FSD Beta Videos (and questions for FSD Beta drivers)

[momo3605]

Yes, 12mph is too slow, and I said it should go faster. In the commit case, going up to 25mph seems fine. In the "waiting" case, I think 15-20mph would be fine and is completely realistic. There's no need to get up to 25mph in that case; it's possible to cross in about 5 seconds with a reasonably rapid increase to 15-20mph. (Again, right now, it's limited to about 12mph, and it attains that speed slowly, and it takes 6-7 seconds to clear the traffic lanes in that case.)

I'll see if I can get an example on my ULT, even though I don't want to stop in the median (don't really like the strategy in most cases, though Chuck's is fine due to the median width), as it's narrower. I can do it with minimal traffic. The requirement is that it be smooth and reassuring.

Here's another link to the earlier example with traffic, but I didn't stop. Though I did have to stop accelerating, to "thread in." So it's somewhere in between. It's about 4.5 seconds to cross here. Started at 2:48, finished crossing at around 2:52.5. I got up to about 17mph before accelerating again.

I could have crossed faster if I were committing, of course - because I was being cautious here (just as well because the BMW changed into the target lane as I went!). But anyway in the "stopping" case with an ample median width, 5 seconds to clear seems fine. That would be a 20% reduction. Just increase all speeds by 30%, peaking at 16mph.

So looking at your video, you're doing about 19 entering the median. And in Chuck's video, FSD gets to about 15mph at the same spot. Not too far off tbh.

 

[AlanSubie4Life]

So looking at your video, you're doing about 19 entering the median. And in Chuck's video, FSD gets to about 15mph at the same spot. Not too far off tbh.

This is 25% faster, which is exactly what I was suggesting would be adequate.

Also the one you pictured of his was a commit case, and mine was not (note the regen bar - it goes to zero because of the BMW, and general caution due to the traffic, 2.5 seconds after the commit - and I would have been accelerating more if there had not been traffic). I can clear the traffic lanes much faster than FSD, without any drama or jerk, when I'm not worried about crashing into someone. I'd estimate somewhere in the 20-25mph range at that point (at the point when I exit the near traffic lanes).

Here's an example of a commit from me. Using FSD sure gets the heart rate up, though (this was after a disengagement). My resting heart rate is about 70 during the day when sitting:

 

[sleepydoc]

Correct me if i'm misremembering, but i recall you complaining that 12mph was too slow and the car should be able to go up to 25?mph? If you get a chance, i'd love to see an example or video from you of accelerating up to 25mph and then stopping perfectly in the median for traffic on the other side. IMO that doesn't seem realistic at all.

He also has to do it only using regenerative braking since using the mechanical brakes at all is apparently an automatic fail!

 

[AlanSubie4Life]

He also has to do it only using regenerative braking since using the mechanical brakes at all is apparently an automatic fail!

I love this comment; so true. To clarify, I would say that using mechanical brakes at some points is physically required. Sometimes it's just the way things go (light timing, lead traffic behavior, downhills, etc.). And using the mechanical brakes in the last 1-0mph is also sometimes good as a way to promptly stop so you can go again (and required if not using hold mode!), when used judiciously. But certainly in the above case using friction brakes would be a travesty (I know exactly what I need to do, so no need for brakes!).

 

[sleepydoc]

I love this comment; so true. To clarify, I would say that using mechanical brakes at some points is physically required. Sometimes it's just the way things go (light timing, lead traffic behavior, downhills, etc.). And using the mechanical brakes in the last 1-0mph is also sometimes good as a way to promptly stop so you can go again (and required if not using hold mode!), when used judiciously. But certainly in the above case using friction brakes would be a travesty (I know exactly what I need to do, so no need for brakes!).

Just how do you propose to accelerate to 25 MPH and then decelerate to 0 MPH in approximately 30-40 feet without using friction brakes?

 

[AlanSubie4Life]

Just how do you propose to accelerate to 25 MPH and then decelerate to 0 MPH in approximately 30-40 feet without using friction brakes?

This is not what is being proposed. The current limit is 12mph. For a pose in the median, the proposal was to up that to 15-20mph peak. 16-17mph would be great. For this median, you would have to take a path like this, to fit comfortably (and only when there is no Miramar left-turning traffic - though sometimes they hang back and let people go; people are very nice, though usually this just causes confusion since they might be doing a u-turn too; I prefer they proceed, and block Padgett left-turning traffic).

And it's not 30-40 ft. It's 115 feet or so. 115ft at an average speed of 11mph (probably possible) is 7 seconds. This path leaves most of it traveled out of the traffic lanes, so I think it's possible to be exposed for around 4 seconds (a big improvement over current 6 seconds).

We'll see! I predict I might have to use my brakes lightly at the end (because going across the double yellow is technically not allowed). Starting slightly further to the right (but not blocking people) is a good idea anyway since it gives (slightly) better visibility.

 

[sleepydoc]

I predict I might have to use my brakes lightly

then you fail by your own criteria.

 

[AlanSubie4Life]

then you fail by your own criteria.

Just a prediction! We'll see!

As I've said, it's actually ok to use the brakes when you're going very slow (like, less than 3mph ideally); it's often appropriate and it's of minimal consequence energy wise; the car can roll for a while if you don't do so, and it can be somewhat wasteful of time, so bringing it to a halt in that case is fine (for example, a stop sign). The real thing I take issue with is jamming them on at 30mph, for no apparent reason at all.

But I'll try to not use my brakes here! I don't think it will be necessary.

 

[AlanSubie4Life]

Would your method trigger the excessive acceleration and braking metrics that insurance companies use in their monitoring? If so, then IMO FSD shouldn't do it.

He also has to do it only using regenerative braking since using the mechanical brakes at all is apparently an automatic fail!

Just how do you propose to accelerate to 25 MPH and then decelerate to 0 MPH in approximately 30-40 feet without using friction brakes?

This is not what is being proposed. The current limit is 12mph. For a pose in the median, the proposal was to up that to 15-20mph peak. 16-17mph would be great.

I did this tonight after getting some delicious Indian food from Punjabi Tandoor. I cleared all three traffic lanes and was established in the median in 4.2 seconds. The remainder of the coast down was in the median. No brakes were used. The double yellow area was not violated. I am not posting the video because I want to get a better one: you can't see my foot is not on the brake, it wasn't an optimal situation - there was some traffic so I couldn't focus on getting the maneuver right, for an undisclosed reason someone will probably point out, and because it was my first attempt; I think I can do it more smoothly and better.

1) Peak velocity 17mph, attained in 2.38s (304-233)/30 = 2.38s
2) Began significant slowing at (314-233)/30 = 2.7s
3) Cleared the traffic lanes around (360-233)/30 = 4.2s
4) Nearly completely stopped (2mph) at (420-233)/30 = 6.23s

The only Safety Score impact here would be "Aggressive Turning" - a low impact factor in Safety Score.

Specific improvement can be to proceed further across the road into the median so I can clear the traffic lanes sooner. I was a little conservative since I didn't want to stop in traffic lanes on the far side (again, there was some traffic in the distance so I didn't want to mess with it). You can see I am definitely more on the left side. I think it is probably possible to clear those lanes in 3.5 seconds or so, but it would require a slightly tighter turn.

It will have to be dark because there is too much traffic on this road at all other times.

Anyway, to compare to that 4.2s/17mph, as a reminder FSD Beta takes closer to 7 seconds to do this by the same metric, peaking at 12mph, and it's much easier to do quickly with a wider median! So there's definitely room to cut off a couple seconds without too much drama. (And they will have to if they want to clear 5-second traffic gaps (when stopping in the median and not rolling all the way through), even with a little preroll before the first vehicle passes.)

The obsession with braking in this situation is silly, anyway, even though it's not necessary. You can clear the lanes faster and establish a safer position more quickly with a small (non-jerky!) amount of brakes at the end. It's definitely possible to do it without exceeding 0.3g (0.3g from 17mph is 2.6s, and this regen-only slowdown from 17mph to 2mph took 3.5s) and avoid a safety score insurance violation.

 

[n.one.one]

Good update from @DirtyT3sla on 10.69.25.2. Some good, some bad, but in his style it seems to be a balanced overview of where things stand.

What I found most interesting was his mention that he's working on an overview video of FSD beta progress over the last two+ years. It's easy for folks to nitpick current FSD beta shortcomings but when you look back at the progress from October 2020 the progress is undeniable (who remembers bounding boxes?).

 

[JupiterMan]

What I found most interesting was his mention that he's working on an overview video of FSD beta progress over the last two+ years. It's easy for folks to nitpick current FSD beta shortcomings but when you look back at the progress from October 2020 the progress is undeniable (who remembers bounding boxes?).

I sometimes go back to the older versions on youtube for comparison and the improvement has been impressive, sometimes it seems everyone is caught up in the moment with the current release - will be interesting to see this overview and his thoughts!

 

[JulienW]

.... It's easy for folks to nitpick current FSD beta shortcomings but when you look back at the progress from October 2020 the progress is undeniable (who remembers bounding boxes?).

There has been undeniable progress BUT......we are no where near L4 Full Self Driving and if the progress continues at the same pace as the last 2 FULL YEARS we are still YEARS away (if ever) from L4. This doesn't even take into account how much slower progress will become once we reach the march of 0.999999s

 

[Durwood101]

There has been undeniable progress BUT......we are no where near L4 Full Self Driving and if the progress continues at the same pace as the last 2 FULL YEARS we are still YEARS away (if ever) from L4. This doesn't even take into account how much slower progress will become once we reach the march of 0.999999s

im afraid you are right on this. I used to be excited about FSD, but after the last few updates, its painfully obvious that although its a long way from where it was 2 years ago, its so far from being useful in a sense of turning on and not having disengagements. Wrong lane choices are the worst things that bother me using it. Once you realize that the logic just isnt there preparing for turns and being in the right lane, I find it hard to even use.

 

[AlanSubie4Life]

Wrong lane choices are the worst things that bother me using it. Once you realize that the logic just isnt there preparing for turns and being in the right lane, I find it hard to even use.

If only that were the worst issue. It would be easy to solve - just let the driver make the lane change requests. No lane changing unless the driver asks for it. And work out the behavior at turn lanes (could be automatic or requested). Definitely totally configurable.

Anyway, a long way to go. As has been said many times starting at the beginning of the public beta, it's been clear that this is a years long process (to get to usable L2), not a one or two year thing. And may never happen. Depends on whether a revolution takes place and whether the hardware has the ability to do the processing (seems questionable).

 

[JulienW]

im afraid you are right on this. I used to be excited about FSD, but after the last few updates, its painfully obvious that although its a long way from where it was 2 years ago, its so far from being useful in a sense of turning on and not having disengagements. Wrong lane choices are the worst things that bother me using it. Once you realize that the logic just isnt there preparing for turns and being in the right lane, I find it hard to even use.

Yea, I'm hopeful on a path to a L3 system and I will be happy but even that seems far away.

Analogy: >10 years ago Siri was released and I remember all the posts about how in just a few years we would be carrying on natural contextual conversations with our iPhones. Well we know know that Apple didn't win but nor did Google or Amazon. It is still years away.

I'm just afraid that we may reach that point with all Full Self Driving software that it seems to nearly stall and only small incremental improvements. While software can be very specialized it struggles with all the human nuances, unknowns and edge cases we take for granted and deal with using little forethought.

 

[gearchruncher]

Well we know know that Apple didn't win but nor did Google or Amazon. It is still years away.

Yeah, but now we have Chat GPT and it looks like microsoft might win the natural language war with their investment there.

And in parallel, we have Waymo and others that ARE actually at L4 in limited environments. They just use a very, very different sensor suite than Tesla. So maybe it's just that FSD actually does occur, but it requires a sensor that Tesla has steadfastly said isn't needed.

Remember, Tesla is highly tied to having shipped 2M+ sets of hardware before the SW is ready, while nobody else has done that, so we can't judge all feasibility of autonomy on that decision.

 

[stopcrazypp]

Yeah, but now we have Chat GPT and it looks like microsoft might win the natural language war with their investment there.

And in parallel, we have Waymo and others that ARE actually at L4 in limited environments. They just use a very, very different sensor suite than Tesla. So maybe it's just that FSD actually does occur, but it requires a sensor that Tesla has steadfastly said isn't needed.

Remember, Tesla is highly tied to having shipped 2M+ sets of hardware before the SW is ready, while nobody else has done that, so we can't judge all feasibility of autonomy on that decision.

I'm not entirely sure the perception side (sensors) is the blame however. It seems most of the dumb stuff FSD Beta does is all planning and execution related. That's pretty much all on the software algorithms (other than stuff they might be able to brute force with more processing power, which no doubt is inevitably coming in future iterations given past history).

 

[powertoold]

There has been undeniable progress BUT......we are no where near L4 Full Self Driving and if the progress continues at the same pace as the last 2 FULL YEARS we are still YEARS away (if ever) from L4. This doesn't even take into account how much slower progress will become once we reach the march of 0.999999s

Do you think there is not single route or locale in the entire USA that Tesla can polish or focus on to deliver a level 4 service?

 

[JulienW]

Do you think there is not single route or locale in the entire USA that Tesla can polish or focus on to deliver a level 4 service?

Of course they probably could....but if Tesla geofences and HD maps to offer L4 to a select few people in limited areas then they go against their entire mission of equal access to all.

As I have said I think at best we will get an L3 system and I will be happy with that. L3 for all on all roads beats L4 to a select few in limited places.

Still at the rate of progress we are likely several years away from an approvable L3 system in all or many states.

 

[legendsk]

Do you think there is not single route or locale in the entire USA that Tesla can polish or focus on to deliver a level 4 service?

Given their apparent current focus, not a single one, however they will likely have the perfect fart noise in . . . . . . . .two weeks?