FSD discussion

[Discoducky]

There are many reports (YouTube, here on TMC) of serious interventions that are needed even with the latest software. Some people have great experiences, and some are having not so great experiences. In order to get reliability needed for robotaxi, they'd need months of virtually no interventions in order to shake out the last sets of bugs. Look how long they have spent testing the semi or other car models before shipping. It's a long way between "hey, this works pretty well most of the time!" to a fully autonomous robotaxi

Basically I think the rate of improvements is not high enough to be close to shipping and it will take many more releases, perhaps a dozen or more, before it is close to being out of beta.

Could you point out a few of these reports to the thread please? Hopefully they have accompanying video for diagnosis.

My feelings are there just a few key things missing currently.

I have a UPL that is fairly complex, but straightforward where FSD has nailed it the last 50 or so times in varying traffic/weather conditions. Why? How can others, nearly identical, fail at a much higher rate? Or why do some FSD drivers have bad failures when others don't? I think it is due to the below and both are not trivial to solve, but are very tractable.

Obviously the Planner rests on top of the entire stack from OccNet/Flow/Obj net to Lane Connectivity to Lanes Net and Merge Net. The planner not only has to improve, but also has to be somewhat resilient to bad output from the stack. These are the first main issue is that the stack is outputting bad info to the planner and the planner is not doing the best it can with the bad data. What happens today is that the bad output is unknown due to lacking short term and long term memory. Neither the stack, nor the planner have enough short and long term memory.

2nd main issue is that not enough very fresh known good static ground truth is being fed into the stack. This is the main hurdle to FSD being solved. I know Tesla is working on this. The first issue, I'm less clear on.

Both of these issues, once solved, will allow FSD to be L4/L5.

 

[jeewee3000]

As someone who never experienced FSDbeta myself, I have a question regarding its ability to comply with the "rules of the road". Are these heuristics based (i.e. hand coded) or is some of this stuff already encompassed in one or more NN's?

For example the rule that one should drive on the right side of the road. FSDb obviously performs this correctly and picks the right lane on roads with just two (marked) lanes where traffic is permitted in both directions.

The reason I ask is because I'm trying to guesstimate how smoothly the transition to other parts of the world will work, hence how quickly I will be able to enjoy the joys of FSD(b) driving on this side of the pond.

For example:
in Belgium, and generally the EU, it is in most cases illegal to overtake the car in front of you on the right side. One must always overtake on the left side. In the US the opposite is true (from my limited experience driving on the West Coast): you can have 6 lanes in the same direction in LA for example and each lane occupant can drive as fast/slow as he likes, disregarding the fact that you might or might not overtake others/be overtaken by others in adjacent lanes.
In other words it boils down to: "give right of way when YOU lane change".

Every system has its up- and downsides, so don't read this as me being judgmental or anything. It just make me wonder how FSDbeta, used to LA traffic, will suddenly switch behaviors.

Seems to me in the current state such behaviours (i.e. of complying with local rules) must be hard coded and geofenced. Therefore I'm a bit pessimistic when I think of how long it will take for FSDb to be applicable in the EU.

And I'm not even talking about driving on the other side of the road.

Could someone with more knowledge on FSD/NN/machine learning than me provide their thoughts on this? Thanks!

 

[DarthPierce]

I'll know tomorrow how 10.69.3.1 does in snow. None of the previous versions slowed at all.

Quick report: Currently 17 degrees and 4 inches of snow on the ground.

Car does not increase follow distance, reduce speed, or otherwise acknowledge the weather.

It's actually even worse since it can't decide where the lanes are and turns abruptly. It also seems to want to avoid the parts of the lane where others have packed down the snow.

Strongly not recommended in snow at this point.

 

[ElectricIAC]

Quick report: Currently 17 degrees and 4 inches of snow on the ground.

Car does not increase follow distance, reduce speed, or otherwise acknowledge the weather.

It's actually even worse since it can't decide where the lanes are and turns abruptly. It also seems to want to avoid the parts of the lane where others have packed down the snow.

Strongly not recommended in snow at this point.

More proof that vision alone isn’t up to the task even if you had 20MP cameras and there was 15 of them placed around the car.

 

[Featsbeyond50]

More proof that vision alone isn’t up to the task even if you had 20MP cameras and there was 15 of them placed around the car.

Vision can be trained to recognize snow and water.

 

[ElectricIAC]

Vision can be trained to recognize snow and water.

Still can’t see through it.

 

[DarthPierce]

Still can’t see through it.

And neither can I. When it's really blizzarding no one should be driving.

In this case however all it shows is that they haven't done any training to work on this situation. The problem is far more about changes in traction than obscured vision.

 

[Bouboule]

Still can’t see through it.

I think, vision, traction control information, acceleration and deceleration based on speed and power should be enough. AWD might need to be mandatory for full FSD.

If I keep it a little on topic, I think the timing of adding multiple cameras will be the step change needed to move to real FSD. to meet redundancy for regulations and bad weather.

 

[SOULPEDL]

People of the North...

As most already know, reaction time on snow events is critical. And so far, I know I can beat Tesla FSD on reaction time all day long. Snow is an added risk factor, but not necessarily for people who don't have snow driving experience. With newbies, quite often the driving and reactions are incorrect (over-steering, braking, etc). For one, steering jerks would need to disappear as that's how we'd initiate a spin as a kid.

Careful with the weather folks, they're still working out good weather FSD, but some weather is good for a road test and initial lessons at least. Rain doesn't seem to be nearly as difficult. We drove up/down Mt Lemon 2 yrs ago on FSD, pouring rain, no problem even then. But snow driving - a steep learning curve it is. I don't expect this reliable for a few years.

However, if Tesla does snow driving analysis in Alberta Canada, many of us have found that they can't drive in snow because they don't have mountains in Calgary or Edmonton (all facts, ask anyone in BC). As a result, FSD would (Edit: In theory) already be an improvement in that region and could even help save lives for my friends in BC!

 

[Bet TSLA]

Just another random data point, but I updated from 69.2.4 to 69.3.1, and none of my old problems went away, and some new ones appeared. Despite my very low expectations, this is the first release that didn't manage to rise to the minimal level. Sure hope V11 is better, but I don't expect it to be.

Two new problems I hadn't seen before:
1. Took a left turn into a left turn lane when it should have turned into a go straight lane (there were two of each). Then came to a complete stop and jerked around unable to decide what to do.
2. With an upcoming right turn needed, it decided to avoid all the stopped traffic in the middle and left lanes (from the red light ahead), and scooted into the empty right lane despite the "lane close ahead" sign and a bunch of cones just past that.

I guess these are just manifestations of its lane selection continuing to be really bad, but I haven't seen these particular behaviors before.

 

[Featsbeyond50]

Still can’t see through it.

Neither can we.

 

[ElectricIAC]

Neither can we.

That’s what wipers are for.

 

[Featsbeyond50]

The most important driving skill for driving in snow or ice is the ability to drive to the tire shop and get proper tires put on the vehicle. As with humans, this will probably be the most important thing for FSD.

 

[Featsbeyond50]

That’s what wipers are for.

Haha, says the guy from Texas.

There are conditions where the snow heavily cakes onto the wipers and renders them nearly useless. Conditions where even your subzero wiper fluid freezes. There are also times it's coming down so heavy you have no visibility even if the snow isn't sticking too the windshield. Especially at night because the snow acts like fog and reflects your lights back at you. Snow may be one word but it's going to present Tesla with a vast variety of edge cases.

 

[ElectricIAC]

Our MX has heated wipers.

 

[Dreadnought]

Competition has its own challenges in unfavourable conditions...

https://twitter.com/x/status/1598901838690283520

 

[Bet TSLA]

I see that on TelsaFi there's a new version 2022.40.4.5 waiting to install on one vehicle that they identify as 10.69.3.2. So it looks like they're revving something that is still not V11. If he continues this much longer it will make a full year Elon's been talking about V11 as coming in "two weeks". It's going to have to be a serious improvement to be worth the wait.

Personally, I'm expecting nothing. Well, parity at best, and I hope no particularly ugly failure modes.

 

[scottf200]

I see that on TelsaFi there's a new version 2022.40.4.5 waiting to install on one vehicle that they identify as 10.69.3.2. So it looks like they're revving something that is still not V11. If he continues this much longer it will make a full year Elon's been talking about V11 as coming in "two weeks". It's going to have to be a serious improvement to be worth the wait.

Personally, I'm expecting nothing. Well, parity at best, and I hope no particularly ugly failure modes.

site: TeslaFi.com Firmware Tracker

 

[Discoducky]

Got a new build today, but it is a dot release, so <sad Mac sound>

 

[FSDtester#1]

Got a new build today, but it is a dot release, so <sad Mac sound>

Version 2022.44.25.20

Full Self Driving Beta Version 11.x

Popped up on teslafi today...