouldn't call this phantom braking. More like being overly cautious which it also is with cars crossing my path or cars slowing down to turn ahead of me. There are levels to this:

1. Slowing down due to something I can see
2. Slowing down for no apparent reason
3. Hard braking due to something I can see
4. Hard braking for no apparent reason

I would only classify (4) as phantom braking. It's been a long time since I've experienced either (3) or (4). The worst case was over a year ago in the summer when I was going down a straight, mostly flat 70 mph two lane highway (one lane in each direction). It would hard brake frequently due to mirages in the road ahead. I solved this by pushing the accelerator gently when I saw a mirage ahead. It was annoying at first but no big deal.

The second worst problem I can remember was braking when it saw an oncoming truck in the opposite lane on a two-lane road or highway. This was a hot topic here at TMC when it was happening. Again, this was easily solved by putting my foot on the accelerator when I saw I truck approaching in the opposite lane. Sure, this was surprising and annoying at first but soon it became second nature to compensate. The problem went away six months or a year ago.

Yes, I totally agree with you.

I often wonder if the 2 & 4 are a result of something we are not capturing or maybe are but dismissing it, whereas AP is still not sure which one of the many actions it should take, and therefore taking the caution of slowing down and/or braking.

In other words, not enough information to make a firm decision.

 

[DrChaos]

Yes, this matches my experience. PB mostly (perhaps always) happens when there is no one else near me, or at least not in front of me. I believe your assessment is spot on.

PB often occurs in empty roads when there is a hill/undulation and the vision system cannot extrapolate forward the lane lines in a sensible way. It's a consequence of trying to be too smart and too safe with limited information.

If it sees a car that reassures it that there is driveable area ahead.

 

[COS Blue]

It's a consequence of trying to be too smart and too safe with limited information.

That, in a nutshell, is my general overall complaint with TACC. It tries to do so much, that it cannot consistently get the basics correct. Most of the time, yes. To me, that's not good enough.

 

[enemji]

That, in a nutshell, is my general overall complaint with TACC. It tries to do so much, that it cannot consistently get the basics correct. Most of the time, yes. To me, that's not good enough.

Imagine if the creator said monkeys are good enough. They are operating within 1 year whereas The new “human” model takes 18 years to get itself going

 

[KArnold]

I don't think it's a bug...
I just know that when I see red lights piling up in front of me at 65-70mph, I start letting go of the throttle and use my judgement.

Not a bug? Then it's a feature by design? Maybe with a toggle switch for PB on or off?

I agree with yout description of behavior and our efforts to mitigate in its current iteration. But IMHO it's definitely one of many many bugs that need to be fixed some day.

Sorry if I missed your point.

 

[enemji]

Not a bug? Then it's a feature by design? Maybe with a toggle switch for PB on or off?

I agree with yout description of behavior and our efforts to mitigate in its current iteration. But IMHO it's definitely one of many many bugs that need to be fixed some day.

Sorry if I missed your point.

Well, the car is not going to crash. It is just making you anxious for sure

 

[sjg98]

That, in a nutshell, is my general overall complaint with TACC. It tries to do so much, that it cannot consistently get the basics correct. Most of the time, yes. To me, that's not good enough.

Agreed, I think FSD/AP remains in an "uncanny valley" of trying to do too much while not doing the basics great.

Imagine you bought an automated coffee machine - the bean bin dispenses to the grinder which grinds and dispenses to the brewer, which boils water and dispenses coffee. When done it dumps the grinds into a refuse bin. It is plumbed in so you don't even have a water basin to fill.
However it requires you to watch it at all times.

While it works perfectly 90% of the time, 10% it does not.

Sometimes It tries to brew without any water and you can smell the heating element overheating.
Other times it fails to close the water valve and water spurts out all over the counter.
Sometimes it only fills half your coffee cup.
Sometimes it over dispenses and the coffee overflows all over your counter.
Sometimes the grinder keeps running forever and you can smell the motor getting hot.
Other times it dispenses beans but fails to grind.
Other times it grinds the coffee and fails to brew.

But don't worry there's buttons for each step and as long as you are watching each step & make sure they work as expected, and override where they don't, it really does make great coffee!

We don't accept this from a $200 coffee machine, but its OK on a $70k 2-ton vehicle moving at highway speed!

 

[KLam]

I think all these reported problems (FB, not slowing down early enough, etc) with AP, NOA, FSD, FSD beta are related to the driving speed and road conditions.

To use the systems in a more relaxed way, we have to know the limitation of the system and try to help it.

On a straight road, we have to understand that an approaching car at a far enough distance will look like a single pixel (if the distance view is on a quiet field) on the forward camera view. This single pixel will bounce around due to road conditions. To detect that its approaching, that single pixel have to become 2 pixels! So the faster we drive, the error of detection becomes bigger. Hence, more phantom braking (false automatic emergency braking), slow in slowing down etc. There are a lot of complaint about these false positives. But, would we want to take away the sensitivity or the function of automatic emergency braking which most older cars do not have? Sure we will have old dumb cruise control and not know about if a real collision happens.

For me, I seldom drive above 73 mph on Freeways like I-5, I-80, 405 etc. On Route 152 which I drive often, I seldom exceed 65 mph and not using NOA due to lots of curves. I stay away of right lane when near a merge/junction. I disable auto lane change and initiate the lane change manually (The problem here is that on the I-5, the system often asked to change lane too late when coming upon a slow truck, where there is a long approaching cars on the pass lane. And the car do not speed up to match pass lane speed) . I take exchanges manually (I just don't want to risk the uncertainty of the system. For example the North bound I-5 exit to Route 152 cuts across a merge into I-5 just in front of the exit). I stay on the left lane of I-5 sometimes for sections where there are a lot of pot-holes on the right lane due to heavy truck traffic. I dial down speed when I see traffic ahead to help the system for a more smooth slow down.

I wish Tesla approach the FSDbeta (City driving capability portion) more conservatively. Manual initiate lane change, stop at signals/stops and ask for confirmations to proceed. Confirms to proceed L/R turns. Confirms to go around obstacles. Ways of step slow downs for speed bumps, etc. In another word, a system to relieve stress on City driving instead of introducing more stress (the uncertainty of whether the system will perform correctly introduce more stress).

 

[Dewg]

The difference is that the coffee maker you describe is hard coded to perform those actions. Most software is hard coded to perform an action. It will do the same thing over and over in the same, predictable manner.

Telsa is developing AI for their cars. They use neural nets to make decisions. The car is "thinking" about what it needs to do based on input it receives from its sensors. It's similar to how humans are "programmed".

If I swing my fist at your face, your sensors detect the fist and the motion towards you. Your brain processes that input and runs it through the various nets in your head (your brain), and a decision is made. The decision is then put into action which controls your body to react to that stimuli. More than likely you'll duck the punch.

Now, let me ask you this question. If I swing at your face again, will you react the exact same way each time? The answer is no, you won't react the same way each time. You might duck this time, you might back up, you might take the punch on the face, you might raise an arm to block it. Even if you decide to duck each time, you won't duck the exact same way each time. You might miss the punch by an inch this time, and several inches the next time.

This is one of the reasons people are surprised that the car takes a right turn slightly differently each time. They think it's a problem - "Why doesn't the car take the right turn perfectly each time?" It's because the car is taking in sensor input and thinking about it, adjusting and making decisions based on the input each time, just like the punch coming at you. It won't necessarily react the same way each time. Just like we don't react the same way each time. We don't take the right turn the exact same way each time when we do it - sometimes we overshoot the lane a bit, sometimes we take it faster, sometimes slower, etc.

 

[sjg98]

The difference is that the coffee maker you describe is hard coded to perform those actions. Most software is hard coded to perform an action. It will do the same thing over and over in the same, predictable manner.

Telsa is developing AI for their cars. They use neural nets to make decisions. The car is "thinking" about what it needs to do based on input it receives from its sensors. It's similar to how humans are "programmed".

If I swing my fist at your face, your sensors detect the fist and the motion towards you. Your brain processes that input and runs it through the various nets in your head (your brain), and a decision is made. The decision is then put into action which controls your body to react to that stimuli. More than likely you'll duck the punch.

Now, let me ask you this question. If I swing at your face again, will you react the exact same way each time? The answer is no, you won't react the same way each time. You might duck this time, you might back up, you might take the punch on the face, you might raise an arm to block it. Even if you decide to duck each time, you won't duck the exact same way each time. You might miss the punch by an inch this time, and several inches the next time.

This is one of the reasons people are surprised that the car takes a right turn slightly differently each time. They think it's a problem - "Why doesn't the car take the right turn perfectly each time?" It's because the car is taking in sensor input and thinking about it, adjusting and making decisions based on the input each time, just like the punch coming at you. It won't necessarily react the same way each time. Just like we don't react the same way each time. We don't take the right turn the exact same way each time when we do it - sometimes we overshoot the lane a bit, sometimes we take it faster, sometimes slower, etc.

Yes, AI/ML models are non-deterministic.
However, for $15k, it ought to be good.
Which it is.. not.

If it drives good, in slightly different ways, fine.

If it drives great, good, ok, awful, often in the exact same road in same weather conditions, it is not confidence inspiring.

Consumers care about outcomes not process.
That is to say, provided it tastes good & is safe to eat - I don't care how the sausage gets made.

Tesla having chosen an approach that no one else has, that may not be working is not my problem. That is their problem.

 

[Dewg]

Yes, AI/ML models are non-deterministic.
However, for $15k, it ought to be good.
Which it is.. not.

If it drives good, in slightly different ways, fine.

If it drives great, good, ok, awful, often in the exact same road in same weather conditions, it is not confidence inspiring.

Consumers care about outcomes not process.
That is to say, provided it tastes good & is safe to eat - I don't care how the sausage gets made.

Tesla having chosen an approach that no one else has, that may not be working is not my problem. That is their problem.

Totally valid viewpoint. You want the result, and I don't mind the journey to get there.

 

[enemji]

I think all these reported problems (FB, not slowing down early enough, etc) with AP, NOA, FSD, FSD beta are related to the driving speed and road conditions.

To use the systems in a more relaxed way, we have to know the limitation of the system and try to help it.

On a straight road, we have to understand that an approaching car at a far enough distance will look like a single pixel (if the distance view is on a quiet field) on the forward camera view. This single pixel will bounce around due to road conditions. To detect that its approaching, that single pixel have to become 2 pixels! So the faster we drive, the error of detection becomes bigger. Hence, more phantom braking (false automatic emergency braking), slow in slowing down etc. There are a lot of complaint about these false positives. But, would we want to take away the sensitivity or the function of automatic emergency braking which most older cars do not have? Sure we will have old dumb cruise control and not know about if a real collision happens.

For me, I seldom drive above 73 mph on Freeways like I-5, I-80, 405 etc. On Route 152 which I drive often, I seldom exceed 65 mph and not using NOA due to lots of curves. I stay away of right lane when near a merge/junction. I disable auto lane change and initiate the lane change manually (The problem here is that on the I-5, the system often asked to change lane too late when coming upon a slow truck, where there is a long approaching cars on the pass lane. And the car do not speed up to match pass lane speed) . I take exchanges manually (I just don't want to risk the uncertainty of the system. For example the North bound I-5 exit to Route 152 cuts across a merge into I-5 just in front of the exit). I stay on the left lane of I-5 sometimes for sections where there are a lot of pot-holes on the right lane due to heavy truck traffic. I dial down speed when I see traffic ahead to help the system for a more smooth slow down.

I wish Tesla approach the FSDbeta (City driving capability portion) more conservatively. Manual initiate lane change, stop at signals/stops and ask for confirmations to proceed. Confirms to proceed L/R turns. Confirms to go around obstacles. Ways of step slow downs for speed bumps, etc. In another word, a system to relieve stress on City driving instead of introducing more stress (the uncertainty of whether the system will perform correctly introduce more stress).

The ONE big change I have made to my driving style is to no longer drive in the left most lane. I am usually in the middle lane. This has helped immensely as it always has lane markers as well as vehicles surrounding to provide itself good references and it works very well.

 

[PVia]

Not a bug? Then it's a feature by design? Maybe with a toggle switch for PB on or off?

Sorry, wasn't talking about PB...

 

[BitJam]

FWIW, I took a 1.5 hour round trip to visit a friend in Cliff, NM. Most of the drive was on a two lane highway (one lane in each direction). Very little traffic going in my direction and no problems with AP.

The paved road my friend lives on eventually peters out into a two-lane dirt road. AP was engaged on the paved road and stayed engaged when it became dirt. Lane keeping was flawless. Here is a picture of the road:

Even without AP enabled the lane showed up in the visualization just fine. I moved to the center of the road to take the picture but the lane the car sees is on the right.

I realize AP stresses some people out and given their stories here, I'm not surprised. For me it's the opposite. I find driving a car without AP enabled much more stressful and tiring. I had AP on for the entire trip except for a short residential section within 400 feet of my house and parts of the dirt road. I also disabled it four times each way to turn onto different roads.

BTW: there was no cell coverage for most of the trip. Spotify did a good job of caching songs for most of the way out. The nearest superchargers are 280 miles apart and AFAIK there are no public fast chargers. The nearest Service Centers are even farther away in Arizona and Texas. It's not an ideal area to visit with a Tesla. I've only seen three other Teslas here and two of them were from out of state.

 

[COS Blue]

There are a lot of complaint about these false positives. But, would we want to take away the sensitivity or the function of automatic emergency braking which most older cars do not have?

Yes.

 

[2101Guy]

Yes.

To make excuses for Elon

 

[electracity]

Does radar still work on cars with the device?

 

[COS Blue]

Yes, this matches my experience. PB mostly (perhaps always) happens when there is no one else near me, or at least not in front of me. I believe your assessment is spot on.

Well, I can no longer say that I don't experience PB when there is a car in front of me. Had my worst PB event yet yesterday afternoon, driving home from a ski trip. I was in a line of cars on CO Hwy 9. My daughter was having a pleasant nap and I had TACC engaged. Bam! Hard braking while simultaneously hearing the beep-beep-beep-beep-beep that I associate with AEB. Usually when PB happens I instantly realize that it's just the car seeing a ghost but this time I was genuinely confused. I had been watching the road but the beeping & braking made me quickly scan to see what it was I missed, ready to take evasive action. Of course, there was nothing. Meanwhile there was a car behind me in the line of cars who I was in the process of inadvertently brake-testing so I put my foot on the accelerator, but then realized I was now closing on the cars in front of me before I pushed the stalk forward to disable TACC. It all happened very quickly and it was probably my most frightening experience in the car (so far).

My daughter, who already hates the car due to other instances like this, was quite upset (she was very rudely awoken from her nap and said that it terrified her), and she wondered why anybody would buy a Tesla. I pointed out that based on TMC, I begin to feel a little crazy because there are so many people who say how wonderful AP is (not to mention FSD), and it makes me feel sometimes like I'm the only one who has these issues (aside from you fine posters on this thread and a couple others). She replied that she had told kids that she skis with about PB and one of them said "Ours does that too!"

What makes it even worse was that I had actually thought PB was getting better. The last couple of ski trips we didn't experience anything really bad, and then BOOM. Since I was in a line of cars and still had someone behind me, I drove the rest of the remaining 75 miles mostly without the benefit of cruise control. I am slowly but surely getting to the point where I believe the number of usable driver assistance tools that my $70k offers me is: zero.