Wiki MASTER THREAD: Actual FSD Beta downloads and experiences

[alee]

Isn't this more a freeway thing - rather than city FSD ?

Anyway - a large number of lanes merging into fewer is not unique to Holland tunnel.

ps : I'm not saying NYC doesn't have anything unique. Every city has some unique things (like Lombard Street or the Monorail). All I'm saying is - overall NYC isn't so unique in so many fundamental ways as to make FSD totally different for NYC compared to other cities.

There are indeed cities that are so different that FSD would have to be differently trained for them - like South Asian cities. NYC is not that.

I guess this is the nuance... there's enough little things that don't happen frequently elsewhere and are not known problems that can be solved for with the software today. NYC has their own, just as Boston, SF, DC etc. do as well. To me, I see these as city specific models. Cues it does not understand today are things like seeing hazard lights on a yellow taxi which is an acknowledgement that a fare is about to be picked up even if that car is still rolling, but is the local norm to go-around as soon as you see it. In many cases during rush hour, it's impossible to pull over for an emergency vehicle, so creating a 5th lane on a 4 lane city street for an emergency vehicle during heavy congestion is an dense urban thing. In comparison, less dense areas the norm is to get off the road completely, or pull over to allow traffic through.

FSD is designed for a highly normalized set of circumstances and traffic rules. As the complexity increases, it requires knowledge/talent it just doesn't have. That includes complex congestion, even though fundamentally it can drive fine in those scenarios. I firmly believe city localization will be necessary even if the car might stumble its way through it ok. If we could see the numbers, I'd bet the average small city might operate at a 70% confidence, but I'm sure we'd see the average confidence drop precipitously for dense city centers.

Different enough to matter anyway.

 

[powertoold]

When I'm driving in San Jose, I'm never wondering which road I should turn into. Many times when I'm driving in SF or LA (or Portland or etc.), I'm confused or thinking that many people would be confused about the road geometry. There are many roads with odd angles or obstructions or confusing signs or lane layouts, etc.

Just as one example in SF, there's a road that splits into one that goes above a tunnel and another that goes into the tunnel but both of them end up on the same road after the tunnel.

It's difficult to create "if then" statements to deal with all scenarios appropriately / smoothly / safely / efficiently.

 

[EVNow]

FSD is designed for a highly normalized set of circumstances and traffic rules.

Not true. As you see, for eg., the car frequently crosses over the yellow line to go around parked cars or even bikes.

Elon is not exactly a stickler for rules, unlike, say Waymo.

NYC has their own, just as Boston, SF, DC etc. do as well.

So, you are saying NYC isn't all that unique. Exactly what I'm saying.

There are a lot of small things that are the local norm. We are still at FD 101 - and you are talking about FSD 901. I doubt Robotaxis will ever get to that level anytime soon - for all we know manual driving will become prohibitively expensive and proper rule based driving becomes the norm everywhere since robotaxis will be the vast majority.

 

[tdiggity]

10.5.1

Took a road trip from Los Angeles to Santa Cruz and back and got 10.5 mid trip. Phantom braking on two lane roads from opposite traffic was greatly improved. Almost a non issue whereas with 10.4 it was unusable! Highway autopilot also did a lot better on 10.5. Felt pretty much like radar AP.

I don’t have much to add to actual city fsd. The hesitations killed the experience this week because I was always followed by someone and disengaged to not annoy others.

Highlight of the week was showing the M3 off to an uncle who had never been in any Tesla before. Rare to show off the acceleration to anyone around here nowadays, i miss seeing those moments of pure excitement. And of course, fsd was fun for all.

 

[alee]

So, you are saying NYC isn't all that unique. Exactly what I'm saying.

There are a lot of small things that are the local norm. We are still at FD 101 - and you are talking about FSD 901.

Same but different. We need localization or we will continue to have to disengage. It's not done until FSD takes a masters class!

I'm closely watching the Waymo testing here in NYC. Was very impressed with my experience with them in AZ, esp. the parking lot drop-off. Want to see if it feels just at home here, and how it balances being assertive with safety.

 

[EVNow]

Same but different. We need localization or we will continue to have to disengage.

I think we have a disengage for a variety of things even before we get to localization. There are probably a million edge case they have to train for - including local ones.

But do you still think NYC is somehow different from all other cities in such a unique way as to make it unfit for FSD - any more than any other city ? If so, I want to know why. For every one of your "unique" cases, someone can post a case that is unique to their location.

 

[nvx1977]

More importantly, the way Tesla FSD works, you can’t really “solve” for CA and not others. Despite what many here think, US is a lot more uniform than many other countries.

Yeah, this always gets me a little. You hear it from the YouTubers all the time. XXstate drivers are different than XXstate. XXstate driving is different than XXstate. In reality the differences are nuance at best and whether you are in a City, Suburban or Rural area will make a MUCH bigger difference that what state you are in.

I have thought about this quite a bit. I have also driven in many areas across the country. WRT to road uniformity, I agree for the most part. The New England area is the exception. Roads and intersections have been preserved as they were during horse-and-carriage days. Lots of really, really strange stuff here.

But road uniformity is not the only factor. Driving culture is quite different. The type of aggression I see in Boston is different than that of NYC is different than that of LA is different than that of Atlanta (which btw I consider to be the most polite of the cities listed here). There are things they have in common, but there are differences. For example, on an LA highway, you tend to have a wide spread of driving behaviors, whereas in Boston and NYC, things are more uniform. I see more slow people in the passing lane in LA than in Boston. Also the distribution of speed is much higher in LA (assuming no slowdown from traffic), whereas most drivers in MA are going around the same speed regardless of lane. In NYC, 2 people trying to change lanes into the same spot is way more common than in Boston or LA. In Atlanta, knowing which lane to be in during traffic makes more difference than in the other cities. This is mainly due to high use of exit-only lanes that appear between exits.

These differences impact how well AP/FSD performs in each locale. Most younger cities have roads in perfect grids and highway interchanges that don't have exit off-on ramp competition. In New England, you have lots more acute/obtuse angles to deal with and also intersections where 3 or more roads converge at angles. Lanes suddenly disappear, forcing you to merge. The main lane can suddenly turn into multiple lanes, and the straightest path can become a turn-only lane. In the residential suburbs, almost all roads are double yellow divided 2-lane roads that do lots of winding through the woods and converge/diverge with other roads in a Y or reverse Y intersection.

So in a nutshell, I think it's a valid observation that FSD performs better or worse depending on locale. I'm definitely not sold on CA having an advantage though. Plenty of CA people reporting poor FSD experiences, even in areas near Tesla HQ. And let's not forget that CA is huge with very different landscapes, congestion, and weather.

 

[alee]

I think we have a disengage for a variety of things even before we get to localization. There are probably a million edge case they have to train for - including local ones.

But do you still think NYC is somehow different from all other cities in such a unique way as to make it unfit for FSD - any more than any other city ? If so, I want to know why. For every one of your "unique" cases, someone can post a case that is unique to their location.

If cities weren't materially different beasts, a new driver would navigate them all with ease.

It's the combination of road dynamics, driving habits of others, casually marked signage, local driving norms (which do not necessarily follow any legal norms), and a certain level of driving acumen (e.g. accelerating to take advantage of an opening). None of these would be acceptable in most driving circumstances but without them, you're the hazard (e.g. the one person going 55 when everyone is doing 80... you may be following the law, but if you're creating a traffic situation, you're the problem).

 

[powertoold]

But do you still think NYC is somehow different from all other cities in such a unique way as to make it unfit for FSD - any more than any other city ? If so, I want to know why.

I don't think we're saying that fsd is unfit for any particular area.

I'm just saying that Tesla is trying to create driving rules and procedures that work well in the entire USA. They're also using training data from all across the USA. That's why we constantly get odd lane changing behavior, like waiting to the last second to change, or changing out of a lane we're supposed to be in, etc.

For example, overtaking a double parked car might be more appropriate in some locales / situations vs others, it's difficult to write one rule that would apply equally to SF vs LA vs rural area.

 

[Bladerskb]

@EVNow has been dropping some truth bombs lately and challenging all the major myths.

Wow I'm loving it.

 

[alee]

@EVNow has been dropping some truth bombs lately and challenging all the major myths.

Wow I'm loving it.

Even if s/he's wrong.

I don't disagree with some of @EVNow's points... but I hold firm that an experienced driver in San Jose, CA that has never driven in Boston would question their driving abilities in Boston. It's very different, and that also carries through to autonomy.

 

[Oil Freedom]

As you see, for eg., the car frequently crosses over the yellow line to go around parked cars or even bikes.

FSD Beta handles lane changes better for upcoming turns, but it still doesn't wait until after it has crossed straight through an intersection. That needs to be fixed since changing lanes while still in he intersection is illegal in many other places.

 

[Oil Freedom]

I don't disagree with some of @EVNow's points... but I hold firm that an experienced driver in San Jose, CA that has never driven in Boston would question their driving abilities in Boston. It's very different, and that also carries through to autonomy.

The effective rules and driving norms are very different between the two places!

It would be cool for FSD to use local driving "accents" to drive "better" based on the locally accepted standards. FSD could could then code switch as the car moved between San Jose and Boston and other locations with their own styles! It sure would help the car to work better with other traffic and blend in.

 

[Bladerskb]

I don't think we're saying that fsd is unfit for any particular area.

I'm just saying that Tesla is trying to create driving rules and procedures that work well in the entire USA. They're also using training data from all across the USA. That's why we constantly get odd lane changing behavior, like waiting to the last second to change, or changing out of a lane we're supposed to be in, etc.

For example, overtaking a double parked car might be more appropriate in some locales / situations vs others, it's difficult to write one rule that would apply equally to SF vs LA vs rural area.

Interesting that Mobileye has a patent on automatically deriving cultural driving behavior from their crowdsourced production fleet (1 million cars) that they can then use to adapt their driving behavior to the local norm. Tesla would have to do something similar.

"Techniques are disclosed to derive local driving behaviors from naturalistic data, which are then incorporated as guidance into the behavioral layer of an Automated Vehicle (AV) for adaptation to local traffic. Moreover, the local driving behaviors are implemented in the AV performance validation process. The techniques facilitate scaling of this process via automatic crowdsourced behavioral data aggregation from human-driven vehicles, as well as ADAS vehicles and autonomous vehicles, and map information. The disclosed techniques also enable the creation of a spatial-behavioral relational database that provides interfaces for efficiently querying geo-bounded local driving information, enabling customization of automated vehicle driving policies to local norms, and enabling traffic behavior analysis."

US20210101619A1 - Safe and scalable model for culturally sensitive driving by automated vehicles - Google Patents

Techniques are disclosed to derive local driving behaviors from naturalistic data, which are then incorporated as guidance into the behavioral layer of an Automated Vehicle (AV) for adaptation to local traffic. Moreover, the local driving behaviors are implemented in the AV performance...

patents.google.com

 

[EVNow]

If cities weren't materially different beasts, a new driver would navigate them all with ease.

The effective rules and driving norms are very different between the two places!

It would be cool for FSD to use local driving "accents" to drive "better" based on the locally accepted standards. FSD could could then code switch as the car moved between San Jose and Boston and other locations with their own styles! It sure would help the car to work better with other traffic and blend in.

But that was not what we have been talking about.

Discussion was about NYC in particular. It wasn’t just about local differences.

Let me state again - NYC is no different than other US cities. They all have some local differences and in that sense they all pose challenges to FSD.

Let us take an example. There are different accents / dialects in each city/region. But Alexa speaks the same way in all of US but they have probably trained with various accents. I don’t think Alexa understands PNW better than Boston area dialect ( or does it ?!).

In the bigger scheme of things i’m Looking at FSD working all over the world. Difference between Boston and NYC and LA pales in comparison to difference been Boston, London, Cairo and Mumbai. Just as the difference in dialects/accents (even if speaking English).

 

[EVNow]

@EVNow has been dropping some truth bombs lately and challenging all the major myths.

Wow I'm loving it.

You should have seen how much downvote I got talking about how Tesla handles minorities

 

[Oil Freedom]

But that was not what we have been talking about.

That's cool, but I'm talkin' about what I'm talkin' about.

 

[JulienW]

I have thought about this quite a bit. I have also driven in many areas across the country. WRT to road uniformity, I agree for the most part. The New England area is the exception. Roads and intersections have been preserved as they were during horse-and-carriage days. Lots of really, really strange stuff here.

But road uniformity is not the only factor. Driving culture is quite different. The type of aggression I see in Boston is different than that of NYC is different than that of LA is different than that of Atlanta (which btw I consider to be the most polite of the cities listed here). There are things they have in common, but there are differences. For example, on an LA highway, you tend to have a wide spread of driving behaviors, whereas in Boston and NYC, things are more uniform. I see more slow people in the passing lane in LA than in Boston. Also the distribution of speed is much higher in LA (assuming no slowdown from traffic), whereas most drivers in MA are going around the same speed regardless of lane. In NYC, 2 people trying to change lanes into the same spot is way more common than in Boston or LA. In Atlanta, knowing which lane to be in during traffic makes more difference than in the other cities. This is mainly due to high use of exit-only lanes that appear between exits.

These differences impact how well AP/FSD performs in each locale. Most younger cities have roads in perfect grids and highway interchanges that don't have exit off-on ramp competition. In New England, you have lots more acute/obtuse angles to deal with and also intersections where 3 or more roads converge at angles. Lanes suddenly disappear, forcing you to merge. The main lane can suddenly turn into multiple lanes, and the straightest path can become a turn-only lane. In the residential suburbs, almost all roads are double yellow divided 2-lane roads that do lots of winding through the woods and converge/diverge with other roads in a Y or reverse Y intersection.

So in a nutshell, I think it's a valid observation that FSD performs better or worse depending on locale. I'm definitely not sold on CA having an advantage though. Plenty of CA people reporting poor FSD experiences, even in areas near Tesla HQ. And let's not forget that CA is huge with very different landscapes, congestion, and weather.

And this proves my point. The minor differences you experiences/noticed were NOT showstoppers and you made it to your destinations. You were able to use your driving knowledge to drive in all theses places despite the small differences. The differences only stand out as large because you are human and that it what we remember. But they did NOT stop or confuse you beyond the ability to drive but only as a memory of the experience.

FSD is aspirational to Level 5 (or at least Level 4) which means if you and I can easily navigate (drive Level 5) there then FSD MUST be able to do so also no matter how many minor differences it encounters.

 

[EVNow]

That's cool, but I'm talkin' about what I'm talkin' about.

Well, i’m saying moving from NYC being uniquely different to all cities have some local flavor is good and truthful.

I’m looking forward to FSD being tolerant and understanding when other drivers make mistakes or are confused, which would be our local flavor (or so we want to believe)

 

[nvx1977]

And this proves my point. The minor differences you experiences/noticed were NOT showstoppers and you made it to your destinations. You were able to use your driving knowledge to drive in all theses places despite the small differences. The differences only stand out as large because you are human and that it what we remember. But they did NOT stop or confuse you beyond the ability to drive but only as a memory of the experience.

FSD is aspirational to Level 5 (or at least Level 4) which means if you and I can easily navigate (drive Level 5) there then FSD MUST be able to do so also no matter how many minor differences it encounters.

Well, I see humans failing here in New England all the time. Not like being stupid aggressive failing, but turning into the wrong lane because the intersection is confusing. I am an optimist when it comes to FSD hitting L4/5 some day, but I don't believe it can get there by just assessing the scene in a stateless manner. At some point, it needs to either remember how certain intersections work, or it can reference some data that tells it what's approaching.