Driving at Los Angeles Korea Town || Drunk Walker Detected || Jay Walker - 25:21 - CMePrint
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Chazman92
Member
#FSDBeta 7 2020.48.10.1
FSDBeta 7 evaluated on several unprotected left turns.
FSDBeta 7 evaluated on several unprotected left turns.
An excellent series of tests with concise editing, excellent camera position, vehicle display magnification, and great explanations of issues that need work.
AlanSubie4Life
Efficiency Obsessed Member
Regarding the car looking at left turn signals and signals on vehicles in general...is there any evidence in any of the FSD evaluation that the car is ever looking at these? Any of the Beta drivers care to comment?
Have there been any example videos showing reaction to signalling, definitively?
I wonder about the resolution of the cameras in the "across the intersection" cases. Is it ever going to be sufficiently reliable to make decisions based on this signaling? Seems like it requires pretty good resolution. Personally, I always assume people are going straight when they are signalling left in this sort of situation (usually it's better to look at the wheels than the turn signals - perhaps this is something the car can do as well?), but undoubtedly it will sometimes be useful to be able to identify blinkers, to be able to assign some sort of likelihood estimate to path prediction.
For highway/AP, I've heard claims that signals are observed, but I've never seen a definitive example myself where I thought the car was responding to a signal.
Have there been any example videos showing reaction to signalling, definitively?
I wonder about the resolution of the cameras in the "across the intersection" cases. Is it ever going to be sufficiently reliable to make decisions based on this signaling? Seems like it requires pretty good resolution. Personally, I always assume people are going straight when they are signalling left in this sort of situation (usually it's better to look at the wheels than the turn signals - perhaps this is something the car can do as well?), but undoubtedly it will sometimes be useful to be able to identify blinkers, to be able to assign some sort of likelihood estimate to path prediction.
For highway/AP, I've heard claims that signals are observed, but I've never seen a definitive example myself where I thought the car was responding to a signal.
powertoold
Active Member
I think the car is able to recognize hazard lights, especially for neighborhood trucks that can be driven around. We've seen many examples of this, although I'm not sure it's because of the hazard lights or the car able to recognize it's a delivery / garbage truck.
As for left and right signaling, I don't trust signaling in general. Based on my own driving experience, 10-15% of the time people signal, they don't carry out their signaled action, as expected.
AlanSubie4Life
Efficiency Obsessed Member
Right, that would be a situation which would be too ambiguous to determine whether there is any effect of the hazards.
It seems like it would be easy enough for a beta driver to construct an experiment and repeat some trials where a vehicle was signalling, vs. not signalling, to determine whether the vehicle changes its response. The easiest example I can think of is a situation with two lanes of traffic in the same direction, driving in the fast lane, closing on a vehicle in the slow lane with 5-10mph speed differential, and then that vehicle signals to move left (but does not move lane position at all) with appropriate timing, where it would be appropriate to yield to allow that vehicle to change lanes. And then compare to the response to the exact same situation, where the vehicle does not signal and does not change lanes.
This is a good example to test because the turn signal would be the only reason the car would respond - otherwise its behavior would be unchanged. So should be easy to determine whether it was a response specifically to the turn signal.
Of course, I agree, as I said above. You just add that info to your situational awareness and use it as a weight in your decision making. Particularly in oncoming traffic situations, it's much better to rely on observation of wheel position, which I assume FSD will do as well, as it's a routine driving task for most drivers. Someone can't turn into you if they do not have their wheels turned (this is particularly relevant when you have right of way, and will be entering the danger zone where they could try to take away your right of way). But looking at new signals (as described in first paragraph) is very relevant when it comes to easing off to let someone make a lane change that they want (where otherwise you would impolitely assert your right of way). Those types of signals tend to be highly reliable indicators. There are different classes of signalling, each with different reliability, and different levels of consequence if you choose to heed/not heed them. Quite a decision tree. I assume this is built in to FSD, or will be soon.
The lane changing yield/no-yield problem is kind of analogous to the “yellow light, do I stop” decision. But with turn signals instead.
But first, wanted to hear about whether we have clearly documented response to signalling. I would think it would have come up in FSD videos at this point. And if not, it would be great to get some videos documenting the response!
It certainly seems like a "must" to have very high capability on this, before wide release. Amongst many other things, of course.
Last edited:
DDotJ
Member
Not sure about current FSD capabilities, but PlaidNet was spotted to have the capability to detect turn signals and brake lights on other vehicles. From just watching FSD Beta videos, it seems like it handles it similarly to NoA where the vehicle only yields once it has detected the vehicle moving into the lane.
Edit: Found another tweet that says current production firmware appears to not detect turn signals. Not sure if Beta firmwares have it enabled or not for FSD, but it I guess there is a non-zero probability that it could be enabled since Tesla had been working on the capability.
Last edited:
powertoold
Active Member
For me, turn signals are only useful in the case someone is turning in front of me, so that I can anticipate stopping (not useful at all for FSD beta, as it's always sensing lead-car distance). Otherwise, I base unprotected left turns on the car's trajectory, not their signaling. For the same reason, I don't do right turns at stop lights when there's a car approaching in the middle lane, as there's a chance it'll change into the right most lane as I turn.
DDotJ
Member
Yeah, I remember in drivers ed they taught us not to trust other people's turn signals (or lack thereof). Too many people change lanes inside an intersection without signalling so I don't make right turns with approaching vehicles either.
I would like the ability for the car to detect a turn signal for a car and yield to let them merge in. Especially on the freeway with NoA the car will stick to the rightmost lanes and vehicles entering will flick on the turn signal to merge but the car won't slow down until they start crossing the lane line. I have to disengage AP to yield to merging traffic or manually turn the speed down to make room for merging vehicles. It would be nice for the car to realize that vehicles want to change lanes and slow down a little bit.
scottf200
Well-Known Member
Andrej Karpathy has mentioned signals in presentations ... context was as one of many indicators factored in related to merge/cut-ins. I doubt it can ever be used as what the other driver is doing but just one of the hints in the decision making analysis.
powertoold
Active Member
On the matter of hazard lights, check out how quickly FSD beta is able to recognize this half-double-parked truck, on a curve too:
At 3:15
At 3:15
90% of the time the turn signal is on. Yes, it's a known issue.
4:22 Good slowing for pedestrians approaching on blind corner. Might have slowed a bit much
6:13 It doesn't let a car in that is halfway into the road. Should it have, it didn't even slow down? Well, out of courtesy it could have, and the car could have dashed its way in.
1:40
LOL, both cars get to the stop sign more-or-less at the same time. Each waits the other out.
Driver of Tesla waves the other car ahead, but Tesla goes too. Funny stuff.
Maybe you need in-car driver gesture monitoring?
Extremely informative. FSD visualization predicts and constructs the world using neuro net. The opposite of HD map approach.
.
Similar threads
