Elon: FSD Beta tweets

[Daniel in SD]

How many humans can safely make this turn in heavy traffic? This turn shouldn't even exist. I don't know why you guys are obsessing over a poorly designed intersection by public planners.

I actually asked Chuck to put up a camera to measure the human success rate. He claimed he was unaware of any collisions at that intersection. Someone needs to put up a camera so we can find the truth. Are any of Chuck's videos done in heavy traffic? I haven't seen that case.

 

[JupiterMan]

How many humans can safely make this turn in heavy traffic? This turn shouldn't even exist. I don't know why you guys are obsessing over a poorly designed intersection by public planners.

because in a world of RoboTaxis without steering wheels it needs to do something at such junctions, either to successfully navigate it or deploy an alternative safer route.

but agree junctions like this should not exist... but they do and there will be others no doubt.

 

[hw13]

because in a world of RoboTaxis without steering wheels it needs to do something at such junctions, either to successfully navigate it or deploy an alternative safer route.

but agree junctions like this should not exist... but they do and there will be others no doubt.

Even UPS plans their routes with right turns only for human drivers.

 

[AlanSubie4Life]

How many humans can safely make this turn in heavy traffic? This turn shouldn't even exist. I don't know why you guys are obsessing over a poorly designed intersection by public planners.

When I said "light traffic," I meant basically non-existent traffic, because in my experience with other turns the car will go in that scenario. It's when there is moderate traffic that it struggles to actually commit to a turn (and I don't know all the reasons for this).

I don’t understand why people think this turn is difficult for a human. You just have to cross three lanes of traffic, with excellent visibility (when positioned correctly of course, not even close to sticking into traffic, but not 5 feet further back than required), while negotiating any left-turning traffic. This is easy stuff.

It's harder if you want to do all the lanes and thread into traffic on the opposite side in a single move, but that's the advantage of the frogger strategy (which personally I would probably not employ unless it was really busy, since it's annoying, and exposes you to traffic traveling by at very close spacing and very high speed). [EDIT: Actually you can see from the video it is pretty wide - a lane PLUS extra space...I might just use that area...]

Here is the excellent visibility from the approximate place a car should stop (without sticking the nose into the lanes of travel); you can see it is very good visibility; no occlusions (other than the post which is not large enough to be a problem of course). The sidewalk width (and the median beyond it!) makes it not a problem at all to get good visibility for the turn - just go until you know you can see, then stop (another key capability that "unlocking" this turn will imply (perhaps) has been mastered - the car will now know when it can see - perhaps a tough problem)!:

Here's a specific example of the car not going to the right place (a human would never stop here; there was not sufficient visibility - just move 1-2 feet forward!), and then further having a very short memory when vision is occluded by turning traffic. These are not mistakes any reasonably competent human would make:

 

[Daniel in SD]

I learned to drive in Los Angeles so I know something about heavy traffic.
To me heavy traffic is when you have to navigate between cars that are basically stopped in traffic. Usually one kind soul will stop and leave you a gap and then the person in the next lane will also stop. You then proceed to go having made eye contact with the two drivers. Then you must wait until a car in the third lane stops to let you through. To me this seems like a near impossible problem for AI. The traffic that I've seen Chuck attempt the turn in doesn't look like it requires AI at all... just assume that all cars will continue traveling at the same speed and you can solve 100% of his attempts.

 

[diplomat33]

When I said "light traffic," I meant basically non-existent traffic, because in my experience with other turns the car will go in that scenario. It's when there is moderate traffic that it struggles to actually commit to a turn (and I don't know all the reasons for this).

I don’t understand why people think this turn is difficult for a human. You just have to cross three lanes of traffic, with excellent visibility (when positioned correctly of course, not even close to sticking into traffic, but not 5 feet further back than required), while negotiating any left turning traffic. This is easy stuff.

It's harder if you want to do all the lanes and thread into traffic on the opposite side in a single move, but that's the advantage of the frogger strategy (which personally I would probably not employ unless it was really busy, since it's annoying, and exposes you to traffic traveling by at very close spacing and very high speed).

Here is the excellent visibility from the approximate place a car should stop (without sticking the nose into the lanes of travel); you can see it is very good visibility; no occlusions (other than the post which is not large enough to be a problem of course). The sidewalk width (and the median beyond it!) makes it not a problem at all to get good visibility for the turn:

View attachment 837117

IMO, the real issue for Tesla is that the side cameras might not have enough range to detect high speed vehicles. According to the AP2 graphic, the B pillar cameras only have a range of about 80 m. That is far less than the range humans have. So maybe FSD Beta is detecting a vehicle just outside the range, at say 90 m, but with not a lot of confidence so it hesitates to make the turn. I calculated that 80 m at 60 mph is about 3 s of reaction time. So if FSD Beta is detecting vehicles at about 3 seconds out, with not a lot of confidence, it could be hesitating to commit to the turn. That's my guess as to why FSD Beta has struggled with Chuck's ULT.

 

[AlanSubie4Life]

IMO, the real issue for Tesla is that the side cameras might not have enough range to detect high speed vehicles. According to the AP2 graphic, the B pillar cameras only have a range of about 80 m. That is far less than the range humans have. So maybe FSD Beta is detecting a vehicle just outside the range, at say 90 m, but with not a lot of confidence so it hesitates to make the turn. I calculated that 80 m at 60 mph is about 3 s of reaction time. So if FSD Beta is detecting vehicles at about 3 seconds out, with not a lot of confidence, it could be hesitating to commit to the turn. That's my guess as to why FSD Beta has struggled with Chuck's ULT.

Yeah, maybe. In the end the cameras will obviously need more than that range, and they probably do already. At 10:00 in the video above you can see it has visibility out to 100m. But then the vehicle is immediately tripped up by occlusion - they really need to reliably identify all the important objects in the scene, immediately, and remember them, just like a human! Another capability unlock which will be enabled by mastering this turn.

It wasn't really even occlusion. I guess you could say it got distracted and target fixated on the turning vehicle. It simply didn't see (if we believe the visualizations, which is questionable) this car at less than 100m range in this case (this is at 10:09, 9 seconds after detecting a vehicle at similar range). It finally shows on the visualization at 60m, at 10:11 (perhaps delayed by occlusion):

Anyway, looking forward to all the detection range, discrimination & consistency improvements in 10.13! The march of 9s has begun!

 

[momo3605]

The march of 9s has begun!

...for one turn Wake me up when the march of 9s begins for the other trillions of scenarios

 

[AlanSubie4Life]

...for one turn Wake me up when the march of 9s begins for the other trillions of scenarios

To be fair, actually improving on this turn in a "scalable" way should mean general capability improvements. Hopefully they'll be able to get that first 9 for this specific case! TBD. Two 9s would be an incredible quantum leap forward (in context). Might even make FSD dangerously good.

 

[Supcom]

because in a world of RoboTaxis without steering wheels it needs to do something at such junctions, either to successfully navigate it or deploy an alternative safer route.

but agree junctions like this should not exist... but they do and there will be others no doubt.

There are many similar intersection near me. They are actually quite common.

 

[mtndrew1]

I learned to drive in Los Angeles so I know something about heavy traffic.
To me heavy traffic is when you have to navigate between cars that are basically stopped in traffic. Usually one kind soul will stop and leave you a gap and then the person in the next lane will also stop. You then proceed to go having made eye contact with the two drivers. Then you must wait until a car in the third lane stops to let you through. To me this seems like a near impossible problem for AI. The traffic that I've seen Chuck attempt the turn in doesn't look like it requires AI at all... just assume that all cars will continue traveling at the same speed and you can solve 100% of his attempts.

Eleven months in and my car with FSD Beta cannot drive 3.8 miles down Figueroa from USC to 2nd street without about five disengagements or interventions.

It’s a straight, well-marked road with no turns or merges. When I hear Elon’s enthusiasm during events like last night’s investor presentation I often think of the inability to drive straight down Figueroa.

 

[DanCar]

Eleven months in and my car with FSD Beta cannot drive 3.8 miles down Figueroa from USC to 2nd street without about five disengagements or interventions.

What do you think of the Mar's guy, Omar, regularly stating 50+ mile drives with no interventions?

 

[mtndrew1]

What do you think of the Mar's guy, Omar, regularly stating 50+ mile drives with no interventions?

I think he desperately wants to be in the inner circle of a certain billionaire and either lies or chooses his routes very carefully.

In the last 11 months I don’t know that I’ve ever gone more than a mile or two without a disengagement or intervention.

 

[2101Guy]

how many of you are still believing ANYTHING Elon says about FSD these days. ANYTHING

 

[mtndrew1]

how many of you are still believing ANYTHING Elon says about FSD these days. ANYTHING

I think they might actually figure it out at some point. I also think that point will be long after I’ve sold my Tesla.

 

[AlanSubie4Life]

I think he desperately wants to be in the inner circle of a certain billionaire

Many do! It's mysterious.

There are inner circles, and then there are inner circles (Elon does respond to this account, too). I guess this is not quite an Elon FSD Beta tweet...sorry to go off topic.

https://twitter.com/x/status/1555043296015269889

 

[GlmnAlyAirCar]

I think he desperately wants to be in the inner circle of a certain billionaire and either lies or chooses his routes very carefully.

In the last 11 months I don’t know that I’ve ever gone more than a mile or two without a disengagement or intervention.

Are you serious? I can frequently drive 30+ miles on mixed roads with no or occasionally one intervention. Frequently with no disconnects. Of course I have been driving these routes on FSD since last October. I probably have 20,000 FSD miles, maybe more. These routes have improved significantly since 10.3. Perhaps its because I’ve been feeding all of this data to the mothership.

i suspect many of those who can‘t go more than a mile without intervention really aren’t that serious about testing.

 

[DarkForest]

I think he desperately wants to be in the inner circle of a certain billionaire and either lies or chooses his routes very carefully.

In the last 11 months I don’t know that I’ve ever gone more than a mile or two without a disengagement or intervention.

I think it depends on your location. In the city I live in, there are certain routes where I could go a few miles without needing to intervene, and there are certain areas where I have to intervene pretty frequently. Also, I just got back from a road trip to South Padre Island, and I barely had to intervene at all while I was in the area. It was nice.

Two months ago, I refused to believe anyone could have an intervention-free drive of any length while using the beta. Now I know it's possible, but it doesn't necessarily mean the beta is good to go. It's still pure garbage in some situations.... but also gold in others.

 

[mtndrew1]

Are you serious? I can frequently drive 30+ miles on mixed roads with no or occasionally one intervention. Frequently with no disconnects. Of course I have been driving these routes on FSD since last October. I probably have 20,000 FSD miles, maybe more. These routes have improved significantly since 10.3. Perhaps its because I’ve been feeding all of this data to the mothership.

i suspect many of those who can‘t go more than a mile without intervention really aren’t that serious about testing.

I’ve driven about 15k miles in total in the last year, maybe 90% on AP/FSD.

The 3.8 mile stretch I referenced in an earlier post is something I’ve done basically every workday since September of 2021 when I got into the program. FSD has never been able to complete that straight shot down a major thoroughfare in Los Angeles that probably sees 5k Teslas a day. To this day it cannot make it straight through two of the intersections on green lights on its own. At one point it gets all the way over into a turn-only lane onto the 110 freeway instead of proceeding through the green light ahead of it.

The car cannot go from my office parking structure down a well-marked road for 500 feet and make a left with a stop sign onto another well-marked street and proceed another 500 feet. That little situation, without traffic, involves three full disengagements. I’ve been doing that for 11 months too.

Last week my car uploaded nearly 50GB of video. FSD is just…it just doesn’t work. It’s a novelty.

 

[diplomat33]

Disengagements will depend a lot on your route and traffic. If the route happens to include scenarios that FSD Beta cannot handle well, then you will get disengagements. And if that scenario happens to be 1 mile into the drive, then you might say that you can't go 1 mile without a disengagement and be telling the truth. Likewise, if your route happens to only have scenarios that FSD Beta can handle well, then you might indeed be able to go 30+ miles with no disengagement. FSD Beta does do well with simple lane keeping so routes that are mostly staying in the lane on a long road with no turns at intersections will likely have less disengagements, especially if traffic is low. Routes that involve unprotected turns or navigating narrow streets with no lane markings, may have more disengagements since those are more diffcult scenarios for FSD Beta.

Also, disengagements are very depending on traffic. If you happen to drive when there is no traffic, you might do the entire route with no disengagements. The exact same route might have frequent disengagements if traffic is busy and FSD beta reacts poorly to other cars. That's why you need to provide context to disengagements. Disengagements are also dependent on the driver as some drivers might get more nervous and disengage more often while others might be more adventurous and disengage less to see what happens.