Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Autopilot leading-car tracking significantly worse after the radar-removal push

This site may earn commission on affiliate links.
My car's a 2018 3 LR/RWD with HW3 and FSD. I use Autopilot & NoAP constantly, on any road I think it'll handle, kinda like bionic driving. I watch everything that's happening, and let the car handle the fine controls. And in that, I watch it very closely... I notice when things change. And things have... definitely changed.

With recent updates coinciding with the push for radar removal (it's been going on for a while - not THAT recent), I've noticed that it's gotten really, really sloppy about detecting and reacting to the car in front of it. I had hoped this would improve, but some updates later, it's still not getting better. Thus, I'm posting about it.
It's my understanding that AP still uses radar on cars that have the hardware, and only uses pure cameras on the cars that do not (Model Y mostly I think). You can tell that by checking to see if you can set follow distance to 1 (camera only wont go below 2). Of course, you could still be right about problems in the newest software releases, but in that case it wont be directly related to the lack of radar.
 
  • Informative
  • Like
Reactions: impastu and Ogre
I'm legitimately surprised that's even... something they bothered changing in the UI. The following distance is so unbelievably lax, and always had been, it has no bearing on a physical measurement - it's just a relative setting whose distance changes based on the speed and mood, with a large scoop of slop on top. So they could have simply flopped the measurements around a bit again, as they've done many times in the past with various updates, and nobody would even have a way to tell a vision system vs. a radar system. For quite some time, the distance setting has only had a suggestive effect on following distance. It was previously about as rigid as a wet piece of spaghetti, but now even that is generous.

Basically what I'm saying is "why'd they even bother making a UI change for that?" 😂

The more I observe it, the more I think it's not so much that its perception (vision vs. radar) is the problem, but the policy (how to act on that perception) is the problem. The visualization shows the cars at reasonably correct relative distances, but AP just chooses to act distracted/drunk in reaction to them. It's more a choice in programming than an effect of sensors...
 
The "following distance" setting is actually based on time. It stays the same amount of seconds behind the lead car, regardless of speed. This has the effect of have a larger distance behind the lead car when moving fast, and a shorter distance behind the lead car when moving slowly.
That's what they say in the manual, but that's not the actual reality of how it operates. It's hardly even worth replying to this kind of boiled-dry manual-recital. It doesn't match reality in any way. It increases the following distance in at least a handful of undocumented circumstances (going through intersections, and when approaching exits on freeway - even if passing them). And whenever it generally feels like it, flopping around like loose noodles. The speed has to do with the baseline/optimal following distance (in a sterile environment, following a car that's also going a fixed speed), but when the dynamic nature of traffic is applied, it completely breaks with any documented behavior.

I wonder how many people are using AP typically in slow-moving heavy traffic, vs. me, typically driving in clear or light traffic. Today, I drove in heavy/slow traffic, and it was completely different. It seemed to be erratically using "green light => accelerate fast" logic at times, accelerating on its own towards the lead car's bumper faster than they were accelerating, then it'd brake as it notices it's approaching them too fast... like it was using what should be used at green lights as "boost", but in a freeway setting. It's all messed up.

I watched what was happening on the visualization as well, and the visualization never deviated from what the car in front was actually doing. It was pretty well accurate there. The vis was showing my car erratically accelerating too close to them, then dropping back, sometimes lagging behind, but generally actually doing an okay job in heavy traffic. Not nearly as frustrating as it is in lighter traffic.
 
That's what they say in the manual, but that's not the actual reality of how it operates. It's hardly even worth replying to this kind of boiled-dry manual-recital. It doesn't match reality in any way.
I wouldn't have stated that if it was not in my experience. Your experience may be different.

I primarily used AP on the highway, and have found the distance to be dictated by time to be pretty accurate, whether in slow moving bumper-to-bumper traffic or at full highway speed.

It sounds like you primarily use AP in city traffic. There, it may act differently. I have tried it in the city, and I do agree that it is slow off the line from a stoplight.
 
It sounds like you primarily use AP in city traffic.
Really, no, it's probably 75% highway and 25% city, if that much. I "drive bionic", using AP anywhere I see it suitable, and flipping between "riding on rails" with AP, and driving manually, as the road and activities require. It's the highway behavior that bugs me about its "drunkenness" and lack of consistency, where I notice it being floppy and incredibly erratic in its behavior.

I wouldn't be bugged as much by it if it had always been this way, too - it's actually changing, constantly, with the updates over the years. I just feel like earlier AP was much more consistent in its following distance; more rigid, robotic, as I would expect/want it to be. (I remember being able to feel that someone was driving stick-shift, as I could feel when their acceleration let-off with the clutch action 😂 Not any more!). In earlier AP's case, yes, it did follow the speed:distance:time formula. Since it's gotten "more advanced", the result has just been getting sloppier at following, while getting better at steering (and not slamming on the brakes for shadows).
 
Just came off driving a 14-hour combined LA area<->SF area trip this weekend, and I have refined a description of the problem.

It seems to have way, way, way overdone the smoothing on detecting false reactions. If the car is approaching a stopped car, it keeps speeding towards it til the VERY last second it can hammer the (real) brakes to come to a stop. I pretty much always have to brake early to disengage AP and stop, or swirl down the speed setting to get it to start slowing down. It's REALLY bad at higher speeds coming to a stop (e.g. a road with a 45mph limit, going 50 and still getting passed) - as it really just does not seem to notice the cars stopped ahead at a light. (This is with stop-light detection disabled, as I got tired of it braking for every god-blessed sign with a flashing yellow light on it)

On accelerating, it behaves similarly - but the behavior can be different. It seems to have two different modes: one, "throw a chain to their bumper and follow it", and two, "be drunk mode". It'll fall out of "chain to bumper" mode and revert to "be drunk mode" if there's any reason for it to decelerate - and most commonly, that's because of a change in acceleration of the car it's following (e.g. it's a stick shift and they clutch to shift), or if it's crossing an intersection and the car decides it wants more following distance. Then, it'll fall back to "lazily resting its robot foot on the accelerator" and ... make traffic behind me angry as the leading car gets further and further away.

When changing lanes, that's probably the most frustrating part of its problem. If you're trying to get around a very slow-moving car (e.g. 50 on a 65 where traffic wants to go 80 but I keep it to 72), you have to make your move, then immediately track to the speed of the car in front of you. That's how it works - you blend in with the traffic you merged into. But AP doesn't... it'll change lanes, be completely in the new lane, and "la dee dee, slowly accelerating, sucks to be the guy behind me 🤡". Extra large yikes from me. I pretty much always have to goose (to put it politely) the accelerator to remind AP to keep up with the car in front.

When in traffic on the freeway, with heavy traffic it generally performs okay, as it will use a lot of the "chain to bumper" mode (sometimes creeping further into their bumper than I'm comfortable, but almost always entirely disregarding my following distance setting and being almost too close), but in medium-speed-moving traffic, especially the more dynamic it is, the more I wish I had an acceleration-inhibit (without disengaging) as well as an acceleration-inducer pedal.

In general, the common theme is that its reaction is significantly delayed, and often comes close to causing its own forward collision warnings. That worries me about how Tesla is planning to use FCWs in FSD Beta eligibility. In this case, I just really wish we could get back the behavior of earlier AP, which was much more robotic and predictable in its acceleration and deceleration, reliably tracking the cars around it.
 
X-post because it's nice to have in this thread: there's a reasonable theory as to why this was changed:
It needed to be more elastic and smooth as it was unnaturally jerky and made passengers car sick. This was frequent complaint by passengers who then didn't want the driver to use AP at all.
To which I say... yeah, sacrificing driver's comfort for passenger comfort probably isn't the greatest. This should be a switchable option for rigidity/elasticity. I'd pick the middle ground closer to rigid, most likely. This elastic sh*t constantly has me on edge and cursing at the thing!
 
X-post because it's nice to have in this thread: there's a reasonable theory as to why this was changed:

To which I say... yeah, sacrificing driver's comfort for passenger comfort probably isn't the greatest. This should be a switchable option for rigidity/elasticity. I'd pick the middle ground closer to rigid, most likely. This elastic sh*t constantly has me on edge and cursing at the thing!
IMO, more people should be pestering Tesla to use Chill or Standard mode for this. I've suggested this in various ways and I've seen a few others.

Vkt2MBj.jpg
 
  • Like
Reactions: FalconFour
What prompted me to post, specifically though, was that one point where I was following a lead car to the stop, AP actually jerked forward from a stop while the lead car was still stopped and the light was still red. The lead car's brake lights didn't turn off, they didn't move forward... AP just decided "Oop, time to go!" and let off the brakes abruptly. I reflexively just slammed the brake in a split second.

So.... this happened to me in 2019 (Read the video description for the details):


Basically the car loses vision of the truck and it disappears from the visualization. The car on AP then starts to go. I sent this video to Tesla support, and they came back and said the AP team was very interested in this incident. They explained that the shading of the truck actually blended with the surroundings. If you watch the vid, you can see the top of the truck lines up with the sky, the middle lines up with the trees, and the bottom lines up with the road. A freak coincidence on a dreary day that made the truck cloak itself.

This was clearly with radar back then. Seems like vision trumped radar. But your experience of AP suddenly going while a car was directly in front was exactly what kept happening to me with this truck.
 
So.... this happened to me in 2019 (Read the video description for the details):
Oh that's wild. I watched the video - just like my case, the situation doesn't look as dire as it's written. It wasn't accelerating into a stopped/invisible vehicle (at least, not with intent that I'd worry about -- you ever take note of how many people just randomly fully let-off their brakes and creep at a solid red light? LOL Tesla is far better than that 😂), and in my case, I uh... I didn't really want to find out, so I only let it go about 3 inches before I slapped the brakes to stop & disengage.

I'd take yesteryear's rougher/rigid AP over today's hyper-elastic "420 Autopilot" any day. haha.
 
  • Like
Reactions: impastu