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:
You have very different standards for how FSD Beta should drive
Based on what I have read here, it sounds like my standards are pretty normal.

Also have my wife to tell me, if she happens to be in the car. (These days, she says to me: “I used to trust your driving and feel safe in the car, but since you started using FSD I no longer feel that way and you are a worse driver.” Running up on stopped traffic even a little bit is a contributor to this. This tells me I should actually be monitoring and disengaging more, and practicing smooth disengagements (they often aren’t: for example, the dives into right turn lanes - I have to get better at anticipating and disengaging before any movement takes place). Anyway. Have to be way better as a safety driver and preempt FSD before it does anything wrong.)

There is no question in my mind that my experiences are extremely normal, and FSD running up on stopped or slowing traffic and not responding appropriately quickly, as a moderately competent human would do, is completely normal. Sometimes it gets it right, but if it does not ease off as expected, I just disengage (actually have to do it a little early since regen braking is briefly delayed on disengagement).
 
Based on what I have read here, it sounds like my standards are pretty normal.

Also have my wife to tell me, if she happens to be in the car. (These days, she says to me: “I used to trust your driving and feel safe in the car, but since you started using FSD I no longer feel that way and you are a worse driver.” Running up on stopped traffic even a little bit is a contributor to this. This tells me I should actually be monitoring and disengaging more, and practicing smooth disengagements (they often aren’t: for example, the dives into right turn lanes - I have to get better at anticipating and disengaging before any movement takes place). Anyway. Have to be way better as a safety driver and preempt FSD before it does anything wrong.)

There is no question in my mind that my experiences are extremely normal, and FSD running up on stopped or slowing traffic and not responding appropriately quickly, as a moderately competent human would do, is completely normal. Sometimes it gets it right, but if it does not ease off as expected, I just disengage (actually have to do it a little early since regen braking is briefly delayed on disengagement).
Again, different standards. On 11.4.4 on a longish trip in and out of urban areas and back and forth on Highways, with FSD-b set on "Average", neither the SO nor I have had any issues with the slow-down aggressiveness. At all.

Back a release or three ago you correctly pointed out an instability (or at least wildly varying braking) on any slow down. Once alerted to the phenomenon, I agreed with your assessment. On this last trip I distinctly remember thinking that, on 11.4.4, The Car Wasn't Doing That.

Hmm..
 
On this last trip I distinctly remember thinking that, on 11.4.4, The Car Wasn't Doing That.
I’m not talking about the oscillating in slowing. That is largely now resolved. Probably not 100%, but way better than before. As expected, a problem (which many here dismissed), now mostly resolved. Gotta admit the issues.

I’m talking about late reactions. Meaning I see something ahead that is happening that would normally result in a human driver at least easing off, and the car is not reacting. It will react, but too late, and in a way that reduces safety and comfort.

In the events where it is late, and I let it do its thing, it results in heavier braking than necessary, of course. But typically not the oscillating behavior.

So this is another example of a problem, which is not being agreed upon by some here (many do think it exists here though - just have to read the complaints).

In this case, unlike the oscillatory behavior, I think this problem is a lot harder to solve - it requires increasing the detection range. That’s a lot harder than optimizing a smooth stop when everything is quite close with well established positions.

The oscillation was super frustrating since it was clearly just a bug - crappy control system design.

The inability to react in time is more disconcerting but less frustrating. It’s less clearly a bug, and may be a limitation of the system. It’s less frustrating (to me) because I recognize there may be no way to make it better. Of course I hope I am wrong.

On 11.4.4 on a longish trip in and out of urban areas and back and forth on Highways, with FSD-b set on "Average", neither the SO nor I have had any issues with the slow-down aggressiveness. At all.
It’s certainly possible to have drives where a relevant scenario does not arise. In many cases, especially with lead traffic, it does just fine.
 
Last edited:
I can post a picture from my Model X showing that I indeed bought FSD. Would that work for you? PM or here? I can write a nice message for you too :)

It would be nice with a convo that doesn't get to the "troll" and poor attempts at character assassination in five posts just because FSD deposn't work very well.
This was the first full post of yours I’ve read. There’s just so very many words.
 
I’m not talking about the oscillating in slowing. That is largely now resolved. Probably not 100%, but way better than before. As expected, a problem (which many here dismissed), now mostly resolved. Gotta admit the issues.

I’m talking about late reactions. Meaning I see something ahead that is happening that would normally result in a human driver at least easing off, and the car is not reacting. It will react, but too late, and in a way that reduces safety and comfort.

In the events where it is late, and I let it do its thing, it results in heavier braking than necessary, of course. But typically not the oscillating behavior.

So this is another example of a problem, which is not being agreed upon by some here (many do think it exists here though - just have to read the complaints).

In this case, unlike the oscillatory behavior, I think this problem is a lot harder to solve - it requires increasing the detection range. That’s a lot harder than optimizing a smooth stop when everything is quite close with well established positions.

The oscillation was super frustrating since it was clearly just a bug - crappy control system design.

The inability to react in time is more disconcerting but less frustrating. It’s less clearly a bug, and may be a limitation of the system. It’s less frustrating (to me) because I recognize there may be no way to make it better. Of course I hope I am wrong.


It’s certainly possible to have drives where a relevant scenario does not arise. In many cases, especially with lead traffic, it does just fine.
I just had the red caliper and upgraded brake pads done on my MS.
I asked the tech to keep my old pads taken off. After 40k miles mostly on FSD they looked like they were barely used. Which I found hard to believe due to all the braking FSD does. Apparently the added regen, although which sucks on the S, does help quite a bit with brake wear.
 
Last edited:
I just had the red caliper and upgraded brake pads done on my MS.
I asked the tech to keep my old pads taken off. After 40k miles mostly on FSD they looked like they were barely used. Which I found hard to believe due to all the braking FSD does. Apparently the added regen, although which sucks on the S, does help quite a bit with brake wear.
Yeah the brake wear isn’t really a concern. Regen does most of the work.

Probably can’t make it to a million miles on the original pads with FSD’s predilections, but not really that important.
 
Huh? I use it regularly in stop and go traffic and slow-moving rush hour traffic. No issues what so ever.
This is not what I am talking about, obviously.

How is that not clear? If nothing else, from further context!!!

We were talking about freeway driving. This implies speed. Further context clearly indicated the scenario. Encountering events that are at the edge of perception range…etc…lead traffic helps…etc.

Can also apply to something as simple as a traffic light.

Not complicated stuff. The fix may well be impossible though!
 
Last edited:
This is not what I am talking about, obviously.

How is that not clear? If nothing else, from further context!!!

We were talking about freeway driving. This implies speed. Further context clearly indicated the scenario. Encountering events that are at the edge of perception range…etc…lead traffic helps…etc.

Can also apply to something as simple as a traffic light.

Not complicated stuff. The fix may well be impossible though!
Odd. I’ve been doing quite a bit of freeway driving on FSD. And, just to get this out of the way, “normal” TACC, up to and including stop and go, hasn’t been an issue.

What I think you’re talking about is coming up on stopped traffic. With, say, nobody in front of one that’s already slowing down.

If this is what you’re talking about, this is precisely the dangerous scenario that used to drive those using the RADAR-based TACC nuts, because the original TACC couldn’t see cars that were traveling faster or slower than 45mph off their own speed. Had to do with reducing RADAR clutter so the car could track other cars off to a quarter mile distant. But it meant that the human driver HAD to intervene and apply the brakes, since the FSD of the time wasn’t going to. (And, yes, there were accidents. The notable ones involving emergency vehicles stopped on the roadway, because of course they do.)

This is a heck of a way to run an FSD application, so I suppose it wasn’t a big surprise that Tesla switched to full vision for TACC, since vision doesn’t have that, “It disappears when it’s stopped” bug baked into the technology.

Early versions of vision were were definitely shaking out bugs in this approach; again, if this is what you’re talking about, those early versions were definitely braking late and hard, not to mention that buggy brake hard/brake light/brake hard sequence we had fun complaining about.

To my mind, most if not all of that seems to have smoothed out. Yeah, a human will come off the gas if they spot, say, stopped traffic a half mile up or something and do a more gradual slow down. But the braking rate on the current rounds of 11.4.x doesn’t seem that hard or uncomfortable to the SO or me. And that covers 11.4.7.2 on the dearly departed 2018 M3, the 11.4.4 on the 2023 M3, or the SO’s 2021 MY running EAP.

Or have you got something else in mind?
 
To my mind, most if not all of that seems to have smoothed out. Yeah, a human will come off the gas if they spot, say, stopped traffic a half mile up or something and do a more gradual slow down.
Yeah this is what I am talking about, and to suggest it is truly fixed is not correct. It works sometimes when sightlines are good, it’s a bit uphill, etc.

It would not be difficult to produce a video showing it does not work in many cases though.

However, I’m not going to do that. I’m sure there are plenty of people here who can confirm my experience. Means I don’t have to put in the effort.

Will it stop? Sure, nearly certainly. Is it comfortable? No. So, disengage and report.

I drove 1000 miles or so recently on FSD, mostly freeway, and there were several instances of this that I can recall.

Different issue - it often stops way too abruptly when someone in front comes to a swift, normal halt. Happened to me just this evening. Came to an annoying halt at least a couple car lengths behind the other vehicle stopped at the light. So annoying. Poor. There seems to be a threshold where it freaks out and decides to leave plenty of margin. And that’s all very well - if there is a legitimate reason. But it’s not very intelligent about it so it seems many normal scenarios are treated as high risk.
 
Last edited:
Based on what I have read here, it sounds like my standards are pretty normal.

Also have my wife to tell me, if she happens to be in the car. (These days, she says to me: “I used to trust your driving and feel safe in the car, but since you started using FSD I no longer feel that way and you are a worse driver.” Running up on stopped traffic even a little bit is a contributor to this. This tells me I should actually be monitoring and disengaging more, and practicing smooth disengagements (they often aren’t: for example, the dives into right turn lanes - I have to get better at anticipating and disengaging before any movement takes place). Anyway. Have to be way better as a safety driver and preempt FSD before it does anything wrong.)

There is no question in my mind that my experiences are extremely normal, and FSD running up on stopped or slowing traffic and not responding appropriately quickly, as a moderately competent human would do, is completely normal. Sometimes it gets it right, but if it does not ease off as expected, I just disengage (actually have to do it a little early since regen braking is briefly delayed on disengagement).
When Tesla reports that the FSD is x times better than a human, how do they compile the numbers? This is an honest question. If a human is obsessively watching over the FSD, this by itself should make the drive safer, no?
 
Um. So, first off, I'm running 11.4.4. When I got the new car it came with 11.3.6, which was Brought Back Bad Memories: abrupt steering, swinging wide right on left turns (even worse than 11.4.4), and, well, not so great; before I got rid of the old M3, it was on 11.4.7.2, I think.

11.4.4 is smoother on turns and handles highway miles pretty darn well; on the Merritt it swung back and forth between passing and traveling lanes and would definitely get out of the way of cars coming up from the rear.

Now, when running around through the interstices of Boston roads.. well, I would let it do its thing if the traffic wasn't too bad, but if things were getting decidedly hairy, I'd take over and drive. Having said that, 11.4.4 went through a number of traffic circles without incident.

But I was still intervening on city streets; oh, say, about once every three to five miles or so. With occasional ten mile or more stints, depending upon traffic, lane markings, and construction. On interstates, pretty much no interventions.

So, don't get the idea that I think the system is faultless: It is a Beta. But compared to, say, Spring of 2022, it's pretty darn good.

This is a bold statement from me, since 11.3.6 was always my favorite.
I think 11.4.7.3 is the best we will get until the v12 branch.
I have had a bunch of improvements, and very few dry wipes.
I keep it in aggressive mode, and boy does it like to change lanes, even appropriately most times, with small gaps. The late braking is my biggest complaint.
Also, why can't MS owners adjust regen like others. The MY regen is far superior to the S. The windshield wipers are also. They actually clean your windshield on the Y. I bought some RainX wipers. 2 sizes for the S, 26 and 20 inches. They are way better than oem blades. Still not as good as the Y.
I agree that I have not had any dry wipes since last Dec except if there is leaf debris on my windshield or my neighbor's never ending stone work or whatever they have been doing for a year. I park my car on the driveway. So I have to expect I guess. My concerns are always lane change issues and the car driving very fast on those lanes that are going to end. Yesterday, am MB was very close to me because it was a merge lane and my car speeded and he was speeding to beat me.
 
  • Informative
Reactions: FSDtester#1
When Tesla reports that the FSD is x times better than a human, how do they compile the numbers? This is an honest question.
It's a good question, because it's mostly marketing and not based on research. Tesla compares Tesla highway-only miles to all miles driven (all roads), and Teslas cars (which are modern) to the whole fleet on the US roads (which includes 20-30 year old cars).
If a human is obsessively watching over the FSD, this by itself should make the drive safer, no?
Yes, if the system is unreliable, which I think it's fair to describe FSD as at present.

When monitoring a poor system, we know it will fail often, and hence we are more alert.

As the system gets better the supervision contributes close to nothing in time-critical applications according to 40 years of research across multiple fields. The system needs to be superhuman by itself to claim "safer than a human" in other words.

This human failure mode is often called "automation complacency". Humans need to be in the loop to be decent at monitoring (and be professionally trained etc).

1699022004997.png


 
Last edited:
  • Informative
Reactions: kabin
Thanks for reminding us that you have very different standards for how FSD Beta should drive. I've never disengaged it just for stopped or slowing traffic.
I do, but that's just a preference. I hate driving at full speed towards slowing cars even if I know the system will suddenly slow in enough time. I prefer a smoother driving experience.

Most likely, FSD doesn't see the upcoming slowdown/stoppage so has to do last moment reactions because, in the words of my driving instructor, it is not looking far enough down the road. If it does see more than my screen indicates, then the only reason for Tesla's choice that I can think of is there is more regen when brakes are applied as opposed to the regen from slowing down.

Aside from the safety that comes from looking down the road and anticipating slowdowns, I mastered driving smoothly because it increased my gas mileage. I recognize that EVs are different but I have trouble believing that the energy savings I experienced in an ICE through smooth driving would not also translate to less energy used by an EV. Certainly when it comes to highway speed the same physics rules apply (more speed = more drag = more energy used). I have a hard time believing the extra energy recovered when driving at a higher speed toward the slowdown and grabbing all that regen, is greater energy than required to return to speed if the slowdown clears while one approaches it slowly so the car never has to come to a stop or extremely low speed.
 
have a hard time believing the extra energy recovered when driving at a higher speed toward the slowdown and grabbing all that regen,
It’s most efficient to regen as little as possible, so it is more efficient to slow down more slowly.

The less regen, the better! (And of course you don’t want to brake at all.)

(There are probably complications due to aero losses - maybe it is better to initially regen a little faster to get rid of most of that loss. But you’d still want to start the process early. Early, faster slowing might also mean you can maintain more speed and never come to a stop, too. But anyway that is not what FSD is doing; it’s so far removed from the optimized stopping profile that the exact details of the absolute most optimal way to stop don’t really matter.)
 
Last edited:
Just a heads up for refresh MS peeps, service just sent me a message regarding a follow up visit to fix some exterior trim. They included this new language:

Concern: Update Vehicle Software to Reduce Vibration During Acceleration Repair Notes: 4 Correction: SB-23-00-006

So there is now a software update to fix vibration????
 
Last edited:
  • Like
  • Informative
Reactions: bhakan and rlsd
Just a heads up for refresh MS peeps, service just sent me a message regarding a follow up visit to fix some exterior trim. They included this new language:

Concern: Update Vehicle Software to Reduce Vibration During Acceleration Repair Notes: 4 Correction: SB-23-00-006

So there is now a software update to fix vibration????
Not that my eyebrows aren't going up a bit as well, but it does make sense. The (two) motors run on pulse-width-modulated pulses with a ridiculous number of electronic control loops, looking at current, temperature, rpm (both of the fronts and rears), with hardware assist on the PWM end of things.

I've worked on fan motors and the design of the surrounding electronics, which aren't nearly as complex as what Tesla plays with. And it's easy to make a mistake somewhere. Some mistake putting poles in the right half plane resulting in surges.. sure, it's possible. And given manufacturing variations on silly things like resistance, inductance, capacitance, and what-all, it's believable that what the engineers played with Worked Just Fine.. but then, given big statistical runs of parts, had corner cases that would lead to instability. Which doesn't take much thought to realize would lead to vibration.

So, they hit some poor engineer with a wet noodle for a while, asking, "Just when did you take that control theory course, anyway?!?", then get a new load of software that keeps the poles in the left half plane, where they belong.