Tesla haven't recognized a train

[Knightshade]

You don't need your hands on the controls all the time in a plane.

You don't need them there in the car either if there's an alternative way to determine the driver is paying attention (see Caddys system for example).

Tesla has simply chose to use "interaction with wheel" as their monitoring system.

 

[duanra]

You don't need them there in the car either if there's an alternative way to determine the driver is paying attention (see Caddys system for example).

Tesla has simply chose to use "interaction with wheel" as their monitoring system.

You have a point, but there are an awful lot of reports of troubles with the AP and FSD.

Also, I could be wrong, but did Tesla have another mean of control acceptable by the government ?

 

[dmurphy]

The numbers are actually published by the FAA, but they are absolute bare minimum values. Airlines typically hire pilots with far more experience than the FAA’s minimums.

Last time I checked, the 1500 hour rule was still in effect. And that's - as you said - bare minimums.

 

[Knightshade]

You have a point, but there are an awful lot of reports of troubles with the AP and FSD.

The vast, vast, vast majority being user error by folks too lazy to read the manual though.

Y
Also, I could be wrong, but did Tesla have another mean of control acceptable by the government ?

The government has nothing to do with it- there's no regulation of L2 systems at all.

(well, I suppose it has to do with companies wanting to AVOID the government deciding they need to regulate it... but there's no specific requirements for any particular system as laws stand today).


The steering wheel system was the cheapest/easiest one available in 2014 when AP1 launched, and they've doggedly stuck with it ever since.

In part because Tesla keeps (so far wrongly) thinking they're just a year or two away from "real" FSD, where you no longer need a driver- so why bother dumping $ into checking their awareness when the car can drive itself?

The 3/Y does have an in-cabin camera, but it's terribly ill-suited for driver awareness (wrong type, in the wrong position) and doesn't even exist in the S/X models.... the stated purpose of the 3/Y camera is... monitoring the passengers for doing damage to the vehicle when operated as an autonomous robotaxi... (see again Tesla keeps making design decisions in their cars as if they have already solved FSD)

 

[Kevy Baby]

The vast, vast, vast majority being user error by folks too lazy to read the manual though.

But it is still somehow Tesla's fault.

 

[dgatwood]

Drive really fast towards the crossing....it always works out for the lead car in the movies. You'll make it.

Only if there's a ramp so you can jump through the empty flatcar.

 

[LACALawyer]

Only if there's a ramp so you can jump through the empty flatcar.

That's very Trans-Am of you to say....

 

[LACALawyer]

Drive really fast towards the crossing....it always works out for the lead car in the movies. You'll make it.

I'm kinda worried about the two people who marked this comment as "informative."

 

[JulienW]

Here is FSD Bata trying to go through or around or under/over a train and he has to disengage after the hood was actually under the crossing gates and no sign of stopping. Oddly the display says "Stopping for traffic control in 25 feet". That is about 10' on the other side.

Since it is clear FSD Beta doesn't recognize trains, there is no way our current AP/FSD will.

Time Stamp 16 minutes:

 

[byeLT4]

Looks really confused at the RR tracks with a train. It seems like it occasionally wants to add a road where the train is going, does it think the train is a bunch of trucks? Definitely needs work.

 

[LACALawyer]

I think this is a really interesting example of how the camera only solution fails. If it had lidar it would see the train. Maybe not understand it is a train but at least know it can't drive through it. It would see the barrier too. HD maps would help, it would know there is a crossing there.

It is a bit worrying that when it can't understand what it is seeing it just drives forwards.

Driving forward when I can't see basically describes my life.

 

[DaveG_NJ]

I think I see a Septa train there

Warminster line maybe?

 

[thewishmaster]

I think this is a really interesting example of how the camera only solution fails. If it had lidar it would see the train. Maybe not understand it is a train but at least know it can't drive through it. It would see the barrier too. HD maps would help, it would know there is a crossing there.

It is a bit worrying that when it can't understand what it is seeing it just drives forwards.

I don't think lidar is required for this specific instance - I might be oversimplifying but it looks like they need to add another classifier for trains to identify them as a single vehicle in the path. Based on the visualization each train car is treated individually and the motion planner attempts to make a run for it through the gaps it keeps seeing but missing due to reacting too slowly. Probably needs some custom behavior or training for knowing when to yield for trains the way it yields for cars/bikes/pedestrians.

 

[Hobojo153]

I think this is a really interesting example of how the camera only solution fails. If it had lidar it would see the train. Maybe not understand it is a train but at least know it can't drive through it. It would see the barrier too. HD maps would help, it would know there is a crossing there.

It is a bit worrying that when it can't understand what it is seeing it just drives forwards.

Lidar wouldn't help here at all. All Lidar does it give you a much higher res, and depth mapped, though monotone view of the world. You would have to do basically the same type of image recognition to "see" a train with it.

Maps also wouldn't solve this specific issue as they can only tell you if there's a train track at all. Not whether or not it's closed. (Assuming you're not hooking it up to a live train map feed)

 

[duanra]

Out of context : "you don't need to see the train"

A myope car will do !

 

[thewishmaster]

It's not really the train that is the problem though, but the time it is in view it's too late. It needs to recognize the crossing and the barrier coming down, and the flashing lights.

Yeah, for the majority of cases, but also incoming / already-visible trains in case for some reason the lights aren't flashing and the barrier is up

 

[Dan D.]

Yeah, for the majority of cases, but also incoming / already-visible trains in case for some reason the lights aren't flashing and the barrier is up

Rule #1 still seems challenging. Don't hit anything. It was going for the train, or at least the barrier.

 

[byeLT4]

Yeah, for the majority of cases, but also incoming / already-visible trains in case for some reason the lights aren't flashing and the barrier is up

Yup, and still several crossings around here with just the old crossbuck sign. No lights or gates.

 

[dgatwood]

Yeah, for the majority of cases, but also incoming / already-visible trains in case for some reason the lights aren't flashing and the barrier is up

It needs to recognize the flashing lights even if there is no barrier. Most railroad crossings in rural areas don't have gates. If it had treated such a crossing like a flashing red light (stop and then go), the occupants of the car would likely be dead a few seconds later. When those lights are on, you do not proceed unless you can see the train and it is stopped. And even then, do not proceed unless there is only a single track.

Ideally, it should recognize the sign, too, and if there are no lights at a crossing, it should stop well before it gets to the tracks, look down the tracks to see if a train is coming, and then proceed only if there isn't one. That said, I'd settle for it simply stopping before old-style unlit crossings and playing the "Take over immediately" chime, at least for now.

 

[Hobojo153]

See above, it's not the train you need to see, it's the crossing and the lights and barrier.

But they already can see that part. Even in the production build we have it rendering RXR markings and treating trail tracks as intersections to slow for.

The point is that the problem is simply that they're not doing enough with the available information, not that they don't have enough information.