FSD Beta 10.13

[jabloomf1230]

Did people see this a few days ago? Apparently from an internal 10.13 tester. Seems to be planning pretty far out, which feels like a good sign. Fingers crossed that will help with lane selection for upcoming turns

https://twitter.com/x/status/1550924652318339072

Traversing a roundabout safely is a far greater need IMO than solving Chuck's ULT. This intersection is such an outlier that it will be a miracle, if the performance of FSD beta at that intersection is even modestly improved. That's what I'm expecting out of 10.13.

 

[AlanSubie4Life]

This intersection is such an outlier that it will be a miracle,

This sort of intersection bears many similarities to extremely common scenarios. The difficulty and hazard of this one is particularly high, but the steps needed to execute such a turn successfully can and must be applied to much simpler scenarios. How many times does one turn onto a road with two lanes of traffic each direction, with a center median? It’s really not uncommon. They can be quite dangerous and difficult to accomplish. It’s fine if you just want to not solve these situations, but that is not the stated goal.

My impression with this turn is the difficulty is not the visibility or the perception. It just was not employing a working strategy (amongst other planning issues and delays)! This issue extends to many other aspects of beta as well. The whole understanding of how to come up with a working strategy for arbitrary situations is lacking. There are a nearly unlimited combination of scenarios. I don’t think the answer is to program and fit each one manually.

 

[jeremymc7]

Traversing a roundabout safely is a far greater need IMO than solving Chuck's ULT. This intersection is such an outlier that it will be a miracle, if the performance of FSD beta at that intersection is even modestly improved. That's what I'm expecting out of 10.13.

In my area of Los Angeles there are hardly any roundabouts. But there is one blocks from me we all call the circle of death - due to how many accidents there are and how many times over the years they’ve tried to redesign the function, but without making any changes to the streets themselves.

There are a huge number of complicated left turns. Many blind or practically blind. The existing camera setup has little to no visibility. As I keep mentioning side view bumper cameras would provide proper visibility. Having had them on my old Audi A7 and they were very helpful is seeing what’s coming without pulling out into the lane.

 

[Dan D.]

Traversing a roundabout safely is a far greater need IMO than solving Chuck's ULT. This intersection is such an outlier that it will be a miracle, if the performance of FSD beta at that intersection is even modestly improved. That's what I'm expecting out of 10.13.

They are in a bit of a bind if they can't "solve" the turn at Chuck's ULT, though I really don't see why they can't solve it. Many times it just sits in the intersection when there is a clear gap, and other times it tries to go slowly in front of traffic. It really should not be so difficult, but seems oddly unable to commit to the moment.

I think their best option at this point is to hardcode the heck out of the turn so the car doesn't have to use much perception to steer the car. Then wait for huge clear gaps in traffic - which do occur every minute or so - and just go. Don't fuss with trying to go between cars, the system isn't ready.

They can get around the fact they have hardcoded the turn with a funny little graphic on the screen that says "Welcome to Chuck's Turn", which will be so funny that we'll forgive that the turn can't be done anywhere else. At least then they can release 10.13.

 

[tmoz]

https://twitter.com/x/status/1552759374644449280

Still working on it! No reason to release an update with known bugs.

I'd buy that guy a dinner if I were Chuck, but a jar of honey is good too! Now the dude knows Chuck is watching.

 

[tmoz]

This sort of intersection bears many similarities to extremely common scenarios. The difficulty and hazard of this one is particularly high, but the steps needed to execute such a turn successfully can and must be applied to much simpler scenarios. How many times does one turn onto a road with two lanes of traffic each direction, with a center median? It’s really not uncommon. They can be quite dangerous and difficult to accomplish. It’s fine if you just want to not solve these situations, but that is not the stated goal.

My impression with this turn is the difficulty is not the visibility or the perception. It just was not employing a working strategy (amongst other planning issues and delays)! This issue extends to many other aspects of beta as well. The whole understanding of how to come up with a working strategy for arbitrary situations is lacking. There are a nearly unlimited combination of scenarios. I don’t think the answer is to program and fit each one manually.

Could you please remind me (us) of the bet you refer to? I don't remember what the conditions were, nor the prize.

 

[AlanSubie4Life]

Could you please remind me (us) of the bet you refer to? I

Bet is for a beer, with @Daniel in SD.

I bet that FSD 10.13 will be successful on any of Chuck’s unprotected lefts (there are multiple different UPLs we attempts; we just take the combined total) less than 90% of the time.

History, embedded link here:

FSD Beta 10.13

Had an easy five-second gap. Six second if you neglect an opportunity with traffic in a different lane. Has to go in this situation. Went with large truck on a five-second gap. If that is safe, the others are safe. Can’t have it both ways. Feeling very good about my bet if that’s the...

teslamotorsclub.com

Since there is tendency to call failures successes, we have to define it. Copied from elsewhere:

Success definition:
1) With traffic that the car has to respond to. Can't be an empty road. (Successes where it goes with no traffic at all don't count, even though that would be an improvement from 10.12.) Any traffic conditions are ok, we get what we get.
2) Car has to go at the first opportunity when a decent human driver would go, when there is a nice relaxed opening. (This is the hardest thing to define but I'll know it when I see it.) (This will be contentious as was seen a few pages back here, where only 1 out of 4 attempts were successful per this metric; a 5-second gap was clearly plenty, and waiting for traffic in the farthest lane to pass is not legitimate.)
3) The car cannot cause any other car to slow down or react at any point in the maneuver (by poking out, moving forward at the wrong time), etc. Including if it stops halfway through in the median - it cannot result in any car slowing or yielding to it or anything like that.
4) The car can't do anything incorrect, like stopping in the first lanes of traffic (even if there is no one coming) to wait for traffic from the other direction.
5) Interventions including accelerator application count as failures. Each discrete intervention before the car starts the crossing counts as a separate attempt and failure. (E.g. pushing the accelerator to creep forward is a failure, then a new attempt immediately begins)

 

[Matias]

I think their best option at this point is to hardcode the heck out of the turn so the car doesn't have to use much perception to steer the car.

That would be cheating and completely useless solution. And Chuck has said, that if this succeeds, he will move to some other UPL to check, if it is general solution.

 

[aronth5]

That would be cheating and completely useless solution. And Chuck has said, that if this succeeds, he will move to some other UPL to check, if it is general solution.

Agree.
Near where my brother lives in Houston, Tx there are similar streets with multiple medians that require the car to handle similar scenarios to Chucks. Certainly not as difficult but nevertheless a requirement for FSD, Hard coding a solution for Chuck would just be foolish and would backfire on Tesla.

4400 Ella Blvd · 4400 Ella Blvd, Houston, TX 77018

 

[legendsk]

this is the adult version of "are we there yet?"

I finally made up the rule: Okay, each kid can ask "are we there yet", or some variation, but only ONCE a trip for each kid. Then I could sit there smiling as I could hear them in the back seat, arguing with each other - You ask him. No, it's your turn, I asked him first last time. . . . . . . .

 

[jabloomf1230]

This sort of intersection bears many similarities to extremely common scenarios. The difficulty and hazard of this one is particularly high, but the steps needed to execute such a turn successfully can and must be applied to much simpler scenarios. How many times does one turn onto a road with two lanes of traffic each direction, with a center median? It’s really not uncommon. They can be quite dangerous and difficult to accomplish. It’s fine if you just want to not solve these situations, but that is not the stated goal.

My impression with this turn is the difficulty is not the visibility or the perception. It just was not employing a working strategy (amongst other planning issues and delays)! This issue extends to many other aspects of beta as well. The whole understanding of how to come up with a working strategy for arbitrary situations is lacking. There are a nearly unlimited combination of scenarios. I don’t think the answer is to program and fit each one manually.

We shall just have to disagree on whether it's an outlier. I've watched every one of Chuck's videos and although that intersection does share some features with others, all the attributes combined are a rarity. In my whole experience, I have seen one other intersection that maybe shares all those attributes with Chuck's ULT.

Let's see if we can agree on the attributes:

1. The intersection is obscured by vegetation.
2. It is controlled by a stop sign.
3. It is a tee with a median that might fit an angled vehicle.
4. The main highway is multilane with a speed limit of 55 mph, but the traffic often exceeds the speed limit.
5. There is neither an upstream or a downstream traffic signal which could create predictable gaps.

Chuck's intersection is an example of poor traffic engineering which could be ameliorated (assuming a traffic light is inappropriate) by cutting back the offending vegetation. The best solution would be to add this sign:

 

[n.one.one]

The latest from Elon.

 

[Dan D.]

We shall just have to disagree on whether it's an outlier. I've watched every one of Chuck's videos and although that intersection does share some features with others, all the attributes combined are a rarity. In my whole experience, I have seen one other intersection that maybe shares all those attributes with Chuck's ULT.

Let's see if we can agree on the attributes:

1. The intersection is obscured by vegetation.
2. It is controlled by a stop sign.
3. It is a tee with a median that might fit an angled vehicle.
4. The main highway is multilane with a speed limit of 55 mph, but the traffic often exceeds the speed limit.
5. There is neither an upstream or a downstream traffic signal which could create predictable gaps.

Chuck's intersection is an example of poor traffic engineering which could be ameliorated (assuming a traffic light is inappropriate) by cutting back the offending vegetation. The best solution would be to add this sign:

View attachment 834182

I had a quick look around Chuck's local area. Roosevelt has dozens of these intersections, so for this area at least it seems pretty common to have risky lefts.

Here's a fun one at Roosevelt & Adrian. The oncoming traffic isn't even visible very far due to the curvature of the road.
https://goo.gl/maps/VdUhtn8W2dD9u38h8


A block over at Rensselaer. Notice the Google camera level is mounted higher than the white truck's roof. Wonder what the view is like down at B-Pillar height. They do like their obstructed views in Jacksonville!
https://goo.gl/maps/K2hWikUBbPTFyzKr9

 

[Dan D.]

Technically this is a no-left turn from Randall to Roosevelt, I guess. There are no signs to prohibit left turns but I think the sharpness of the median is supposed to be your clue. Conceivably I could see traffic also trying to go straight. The Google maps know it's one way but there are no signs from Randall to tell you that.
https://goo.gl/maps/zZUBjCGJdfY9ZzVN6

I don't really have a point about this intersection except that you have to decide for yourself if left turns are allowed from Randall based on your own interpretation. Mapping would be of better use here rather than vision for FSD as it would be marked already as no left turns.





Curiously there used to be a Do Not Enter sign on the straight direction in 2011 that has since been removed.

 

[legendsk]

I live out in the country, about 25 miles from town, 2 miles from a 4 lane divided highway that looks a lot like Chuck's choice. The differences are: the speed limit is 75 mph and yes, the majority of traffic is running around 80 - 85 mph. Vegetation isn't as robust and only about 1/2 of the estimated 25 intersections, are as obscured as Chuck's example. Some of those intersections have flashing yellow lights overhead on the highway to alert / warn highway drivers, which are totally and profoundly ignored by all of the highway drivers. My MYLR violently brakes at about 1/2 of them and ignores the rest, so I have to be on high alert to punch it, to prevent being rear-ended by the traffic following much to closely. Many of the local's solution seems to be to drive around 50 mph with their hazard lights flashing, sometimes in the traffic lane, sometimes on the shoulder, which makes using any driver assist functions dangerous and useless. The best solution seems to be, when there is a frontage road available, take it. They usually have 2 to 4 lanes, one way with a 55 mph speed limit.

 

[LowlyOilBurner]

Not for nothing, personally, I would avoid those type of left turns altogether manually driving.

 

[legendsk]

Not for nothing, personally, I would avoid those type of left turns altogether manually driving.

Indeed. Far better to get on with a right turn, then make a U-turn at the next opportunity and hide in the dividing median, waiting for a gap. However, doing that must be done manually. Have to creep out and lean forward enough to see far enough to the left to gauge the speed of oncoming traffic and spot a gap large enough to do 0 -> 75 mph while merging without risk or being a hassle for oncoming traffic. The Y is great for this, if driven manually.

 

[AlanSubie4Life]

, so for this area at least it seems pretty common to have risky lefts.

Also, risky left turns are everywhere. It’s very very common, and humans generally do them quite skillfully.

That was my point. Three lanes of traffic each direction is more unusual (obviously also common, though unusual), but there’s nothing else really different about this including the visibility (which is really not bad on Chuck’s corner - it is not obscured, you just have to be in the right place, which is easy).

 

[GalacticHero]

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 :-(

 

[legendsk]

I've been viewing the experience as sort of like teaching a teenager to drive. Now I think it's more like teaching a 5 y/o to drive. The car just doesn't have the intelligence, common sense, experience, hazard avoidance training of a typical teenager. Hopefully that does not mean AI will take 12 or 13 years to reach that level.

On a different note...
I remember around 40 years ago?? UPS modified their route planner software to avoid left turns, make 3 right turns instead? Don't know if it's still that way or not. Any UPS drivers on here?