FSD Beta 10.13

[jabloomf1230]

We Should Abolish the Left Turn, Science Suggests

It does make a lot of sense.

www.popularmechanics.com

 

[Daniel in SD]

I think to avoid that left turn he would have to go straight across the road and make three rights on the grass... doesn't seem much easier or safer!
He could go right and make a u-turn but a u-turn is a left turn! Chuck also states that he's aware of multiple collisions at the signalized intersection and zero collisions at his infamous left.

Maybe there's a spot further down where he could do six right turns...

 

[jabloomf1230]

I think to avoid that left turn he would have to go straight across the road and make three rights on the grass... doesn't seem much easier or safer!
He could go right and make a u-turn but a u-turn is a left turn! Chuck also states that he's aware of multiple collisions at the signalized intersection and zero collisions at his infamous left.

Maybe there's a spot further down where he could do six right turns...

The "most dangerous" intersections are usually high volume. Considering that in all of Chuck's videos there are only a few other cars making the ULT, it's probably why there have been no accidents there... yet.

 

[g1000x]

The latest from Elon.

View attachment 834213

Cool so probably another month of waiting

 

[AlanSubie4Life]

If they nail this it sure seems there must be some improvement in the easier maneuvers. As I keep bitching, I'm constantly overriding it failing on protected right turns :-(

Yep, success would mean so many improvements:

1) Consistently positioning perfectly prior to turn; no angling and blocking turning traffic into the side street.
2) Rapid and assertive progression across lanes of traffic (not 7 seconds like we saw so far!).
3) Ability to understand using the center reservation with proper positioning, and able to quickly and accurately occupy it.
4) No turning of the wheel before turning.
5) Ability to fit into reasonable gaps in traffic safely and assertively.
6) Very good perception and ability to accurately gauge vehicle speed at substantial distance using pillar camera and repeater cameras. (Otherwise it will require gaps that are too large.)
7) Ability to see and anticipate turn signals for traffic in far lanes at substantial distance (>4 seconds)

These substantial skills will definitely translate to other scenarios. I hope they all pan out!


I feel like they’ll also need to remove a lot of the hobbling they have currently implemented, to make this work, which will be a good thing. (For example they haven’t been able to make it move aggressively because then someone might not intervene in time - now it will just have to be right, and users will need to be paying attention - maybe they’ll add some audio feedback to alert people of otherwise reduce risk? Another possibly weird option is to move slowly for the first ~300ms, then rapidly ramp up. )

 

[Dewg]

Yep, success would mean so many improvements:

1) Consistently positioning perfectly prior to turn; no angling and blocking turning traffic into the side street.
2) Rapid and assertive progression across lanes of traffic (not 7 seconds like we saw so far!).
3) Ability to understand using the center reservation with proper positioning, and able to quickly and accurately occupy it.
4) No turning of the wheel before turning.
5) Ability to fit into reasonable gaps in traffic safely and assertively.
6) Very good perception and ability to accurately gauge vehicle speed at substantial distance using pillar camera and repeater cameras. (Otherwise it will require gaps that are too large.)
7) Ability to see and anticipate turn signals for traffic in far lanes at substantial distance (>4 seconds)

These substantial skills will definitely translate to other scenarios. I hope they all pan out!


I feel like they’ll also need to remove a lot of the hobbling they have currently implemented, to make this work, which will be a good thing. (For example they haven’t been able to make it move aggressively because then someone might not intervene in time - now it will just have to be right, and users will need to be paying attention - maybe they’ll add some audio feedback to alert people of otherwise reduce risk? Another possibly weird option is to move slowly for the first ~300ms, then rapidly ramp up. )

An excellent set of goals! The turn signal one is hard though - SOOOO many people turn their signals on and just forget them. If the Tesla sees a car with it's signals on and waits, assuming it will change lanes into the lane it wants to occupy, but it never changes lanes - does that count as a failure?

 

[AlanSubie4Life]

If the Tesla sees a car with it's signals on and waits, assuming it will change lanes into the lane it wants to occupy, but it never changes lanes - does that count as a failure?

Actually the turn signal perception is for people turning left onto the Tesla’s street. Very important, less important the faster the Tesla moves to the median.

I would not expect the car to go if there is a car oncoming and close in the middle lane (dangerous) regardless of signal. Far lane seems safe enough (if the vehicle in the far lane is close enough and not signaling I don’t see that there is any risk, and if it is further away it is not a risk). So detecting signaling for that go/no-go decision is less important. I would not be paying that much attention to that as a human; would just go when two lanes are clear and the third does not pose any obvious danger. If paying attention to turn signals maybe you could be slightly less conservative with cars in the middle lane but caution would still be advised (basically would time it so that that middle lane vehicle would come to zero “following” distance during the Tesla acceleration phase to full speed in the travel lanes, rather than ending up at 2-3 seconds behind as would be required for vehicles in the close lane). All easy to calculate.

In general I assume people are turning or changing lanes (or going straight if that would result in a collision with me) when they are not signaling, and they are going straight (or turning if that would result in a collision) when signaling, and plan accordingly.

So in this case, the worst case would be a person who was signaling to turn left (falsely) would mean a potential delay on initiating the turn if it coincided with an opening in the three near lanes. And any traffic in the nearest of the far three lanes would be assumed to be turning left regardless of signal. So really maybe the signal detection doesn’t matter that much.

 

[Dewg]

Actually the turn signal perception is for people turning left onto the Tesla’s street. Very important, less important the faster the Tesla moves to the median.

I would not expect the car to go if there is a car oncoming and close in the middle lane (dangerous) regardless of signal. Far lane seems safe enough (if the vehicle in the far lane is close enough and not signaling I don’t see that there is any risk, and if it is further away it is not a risk). So detecting signaling for that go/no-go decision is less important. I would not be paying that much attention to that as a human; would just go when two lanes are clear and the third does not pose any obvious danger. If paying attention to turn signals maybe you could be slightly less conservative with cars in the middle lane but caution would still be advised (basically would time it do that that would come to zero following distance during the acceleration phase, rather than ending up at 2-3 seconds behind as would be required for vehicles in the close lane).

Also thinking about unprotected right turns too - if there is a car in the middle/left lane with their right-turn signal on, indicating they may move into the right lane, does the Tesla wait before turning right? If they don't change lanes, causing the Tesla to sit there with an empty lane - does that count as a failure?

 

[AlanSubie4Life]

too - if there is a car in the middle/left lane with their right-turn signal on, indicating they may move into the right lane, does the Tesla wait before turning right? If they don't change lanes, causing the Tesla to sit there with an empty lane - does that count as a failure?

Yes, and no (assuming the vehicle is close enough to qualify as a hazard). I’d wait regardless of signal. Too risky. Just wait for an adequate opening! Which lane is occupied just affects how large a gap you maintain post turn. With the vehicle in the far lane (two lane case) you should ensure that there will be zero gap after you get up to speed. And when the vehicle is in the near lane (your destination lane) you have to ensure the vehicle does not get closer than 2-3 seconds. So it means the vehicles in the far lane can be closer by 2-3 seconds (assuming equal speeds, not much greater than your target speed, etc.). But if they change into your lane unexpectedly there would still be no collision (just zero following distance at the moment you reach your target speed (which is equal to that vehicle’s speed)); they would be neck to butt with you in the adjacent lane (or same lane), basically.

So as far as the turn signal is concerned, I guess if it is on, maybe as a courtesy you budget another couple seconds for vehicles in that far lane, to accommodate their indicated intention (and make sure they don’t have zero following distance were they to follow through). They basically are treated the same as a vehicle in the near lane. But even if you didn’t do that there would be no collision risk.

Again, all trivial for the car to calculate. We have to do it at intuitively!

 

[JHCCAZ]

...
3) Ability to understand using the center reservation with proper positioning, and able to quickly and accurately occupy it.
...

All of those numbered points are good, but I'm isolating this one as I've repeatedly mentioned this and it's not legal everywhere. AZ traffic school teaches that you may not use the median area as a stopping point aka "reservation" for these left turns. There are exceptions with very wide medians and special road markings, even stop signs in those special cases. But this proscription is widely ignored.

Medians, in most divided roads, are not wide enough to accommodate a vehicle waiting in that area without interfering somewhat with the high-speed left lane traffic. They only get physically wide enough when there is at least one left-turn bay for the main road lanes - and then it's a real mess if someone is actually occupying said left turn bay, waiting to turn. In the latter case, the visibility becomes is heavily blocked by any car using the median as a turn-in reservation spot, and the right-of-way rules for the various turning cars become confused.

There's also a problem in such roads, even without the presence of a turning-in vehicle, whenever both of the opposite-direction left-turn bays are occupied. This commonly results in unsafe, vision-blocked ULTs in both directions. Again, throwing in a third or fourth car, stopped in the middle of a turn-in action, makes for a chaotic and stressful interaction.

I completely understand that these situation is very common, here in AZ and everywhere else, regardless of what the law might say. Common because it's cumbersome and expensive to engineer a safer infrastructure for multiple left-turning vehicles in and out of high speed traffic. Installing a traffic light every block or two is not acceptable outside of dense urban environments, and of course multi-level entrances and exits are far too expensive and ugly for these suburban commuting routes. Many of these roads were originally built when traffic was much lighter, and have evolved into semi-highways replete with intersections like this.

So why is it true, as Chuck has noted, that there aren't so many collisions at these kind of secondary intersections? I'm not completely sure, but I think has to do with the instinctive alertness required when executing the turns.

And then why are there more commonly accidents at the traffic-light intersections? I'd say it's because in those intersections, there is a higher density of turning activity, and also a false sense of security from the traffic light control. I think that many of those collisions happen because inattentive drivers in the through-traffic lanes either ignore or try to beat the light, while drivers executing the "protected" turning or crossing are simply taking their legal right-of-way and not yielding to the high-speed main lanes. In the unprotected intersections, it's clear that crossing/turning drivers must yield or die. The unprotected case is indeed risky, but experienced humans are able to accomplish it almost all the time.

Noting these points, we could conclude that never-inattentive AVs will solve the most common problems at protected intersections. But they may struggle more than humans do, at difficult unprotected intersections, because they will be more risk-averse. If properly programmed, they won't proceed with a questionable unprotected turn just because they've been waiting a long time and/or don't want to frustrate drivers behind them. And this defensive behavior will be seen as annoying and non-human-like, even if humans would be advised to do the same. (Here I'm speaking of a finished-product AV, not one that's still learning how to actually decide and then execute the maneuver.)

 

[AlanSubie4Life]

AZ traffic school teaches you that you may not use the median area as a stopping point aka "reservation" for these left turns.

It should only do this when it is legal. Personally I avoid using the frogger technique whenever possible. But in this specific case I might well use it.

Double yellow lines (specifically center reservations surrounded by double yellow lines) of course cannot be crossed/entered. They are treated as solid obstacles; they may as well have curbs and shrubbery.

 

[AlanSubie4Life]

But they may struggle more than humans do, at difficult unprotected intersections, because they will be more risk-averse.

This doesn’t seem like that big a problem. They know exactly how fast all traffic is going and know what is obscured (what they can’t see) and make appropriate worst-case assumptions. And they can calculate how long it will take to make a maneuver, and they know how much margin to leave. So it seems easy to know when it is safe, and it would not be frustrating at all. Completely eliminates the “darn, that was a good opportunity” phenomenon.

I am not sure why self-driving companies avoid certain left turns, though (except to save time in some cases).

 

[JHCCAZ]

Also thinking about unprotected right turns too - if there is a car in the middle/left lane with their right-turn signal on, indicating they may move into the right lane, does the Tesla wait before turning right? If they don't change lanes, causing the Tesla to sit there with an empty lane - does that count as a failure?

It actually shouldn't matter whether there's a turn signal indicating impending occupancy of an empty lane. Though common, it's incorrect to turn into the empty lane when someone is coming in the adjacent lane. First, because you can't know whether they're in the process of checking their blind spot and just about to make a lane change; you don't know their intent. Second, because they don't know your intent, and you shouldn't put them in the position of reacting to your turn-in. If just a little different steering wheel input would cause you to cross the empty lane and end up colliding with them as they pass by, then they also can't read your intended path with enough certainty.

These discussions will of course never end, because one man's failure to act is another man's defensive-driving pause. But that's my view of turning into unoccupied lanes. My dad taught me not to do it, and Tesla shouldn't do it.

 

[AlanSubie4Life]

But that's my view of turning into unoccupied lanes. My dad taught me not to do it, and Tesla shouldn't do it.

Yes the only thing that is different is the margin you leave. You might not reduce that margin if there were a turn signal (as a courtesy).

But in both cases (lane change occurs without signal, or going in spite of a signal being shown) there would be zero collision risk, even in the event of reducing the margin, with no oncoming vehicle speed changes required.

 

[JHCCAZ]

This doesn’t seem like that big a problem. They know exactly how fast all traffic is going and know what is obscured (what they can’t see) and make appropriate worst-case assumptions. And they can calculate how long it will take to make a maneuver, and they know how much margin to leave. So it seems easy to know when it is safe, and it would not be frustrating at all. Completely eliminates the “darn, that was a good opportunity” phenomenon.

I am not sure why self-driving companies avoid certain left turns, though (except to save time in some cases).

I think that, even considering perfect physics calculations, there is a safety and risk reduction margin that needs to be factored into the calculation. Humans commonly may choose to shave down these margins when traffic is heavy and there's pressure of time and other drivers' impatience. I'm saying the AV shouldn't cut those margins anytime humans are involved on either side of the maneuver. Think about how many small risks we take in heavy traffic. If your programming and selling an AV you just can't do it, at least not yet.

Sometimes you see discussions here where people talk about the AV being able to calculate speed and acceleration better than humans, so they'll be able to shoot the gaps with high precision. But you can't scare other drivers just because you know you you'd have an inch to spare.

 

[KArnold]

They know exactly how fast all traffic is going and know what is obscured (what they can’t see) and make appropriate worst-case assumptions.

I'm not so sure, at least in the current FSDB.

Leaving my neighborhood there is a stop sign entering a 4-lane (2 each way) 40 MPH (mostly driven @ 50) road. Looking left it is seriously obstructed by vegetation. Driving manually and leaning forward there is about a 2-foot section where you can stop to see incoming traffic to the left without impacting cross traffic.

FSDB stops at the stop line and creeps, as expected. And keeps creeping until the pillar cams can see, which is already too far in the road if a car is coming. It's a 100% fall if there is any traffic at all to the left.

On this UPL setup as it is with vegetation and pillar cams, I'm not sure FSD could ever make this turn safely. Honestly I'm thinking about trimming those trees myself just for FSD.

But my point is, I'm not sure they can always know when traffic is obscured. At least, not yet.

 

[AlanSubie4Life]

Sometimes you see discussions here where people talk about the AV being able to calculate speed and acceleration better than humans, so they'll be able to shoot the gaps with high precision. But you can't scare other drivers just because you know you you'd have an inch to spare.

Sure. Definitely not talking about that. It’s required that you not require any other drive to react (except perhaps very slightly ease off out of an abundance of caution (defensive driving!)).

there is a safety and risk reduction margin that needs to be factored into the calculation.

Yes, hence calculating the appropriate margin.

 

[AlanSubie4Life]

I'm not so sure, at least in the current FSDB.

No, was not talking about that. I have no idea what the capabilities are of FSDb. Talking about a system which can exactly determine all vehicle speeds and distances (and by extension acceleration of course).

And of course a system that knows what it cannot see. This is (in theory) easy for a human. Not sure how it is done with a visual perception system.

 

[Wrkent100]

It’s hard to imagine they will ever conquer fsd when they can’t even correct phantom breaking…

 

[n.one.one]

Time will tell...