Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.


Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.


Here is the AI Day explanation by in 4 parts.


screenshot-teslamotorsclub.com-2022.01.26-21_30_17.png


Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

 
Last edited:
Thats why the gap needs to be dependent on speed. Yesterday FSD decided to stop atleast a car length behind the stop line at a traffic light - and I had to disengage and move forward since the left turn signal was not getting triggered !
Yes, going 20MPH and leaving 2 car lengths on the Interstate is an invitation for cut ins. Odd that it seems to keep a much tighter gap at the same slower speeds on streets than Interstates. Probably just being extra conservative on the first few releases of the highway stack.
 
Yes, going 20MPH and leaving 2 car lengths on the Interstate is an invitation for cut ins. Odd that it seems to keep a much tighter gap at the same slower speeds on streets than Interstates. Probably just being extra conservative on the first few releases of the highway stack.
The "2 second rule" is good and works for all speeds. FSD should allow us to set the following distance in seconds rather than car lengths.
 
A bit of a rhetorical question, but partially serious- why is it a problem if someone cuts in? The goal should be to arrive at your destination safely, not necessarily before everyone else. If someone cuts in how much does it affect you? A second at the most?

Back when I took driver’s ed I was taught to leave 3 seconds between me and the next car. It seems like most people want to leave about 500 milliseconds.

Ultimately this comes down to driving styles, philosophy and people being irrationally competitive.
I can only speak for myself but it's not about being "irrationally competitive". If there's particularly slow traffic on the right lane of a 4 lane highway (2 lanes each way) and you're in a line of cars in the left lane trying to get past them, then cars repeatedly speed up on the right and cut back in front of you, it's annoying to you and the line of cars behind you, ultimately creating an unsafe situation in its own way if you will. All I'm trying to do is get around the truck/camper etc... backing up traffic on the right then plan to pull back in the right lane and cruise at a reasonable speed at a reasonable following distance.

Sure, the absolute safest thing may be to creep along in the right lane at 55 mph in a 75 mph for an interminable time but I'd rather not. Safer yet to not even drive in heavy traffic I suppose but reasonable people make reasonable decisions about safety every day. Thankfully we don't all have to abide by the dictates of the most risk-adverse among us at all times. I see "safety" bandied about here a lot as if it's the ultimate trump card, discussion closed, but it's not IMO. Reasonable people can, and do, disagree and that's OK. And don't get me started on book banning.... ;) 😵‍💫
 
  • Funny
  • Disagree
Reactions: sleepydoc and EVNow
I see "safety" bandied about here a lot as if it's the ultimate trump card, discussion closed, but it's not IMO. Reasonable people can, and do, disagree and that's OK.
Talking about using safety as the ultimate trump card ... ;)

you are now also going to beep at me every time I press accelerator to actually close the following distance to the safe range,
 
On the topic of disengagement: After a go pedal nudge, FSD beta often decelerates to a fraction of the set speed. It seems like if it’s set to 70 and parks itself in somebody’s blind spot while blocking the left lane, I nudge it momentarily, to 75 to get out and ease off the pedal, it will go down to 30 and then slowly accelerate back to 70, even on the highway with cars whooshing by at 70 mph. The only way I’ve found out of this is to disengage, accelerate and re engage. Anyone found a way around this?
I offer this only as a possible explanation. I have not experimented to confirm my hypothisis.

Sometimes, when I use go pedal and then slowly back off to slow down, the accelerator continues to control the speed even as we decelerate below the set speed. I notice this when I then release the accelerator and the car speeds up. Odd that coming off the accelerator results in increased acceleration! Perhaps this is what you experienced.

Perhaps when we push the accelerator, FSD changes speed control mode into pedal control. It only switches back to TACC when the accelerator is released.
 
I offer this only as a possible explanation. I have not experimented to confirm my hypothisis.

Sometimes, when I use go pedal and then slowly back off to slow down, the accelerator continues to control the speed even as we decelerate below the set speed. I notice this when I then release the accelerator and the car speeds up. Odd that coming off the accelerator results in increased acceleration! Perhaps this is what you experienced.

Perhaps when we push the accelerator, FSD changes speed control mode into pedal control. It only switches back to TACC when the accelerator is released.
I've noticed the same recently - though I've not noticed this before the past weeks/months, maybe something has changed in this area.

BTW: my Nissan Leaf does exactly this, adaptive cruise control will only control speed if your feet is fully off the accelerator.
 
  • Informative
Reactions: old pilot
Be glad when this is fixed: Makes a RH turn onto the rightmost lane which turns into a well marked right turn only lane. Rather than moving to the LH lane and continuing straight it stays in the right turn only lane 'til it ends and then brakes, unsure where to go, does the "what do I do?" dance and I have to take over.
View attachment 927704
On 4.2 now and this issue is fixed - signals and moves to the left to continue straight well before the RH lane ends very smoothly!
 
  • Like
Reactions: FSDtester#1
Let me start with I haven’t driven 11.4.2 yet. It’s installing now.

I simply do not understand how anyone out there still believes this crap, “At this pace, we will achieve self driving safer than a human this year.” Seriously!? What pace? Most of my trouble spots from 10.2 all those years (year) ago are still there in 11.3.6? How can this be called pace at all? There has been no “eureka moment,” and there never will be. I know Chuck talked about a chatGPT moment, but chatGPT is a fundamentally different product that doesn’t kill people when it makes mistakes, it only gets people Fs on exams.

With each update we continue to have hope. We continue to believe that it is 🔥or even 🔥🔥. Oh, but yet we are always disheartened with each drive after an update. An update supposed to be 🔥 is nothing more than a slight improvement, or sometimes a regression.

And now, our suspicions have been confirmed. We HW3 owners are doomed. HW4 has arrived.
Yes, they are still saying that HW3 will achieve FSD (of some sort), but as soon as all cars have moved to HW4 and people have bought all the HW3 cars, Tesla’s attitude toward HW3 will start to change.
We have seen this movie before. How is HW2.5 doing these days? I thought that those cars would be able to drive themselves? Oh wait.
I say T-minus one year before HW4 starts on a different, better, software track than HW3. The problem of self driving was not solvable on HW3, so they fixed it with hardware.
My dad taught us what he called the "Socratic" method, so here is one cup-half-full, Elon's Advocate speculation.

Perhaps they are largely disabling map and sign reading inputs so they can improve FSD's ability to deal with situations where map and sign info is unavailable or incorrect. Once it can safely handle driving on pure, illiterate vision, they will add map and sign reading, and we will see (no pun intended) a sudden vast improvement overall. i.e. They are solving the hard problem first.

The current robotaxi firms have used extremely fine detailed maps plus lidar, plus geofencing, and plus remote monitoring and intervention, none of which is compatible with consumer priced FSD. (And after the next big quake in SF, all those aids will stop working, and we can add robotaxi patrons to the number of people stuck in inoperative elevators. ;))

So, perhaps Tesla's development strategy is to make FSD learn to walk without crutches because that is faster and safer than letting it use crutches while it is trying to learn how to walk without crutches. In the mean time, us FSD β testers are the crutches, and every time we disengage, Tesla gets another data point about an opportunity to improve. Maybe we could think of this a a bunch of cars limping around while learning to walk without running into things, and driver's licenses won't be given out till they know how to walk.
 
I offer this only as a possible explanation. I have not experimented to confirm my hypothisis.

Sometimes, when I use go pedal and then slowly back off to slow down, the accelerator continues to control the speed even as we decelerate below the set speed. I notice this when I then release the accelerator and the car speeds up. Odd that coming off the accelerator results in increased acceleration! Perhaps this is what you experienced.

Perhaps when we push the accelerator, FSD changes speed control mode into pedal control. It only switches back to TACC when the accelerator is released.
I think that's exactly it.
 
I can only speak for myself but it's not about being "irrationally competitive". If there's particularly slow traffic on the right lane of a 4 lane highway (2 lanes each way) and you're in a line of cars in the left lane trying to get past them, then cars repeatedly speed up on the right and cut back in front of you, it's annoying to you and the line of cars behind you, ultimately creating an unsafe situation in its own way if you will. All I'm trying to do is get around the truck/camper etc... backing up traffic on the right then plan to pull back in the right lane and cruise at a reasonable speed at a reasonable following distance.

Sure, the absolute safest thing may be to creep along in the right lane at 55 mph in a 75 mph for an interminable time but I'd rather not. Safer yet to not even drive in heavy traffic I suppose but reasonable people make reasonable decisions about safety every day. Thankfully we don't all have to abide by the dictates of the most risk-adverse among us at all times. I see "safety" bandied about here a lot as if it's the ultimate trump card, discussion closed, but it's not IMO. Reasonable people can, and do, disagree and that's OK. And don't get me started on book banning.... ;) 😵‍💫
why not just tap the accelerator for the occasions when you need to go a bit faster?
 
I do, though it’s generally not a “tap”, it’s a push until the traffic clears which sometimes takes time. Why not just let us choose our own following distance? It used to give us a better range so it’s a simple fix.
Its definitely not a "simple fix". Tesla does not feel confident they can react fast enough to make it safe.
 
The "2 second rule" is good and works for all speeds. FSD should allow us to set the following distance in seconds rather than car lengths.
Even when FSD following distance was based on a number, its quite clear that it did not represent car lengths. It was an abstracted number, a relative distance that lengthened with speed. It was more of an adjustable time-based gap, but not strictly that either.

Now it's not a number anyway, so that discussion point has become moot.

I agree that the two-second rule is good for safety, but as I've said before, I see almost no one follow a two-second rule, much less three seconds which is being taught now. As others said it's really closer to a half-second for many people. Tailgaters, even less.
 
I agree that the two-second rule is good for safety, but as I've said before, I see almost no one follow a two-second rule, much less three seconds which is being taught now. As others said it's really closer to a half-second for many people. Tailgaters, even less.
That is insane - and not at all common here. As ADAS becomes more common - I think the follow distances will get back to being sane ;)
 
I offer this only as a possible explanation. I have not experimented to confirm my hypothisis.

Sometimes, when I use go pedal and then slowly back off to slow down, the accelerator continues to control the speed even as we decelerate below the set speed. I notice this when I then release the accelerator and the car speeds up. Odd that coming off the accelerator results in increased acceleration! Perhaps this is what you experienced.

Perhaps when we push the accelerator, FSD changes speed control mode into pedal control. It only switches back to TACC when the accelerator is released.
With both FSDb and TACC if you use the accelerator to exceed the set speed for more than a few seconds the car will cede speed control to the accelerator pedal. Unlike traditional cruise control it won’t pick up if you drop below the set limit, either. Only when you completely let off will it resume speed control. It will also transiently decelerate before doing so.
 
  • Informative
Reactions: JHCCAZ
All right. Got 11.4.2 late last night and took it out for a drive. Then, today, the S.O. and I went out for lunch and a trip to a grocery store.

Synopsis: The 40 mile trip last night was near perfect. The ten mile trip today.. two nasty interventions and a discovered regression.

The long trip:
  • Right turn onto heavy traffic with an immediate slew into an on ramp. People who live around here sometimes wait five minutes or so for a stop light roughly a mile away to turn red, at which point one can turn right. 11.4.2 did see the piles of cars, trucks, and semis heading for the interstate at 50mph+ and, correctly, stopped and waited. But there must be some kind of a timer: After the second or third semi-gap (that is, just enough of a gap where if one floored it, one wouldn't be turned into a pancake by rear-coming traffic), it tried to go into traffic but without flooring it. I intervened with the gas pedal and got out of there.
  • The next thing was a right exit onto an interstate. Smoother than 11.3.6, better tracking down the middle. Merge was OK, about the same as 11.3.6.
  • Twelve miles on a 3-lane interstate, left lane pulling 70+, right lane around 60. Contrary to other reports around here, with the speed set to 70, the car would go into the left lane to pass people. And, when either a gap opened up on the right or people approached from the rear, would move out of the passing lane into the middle on its own. It did this several times; so, those reports about 11.4.2 hogging the left lane appear, to be on the face of it, not quite up to snuff.
  • Got onto the exit ramp; stopped at the light a bit further up; turned left sans problems, right a few miles up, and another couple of rights ending up in the parking lot of a supermarket I had picked at random. No interventions. Smooth on the turns, just a trace of jerkiness on the final right into the supermarket. So far, so good.
  • The above was pretty much all interstate. Figured a decent test would be better with 90% local roads, so did that. Navigated out of the parking lot and retraced steps to the interstate. The on-ramp turned almost immediately into an off-ramp onto River Road, which is just about as curvy as one would expect.
  • Transitions from one lane to two lanes and back handled well; the 11.4.2 was picking the appropriate lane sans problems. At the end of the River Road, protected left onto a heavily traveled business district. Handled lights, other cars, and all that jazz no problems.
  • Handled the fork to the right correctly at the end of the business district; navigated and didn't run into pedestrians who, in two cases, attempted suicide by leaping between cars into the road on their ways to the driver's side doors on their respective cars. First one the car came to a near halt; second one it swerved away from the death cult guy, both without fault.
  • Ended up on the three-laner, 50 mph, local highway with Jersey Barrier and lights. It went into the left lane at the appropriate spots, back out again, and, in a first, got over into the rightmost lane for the exit 2/3 of a mile from said exit, rather than trying to wait until the last minute. Good boy, have a scooby snack!
  • Nice exit ramp, came up the hill, and, with no traffic coming up from the rear, went through the yield sign at speed. OK, no traffic, so that was fine.. but I would have slowed down a bit for a double-take. But I only have two eyes, the car's got more, so, maybe OK?
  • Back to the house was fine and no interventions.
Short trip: Not so great.
  • Got out of the development sans problems, a bit faster than 11.3.6. Next intersection was a stop sign with an unprotected left. Took it a bit, went through, faster than 11.3.6, but not as fast as I would have done it.
  • First big problem. So, this one lane road goes over two big bridges and continues on. On the bridges, there's the double yellow in the middle, but the road probably should have two lanes marked on both sides. Doesn't, though. So, cars wanting to take the unprotected left on the second bridge to get on the Major Highway NorthBound are advised to hug the double yellow so $RANDOM_DRIVERS can pass on the right. On early versions of FSD-b, the car had a habit of swerving all the way to the right curb (two lanes, remember!), halting, and then going through to the left.
  • But this, "swerve right on a left" was a bad habit seen nearly everywhere. In a protected left turn, waiting at the head of a line of cars with for the left turn signal to go green? The steering wheels would be canted to the right and the car would hug the right lane marker, on a be-jeezing left turn. In fact, on any road, with a left, the car would attempt to swerve right.
  • On 11.3.x, this behavior had, finally, been reduced, but not eliminated. The car would stay move away from the double yellow, but not too much. Today, on 11.4.2, the unexpected swerving darn near caused an accident. Approaching the left turn, the car went a full lane to the right. Cars behind me saw the right swerve and slowed down, except for the fellow who was directly behind me who, clearly, thought, "Jeez, why is that idiot doing that? Let me go around him!" and proceeded to do so. Just as the car tried to turn left, directly into him. I intervened and got up against the right curb, got strange looks as the massive crowd passed me on the left, and then had to proceed up half a block to pull a U-turn to go back. That was.. bad. Very bad. That's a regression, gentlepeople. I was too busy not getting killed to leave a report.
  • So, on the exit. This is an on-ramp, interstate style. One has to get to the left one lane, otherwise one is going off into never-never land. There was a car to my right as 11.4.2 sped up. What it should have done: Slowed down, then moved in behind the guy on the left. What it did: Steady speed and tried to go for never-never lane. Intervened. And this time left a snippy report.
  • Rest of the trip for food was fine, smooth driving, moved into the exit lanes correctly, and all that. Return trip was fine, too.
I had been hoping that this little excursion today would have convinced the S.O. that Tesla's getting better. Well, so much for that.

The car is driving smoother and, when not trying to make an accident, does a better job of 11.3.6 getting in and out of lanes and all. But it's still not ready for prime time.
 
Last edited:
Its definitely not a "simple fix". Tesla does not feel confident they can react fast enough to make it safe.
I disagree, they've done it before. I don't recall reading about accidents being caused by the previous setting - perhaps you have a reference?

I'm not advocating unsafe driving but the more the system drives like real people the faster adoption by mainstream users will occur IMO. This goes for the tentativeness at stop signs, roundabouts etc... Many here will attest their SO's can't stand riding with it engaged for just that reason. Edit: And WHY does it hate getting in the right lane? But I digress. Those of us whom have bought in are probably more patient and forgiving than most.

I'm not trying to convince you or anyone else of my opinion, just stating my preference for system performance in a specific situation, just as you and others have taken liberty to do.
 
And, when either a gap opened up on the right or people approached from the rear, would move out of the passing lane into the middle on its own. It did this several times; so, those reports about 11.4.2 hogging the left lane appear, to be on the face of it, not quite up to snuff.
I hope that's finally fixed! Just finished a 2700 mile trip on 3.6 and having to constantly signal to move to the right lane got old quickly.
 
Except it does so even when there the view is unobstructed. That indicates it’s more of a software issue than due to the camera location.
It's hard to know but it could be a bunch of things. One thing is for sure. FSD's visual perpendicular kinematic estimation has been consistently crap which can't help when the ego approaches stop signs. Maybe image motion blur, low camera resolution, inadequate multitasking, poor estimation of target vehicle paths and kinematics, system latency... It's a WAG but I suspect these unsolved issues are the main reason for crutches like slow approaches to stop signs, stopping short, and creeping to stop signs.
 
  • Like
Reactions: Sharps97