Think we'll ever get follow distance of "1" again?

[ScoobyDoo82]

Can I just say that using car lengths as the metric is stupid? Following distance should be based on time not distance.


It should be a measure of your reaction time, and ability to stop before plowing into what is in front of you. See the recent crash in the tunnel, where people behind were choosing a poor follow distance.

If you use time then you don't need to adjust it for different speeds, and you would automatically be able to follow closely at low speeds, because 200ms of time is not very much pavement at 10 mph.


For someone driving aggressively and video game caliber reflexes, 200ms would be OK. If you are driving casual, let's bump that up and give you 500ms of reaction time. If you are on TACC, let's bump it up to 1000ms, because you want some leeway.

Am I asking too much to have sane measurement units?

The AP follow numbers are not actually measured in car lengths. They are just arbitrary numbers similar to a stereo volume knob. The number doesn’t reflect any measure of length or time. It’s simply a setting to indicate more (or less) space from the number before (or after) whatever you have AP set at. Setting your AP to “2” will keep you actually a little bit more distanced from the vehicle ahead of you than only 2 “car lengths”.
I mean what actually is a “car length”? 8ft, 10.8ft, 12ft? It’s subjective.
I’ve had my AP set at “2” for as long as I’ve had Tesla’s.. the following distance has always varied.

 

[AlanSubie4Life]

Except I didn't attempt to define either. Tesla refers to the setting as "follow distance."

My point was just that AP is broken. This is unrelated to follow distance. You could have a setting of zero and you’d have the same problem you describe (which is clearly a problem and has been described by many).

And yes, it is follow distance. Because it sets the follow *distance*, which is dependent on speed, of course. It’s also sometimes extremely long, not controlled by this setting, but by the broken slow-settling control loop.

I doubt it has much to do with radar vs. vision; it is just broken.

Humans may be smooth when stopping time and distance are well known, but in a panic stop, they... panic and perform poorly. Or in the worst cases they don't realize the car in front of them has panic stopped and they don't brake appropriately and run into the leading car.

That is why I specified an attentive (100%) human. Which the car can do. Unfortunately the car lack other aspects of human anticipation, which *far* exceed FSD capability.

 

[BitJam]

They broke it because they are off in the weeds desperately trying to solve the jerk problem (and failing of course).

There is a separate thread for Elon & Twitter.

 

[TarmacSurfer]

My default is “5”. I hate tailgaters and don’t care if someone else really wants to slide into the gap. Most of the time they don’t anyway. Much, much smoother than with lower following distances.

 

[AlanSubie4Life]

My default is “5”. I hate tailgaters and don’t care if someone else really wants to slide into the gap. Most of the time they don’t anyway. Much, much smoother than with lower following distances.

I use 7 and can report that it is not smooth at all in any situation which presents any sort of challenge.

 

[DayTrippin]

I was just reading this thread and thinking about the difference between my 3 and S. On my S (with radar) I purposely didn't upgrade the software when the removed radar from the stack.

It is SO much better than my 3 which came without radar, even running the latest version. As luck would have it, my S automatically (without any intervention from me) started downloading the latest version of SW. It is at 50% downloaded now, is there any way of stopping and canceling it or is too late? I can disconnect it from WiFi if that will buy me any time.

 

[gsmith123]

I'd much rather have >85mph AP than follow distance 1.

 

[DayTrippin]

That is very handy too, especially in Texas. It is even more important since the speedometer is off a bit. So on one of the highways with an 85 mph speed limit, if I were limited to 85, it would actually be about 82-83 mph and I'd be blocking a lot of traffic. Setting it at 90 mph tends to work pretty well.

 

[nvx1977]

Given that the "1" setting was removed at the time Tesla disabled radar, it should be obvious that vision cannot identify distance as well as radar. And yes you can call it a latency problem because the triangulation calculations needed via vision takes longer to get an answer of confidence than a single ping/pong from a radar.

 

[AlanSubie4Life]

Given that the "1" setting was removed at the time Tesla disabled radar, it should be obvious that vision cannot identify distance as well as radar. And yes you can call it a latency problem because the triangulation calculations needed via vision takes longer to get an answer of confidence than a single ping/pong from a radar.

The thing is, the issue being discussed here is the inability of the current system to get to the right following distance in a timely manner. Meaning it takes 20-30 seconds or something. I have not measured it, but everyone knows the phenomenon - which is not specifically related to radar or vision of course. It’s very frustrating particularly when traveling in the fast lane when you are expected to keep up after making a move that puts you in a position of needing to close following distance (at a reasonable distance).

 

[nvx1977]

The thing is, the issue being discussed here is the inability of the current system to get to the right following distance in a timely manner. Meaning it takes 20-30 seconds or something. I have not measured it, but everyone knows the phenomenon - which is not specifically related to radar or vision of course. It’s very frustrating particularly when traveling in the fast lane when you are expected to keep up after making a move that puts you in a position of needing to close following distance (at a reasonable distance).

i feel like this slow-to-get-back-to-speed issue also started after disabling radar. I tend to agree that this problem can be eliminated via vision-only, so the radar isn't essential. Not sure why the bug still persists. My best guess is that it won't be there for single stack, and there's not a lot of emphasis on the old code stack anymore.

 

[gsmith123]

My best guess is that it won't be there for single stack, and there's not a lot of emphasis on the old code stack anymore.

FSD Beta is noticeable better at keeping up in traffic so I'd agree with this. AP isn't unsafe it's just.. annoying and you have to hit the accelerator.

It's pretty clearly a decision making issue, not a perception issue.

 

[beatle]

It seems silly that Tesla cannot crib from their own codebase and have the following behavior mimic that of the old cars. It's one thing to have to work within the confines of the system with only vision to measure distance, but it's another thing to execute driving inputs based on that distance. I am generalizing here, but as the following behavior was "solved" years ago, they just need to rip out the existing driving code and replace it with the code used on the radar system or even AP1. Vision only exists to provide the data necessary for the car to execute commands; it doesn't actually drive the car.

This is assuming the data on distance coming from the cameras isn't wildly, incredibly poor. We like to bash it, but it's been around for years and I believe it should be good enough.

 

[**BatteryLow**]

It seems silly that Tesla cannot crib from their own codebase and have the following behavior mimic that of the old cars. It's one thing to have to work within the confines of the system with only vision to measure distance, but it's another thing to execute driving inputs based on that distance. I am generalizing here, but as the following behavior was "solved" years ago, they just need to rip out the existing driving code and replace it with the code used on the radar system or even AP1. Vision only exists to provide the data necessary for the car to execute commands; it doesn't actually drive the car.

This is assuming the data on distance coming from the cameras isn't wildly, incredibly poor. We like to bash it, but it's been around for years and I believe it should be good enough.

I think you leave out Tesla is leaving out radar to produce “premium” cars at lower production costs and that also includes the cost reduction on OTA software updates to no longer include the support for radar.

 

[Supcom]

Given that the "1" setting was removed at the time Tesla disabled radar, it should be obvious that vision cannot identify distance as well as radar. And yes you can call it a latency problem because the triangulation calculations needed via vision takes longer to get an answer of confidence than a single ping/pong from a radar.

The latency issue is not the time it takes to estimate distance. Tesla does not 'triangulate' range. That would be useless for objects in front of the car where the angle or arrival is constant.

The latency issue is the simple fact that the car is a half second behind the universe, so there is an ambiguity zone of position, velocity and acceleration of all dynamic objects seen by the car. The planner needs to take into account that the lead car may have slammed on it's brakes a half second ago. At 80 mph, this means that the Tesla could be 58 ft closer to the lead car than the Tesla knows about and could be rapidly closing the gap. So, the Tesla needs to allow and additional half second of time to stop the car to guarantee it does not rear-end the lead car.

 

[beatle]

I think you leave out Tesla is leaving out radar to produce “premium” cars at lower production costs and that also includes the cost reduction on OTA software updates to no longer include the support for radar.

Yes, I know their reasoning. My question is why the AP programming that determines the driving inputs necessary for good following performance can't just be reused. How you calculate the distance of the car in front of you should be meaningless to the driving function.

There are two parts of the AP code:

1. Figure out what the world looks like.
2. Manipulate steering and acceleration/deceleration.

You need #1 in order for #2 to have any context, but if they've already solved #2 and are just doing #1 via a different method, that should be okay and should still produce a car that drives the same.

The latency issue is the simple fact that the car is a half second behind the universe

But that's exactly what latency is. I don't know how the car could have 500ms of latency with as much processing power as the AP computer has. A human has an average reaction time of 250ms to visual cues.

 

[Supcom]

But that's exactly what latency is. I don't know how the car could have 500ms of latency with as much processing power as the AP computer has. A human has an average reaction time of 250ms to visual cues.

How can you assess the latency of the AP computer when you are not familiar with the code that it executes? Do you have some detailed insight into Tesla's software?

I can see the latency at work simply be watching how long it takes for the car to react to various situations. For example, when a lead car slows and turns into a parking lot, my car continues to apply brakes for at least a half second after the lead car has completely departed the lane.

BTW, human reaction time of 0.25 sec is for unexpected, or sudden changes that are not predictable. When a human is performing an action where the situation is predictable, reaction times are much better. You can steer a car smoothly around corners because you continually assess the position of the car in response to your steering and can correct without a long reaction time. Same goes for the example I gave above where a car turns out of your lane. You anticipate the other car's turn so can stop braking and start speeding up the instant the lead car is clear of your lane.

Even so, 0.25 sec latency is far better than 0.5 sec in a driving situation, right?

 

[beatle]

What makes you think the computer's latency is 500ms? The processors run at 1.6Ghz and are capable of 72 tflops. This isn't some vacuum tubed dinosaur. The cameras aren't great which does affects distance vision, but they'd have to be downright russet potatoes as to not provide enough data to allow for the cameras to determine distance in order to maintain human like following behavior. Remember, Tesla originally billed HW3 as being capable of "full self driving." I know most all laugh at that now, but lane centering and traffic aware cruise control are some of the most basic driver aids, and their performance was the point of this thread.

Also, if you're talking about a human predicting a move, then you're not talking about reaction time. That means you've decided to take action in advance of an event occurring so your "reaction time" would be negative, but that's not what we're talking about here (though the car suddenly slamming on the brakes as a slower leading car approaches may be indicative of terrible prediction by the computer).

 

[bo3bdar]

I was just reading this thread and thinking about the difference between my 3 and S. On my S (with radar) I purposely didn't upgrade the software when the removed radar from the stack.

It is SO much better than my 3 which came without radar, even running the latest version. As luck would have it, my S automatically (without any intervention from me) started downloading the latest version of SW. It is at 50% downloaded now, is there any way of stopping and canceling it or is too late? I can disconnect it from WiFi if that will buy me any time.

If you disconnect from WiFi it will stop trying to download new versions, nor finish the one in progress. It will notify there are new updates available, but not download. However, there are scenarios where they will force download a version anyway, over LTE. The one I saw force fed to me was to fix a recall notice.

It did not force update, still needed an OK, but then would nag on every drive about being available, and one click away from giving me an update.

One thing I found is that if you factory reset the car, then any pending download is dropped. You don't get old firmware back, but you can clear a pending install this way. (Model 3 here, can't be positive S behaves same way.)

 

[DarkForest]

How can you assess the latency of the AP computer when you are not familiar with the code that it executes?

What makes you think the computer's latency is 500ms?

Do you even Beta, bro?

(sorry, I had to haha)

I use the Beta for all of my driving, and the approximately half-second latency is painfully obvious in numerous scenarios.

It may not be across the board and may only affect certain NNs, but it’s definitely there. It’s most obvious when there are vehicles up ahead driving perpendicular to you or otherwise could be in your way. A lot of times I’ll know the car will be out of the way by the time we cross their path before the Beta realizes this leading to unnecessary braking. I’ve also seen it happen with traffic signals in rare cases (eg like in Chuck’s traffic light phantom braking video).