Navigate on Autopilot is Going to Cause Accidents

[Skotty]

It will never beat a good driver. A good driver is equally attentive and has advanced anticipation, prediction, and improvisational abilities that an affordable autopilot system won't be able to compete with. The only advantage an autopilot system can have is a better sensor suite and the potential for mathematical precision.

Of course, a lot of drivers are not good drivers though. So it's still a worthy thing.

 

[barjohn]

Its a love/hate thing. I love that when I am in the right lane and NOA is engaged it won't try and take every exit like AP does and keeps you going straight ahead but I hate the fact that it doesn't look ahead very far before make lane change decisions. Most of us, if we are in a hurry, can see that if we maneuver around the driver ahead of us and there is a hole in front of hi and it will allow us to also get into the faster lane that is being backed by a car to our left can decide whether it is worth it or not but NOA has no clue and will often try to move to a lane that we can see is stopped just 50 or so yards ahead while our lane is moving though slower than we would like. In very heavy traffic it wanted to change lanes into the lane to my right (I was in the HOV lane) that was bumper to bumper and I had to keep ignoring its request. I kept thinking why would it want to go into a lane that was loaded with cars and no hole I could see to get into safely.

 

[insaneoctane]

I tried NoA a few times today. More EAP progress has been made and It was impressive to have the car signal and exit the freeway entirely by itself! But you can tell it's definitely "BETA" still. It makes very early suggestions to move the right most lane in preparation for the exit. Like 2 miles ahead. In my case the far right (slow lane) had lots of trucks and slow cars easily going 15 MPH slower than the other lanes, so I ignored the suggested lane change and beeps for a couple minutes and passed a bunch of cars before I confirmed the lane change about 1/4 mile before the exit. Saved noticeable time and annoyance by delaying the move. I understand that it was a suggestion but soon enough there will be a setting and the car will just do these things without confirmation. Changing lanes when it suggested would have just been stupid and really highlights how non-traffic aware the computer was. This experience tells me it has more learning to do.

I also agree with the assessment that as EAP does more that it's harder to predict when it is likely weak and monitor it's moves as well. Stay vigilant!

 

[barjohn]

Traffic awareness and when you need to start moving over in other lanes to reach an exit is difficult for us humans to do and sometimes require risky maneuvers to execute. I don't expect NOA to be as capable in this regard and would expect the program to be conservative and get into the rightmost lane well in advance of the exit to avoid the car having to try the kind of tricky and aggressive maneuvers I might execute. It needs to be a little more aggressive and it needs to speed up rather than slow down to merge. The slow down will certainly piss off someone behind you and sooner or later result in road rage and an incident because you did an action the driver behind you was not expecting. The way it merges is the way a new driver to freeways merges by slowing down in fear instead of accelerating to slightly above the speed of the traffic you are merging into.

 

[djgarrett21]

For me it comes down to how much work is required to babysit NoA. With NoA activated I'm require to monitor it, in addition to monitoring traffic as I normally would when I'm in control of the vehicle - more mental energy/stress. Thus for me NoA is not an enhancement but rather a novelty. This is certainty not a feature I'd want someone who is elderly or generally a poor driver utilizing, way too easy to get into trouble right now.

Yes, we all know these features are "beta" so no need to get defensive and state the obvious.

I think its important to discuss the limitations and issues we encounter as these updates get rolled out. Folks need to be aware as to what to expect and be informed regarding how the system behaves.

 

[Abstrakt1]

I saw this post yesterday and figured it was just another one of those autopilot isn’t perfect threads that pop up daily. I tried nav on autopilot today and you are right. This feature is not ready for the road yet.

Every time there was a crucial lane change to stay on the route it put itself in a position where it couldn’t change lanes. It would only start trying about 200 feet before it had to be in the other lane. I had to take over and make a tight squeeze because it wouldn’t change lanes due to the (light) traffic.

It would tell me to change lanes to stay on the route and then immediately tell me to change back to the lane I was in (both lanes would keep me on route).

Getting off the ramp it said take over in 200 yards and then just stopped in the middle of the ramp.

It cancelled lane change on three different attempts. With no cars anywhere in site. It slowed down from 70 to about 40 to try a lane change instead of speeding up to about 72 to pass with cruise set to 78 and no cars ahead of me.

All this in less than 100 miles. This feature seems to have a long way to go and probably shouldn’t be in customers cars right now.

 

[sperkin]

This is what happened to me today. What do you guys think? I don't want to be an unpaid Beta tester anymore.

Steve on Twitter

People behind me passed me later and was laughing their a$$ off. There was another bad turn on the wide sweep to get on the 55S from the 5S. I forgot to save the video but car almost hit the left wall at 50mph.

 

[NuclearPowered]

This is what happened to me today. What do you guys think? I don't want to be an unpaid Beta tester anymore.

Steve on Twitter

People behind me passed me later and was laughing their a$$ off. There was another bad turn on the wide sweep to get on the 55S from the 5S. I forgot to save the video but car almost hit the left wall at 50mph.

IMO, this is the issue that is going to cause an accident. There is a spot in my commute where one lane splits to two like this and it swerves hard into whichever lane it randomly selects. Drivers in the adjacent lanes do not appreciate. A hardcoded right or left rule would be better than whatever it is doing currently.

 

[brandonee916]

It's beta... my left turn signal fires up on an exit only lane to the right... "Bug Report" followed by a brief situation. Most likely they didn't code the map data properly for that lane split. That said I'm also noticing the strange swerving behavior on lanes that merge in... exactly like Nuclear said, it should be hard coded to stay to the left (or right depending on the lane action).

This is what happened to me today. What do you guys think? I don't want to be an unpaid Beta tester anymore.

Steve on Twitter

People behind me passed me later and was laughing their a$$ off. There was another bad turn on the wide sweep to get on the 55S from the 5S. I forgot to save the video but car almost hit the left wall at 50mph.

 

[Kermee]

Unfortunately, the white striping of merge lanes, off-ramps and on-ramps and regulations are extremely varied in the U.S. Primary function of EAP still uses those lines in determining which lanes to take and which lane the vehicle is in. Where I live, the standard is all over the place (i.e. no striping at all, some striping, or 100% fully striped and EAP *loves* 100% fully striped).

The example in "Steve on Twitter" video, you can see the off-ramp lane starts to widen, then about a hundred feet later, white striped lane demarcations start to appear. At that point, EAP and cameras needs to figure out which lane to split into and EAP guesses (incorrectly) to go left.

Obviously, all of this information could be coded into the HUGE maps data that EAP is starting to use along with the cameras to infer the correct course of action, especially with NoA. But until the information in maps is more accurate, I'm afraid more of these incidents are going to happen.

For example, I have a U.S. spec'ed Model 3 and traveled to Canada over the weekend where NoA is still unavailable. Probably the reason why is the maps data that Tesla needs for NoA in Canada isn't up to par yet. My Model 3 tried to kill me the other day in Canada with NoA because the off-ramp was moved 200m earlier down the freeway (I posted here), two years ago, than what the maps thinks where it is. The old location of the merging lane for the exit has been replaced with... a jersey barrier. Which my Model 3 was trying to merge into after it blew by the actual merge lane earlier.

Hilarity ensued.

 

[xav-]

This is what happened to me today. What do you guys think? I don't want to be an unpaid Beta tester anymore.

Steve on Twitter

People behind me passed me later and was laughing their a$$ off. There was another bad turn on the wide sweep to get on the 55S from the 5S. I forgot to save the video but car almost hit the left wall at 50mph.

Granted I don’t have “drive on autopilot” yet... but I have used autopilot enough to know that I would never ever use it at that location (I live close by)

What you tried here is a major no no. My advice is stay out of the carpool lane (especially if there isn’t any traffic like in your video) or stay in the carpool lane but temporarily disengage AP on those situations (lane splitting / merging)

 

[Az_Rael]

3rd, even when I don't want to take the suggested lane, AP will slow down a lot and wait for me to confirm. Almost got rear ended.

THIS is so annoying. I can ignore beeping and messages, but when the car starts slowing me down so I can make the lane change, that is very annoying to me and the guy behind me.

Speaking of which, what situations would you want to have a follow distance of 7?!?! I

I use follow of 7 all the time. It is a long way behind in freeway speed traffic (which I like - lots of time to correct it if needed), but in stop and go, it is my normal follow distance. I don’t have much trouble with people cutting in even down in LA, so I think the differences in the settings at slow speeds must be very small.

 

[sperkin]

Granted I don’t have “drive on autopilot” yet... but I have used autopilot enough to know that I would never ever use it at that location (I live close by)

What you tried here is a major no no. My advice is stay out of the carpool lane (especially if there isn’t any traffic like in your video) or stay in the carpool lane but temporarily disengage AP on those situations (lane splitting / merging)

Nav on AP is suppose to fix this. It goes On ramp to Off ramp and navigate interchanges. It does this 95% of the time. It's the 5% where it fails that will scare you or cause an accident.

I understand Elon put these out for us to test to get millions of miles but it's not safe when that one driver starts to text and not pay any attention.

 

[barjohn]

Driving with it on NOA from Riverside to Newport Beach this afternoon and its behavior was so startling that my wife threatened to never ride with me again if I used it while she was in the car. What does NOA do wrong you might ask? I was in the HOV lane and there is an underpass where the concrete barrier widens out to the edge of the lane, the car tries to swerve first to the right and then to the left forcing me to take control away from it. That was the first scare. The next scare occurs because it wants confirmation to make a lane change even though there is no one directly ahead slowing you down and you are in a two lane section of toll road so you confirm the lane change and it abruptly slows down from 75 to about 45 before executing the lane change. My wife's reaction was immediate as it was unexpected. Then further up on the 91 toll road where it approaches the 55 split and the left lane goes to an overpass keeping you on the 91 and the right lane leads to a merge into the 55 fwy, about 1 mile before the split the road curves to the left and there is heavy traffic on the lanes adjacent to the toll lanes that is stopped or crawling along and even though there is no car in your lane ahead of you, it slams on the brakes, not once but twice. with a slight pause between where it accelerates before doing it the second time. It was after that last event that my wife insisted I turn NOA/AP off.

This kind of erratic behavior is what must be fixed. I am on 42.4 so it is definitely not fixed yet. I also keep finding its routing to be strange and unpredictable. While I can live with the fact that its routing engine is not always the best, its constantly telling me it wants to change lanes to stay on route when I am on a restricted toll road and there is no one ahead slowing me down it a real pain. You no sooner change lanes to make it happy and it wants you to change back. It doesn't appear to understand HOV or toll roads and completely ignores directing you onto them. Also when a route changes names (hwy or street) it tells you to take an exit to the right or left even though there is no exit, it is just a change in name. Near my home, Alesandro Blvd becomes Central Avenue after a bend in the road. It says I should exit onto one or the other. The 55 fwy from Newport Beach becomes the 91 freeway where it ends and intersects the 91 fwy east bound. It calls it an exit to the right but the exit to the right is the 91 going west bound and the toll road lanes are to the left. If the car is going to explicitly follow navigation, the navigation has to be right. It has to know where toll/HOV lane entrances and exits are and to avoid entrance areas by staying in the lane opposite of the entrance lane. When exiting on to an exit lane that is not a merge to another freeway but a lane with a possible stop, when it disengages it should do so completely and ask the driver to take over. Merely turning off the NOA but leaving it in AP makes no sense and could easily lead to an accident as you go flying through the stop sign or traffic light at the intersection with the surface street or it barrels into the stopped cars waiting at the light. I know that in some places on some occasions it seems to know it is a yield or stop and slow down and even stop but that is the exception. It has never done that for me in my car and I have had to quickly take control and apply the brakes to prevent an accident.

In short, it makes a lot of mistakes and it makes them so abruptly that it could cause an accident. I am a huge technology fan having worked in tech my entire life and I see the potential benefits this technology could bring but it needs some very rapid improvements to at least be safer and please stop the braking for a lane change. If it isn't clear and safe to make a lane change, then wait or don't do it. The braking lane change maneuver is a last resort maneuver that should only be executed when you have no other choice and if you don't you will miss an exit, not to get into a faster moving lane. Changing lanes for better speed should only be done if it can be executed at the present speed or by accelerating to a higher speed while executing.

 

[cwerdna]

A few of the problems reported (like from barjohn) could possibly be mitigated if they took of the approach of GM Super Cruise, where the feature can only be enabled in areas that are properly mapped ahead of time.

 

[eli_]

Lane change control needs a lot of work, I think the main problem is every lane change maneuver is performed with the same amount of urgency. It should modulate aggressiveness (i.e. willingness to change speed, lateral movement rate, blind spot gap allowance) based on how urgent the change is needed, like if it's just intended for passing or miles before taking an exit, you can just maintain your speed until an opportunity passively presents itself. With the way it works now, it drives obnoxiously and abruptly slows so it can cut into gaps. It would be better at playing Tetris than driving.

 

[ForeverFree]

Agree with all of the above.

I'm an experienced S/3 AP 1 and 2.5 driver and have been very pleased with AP since its toddling start in 2015.

NOA is actually far from terrible. However, changing lanes in preparation for exiting is one of the tougher driving challenges which all of us face. In SoCal, where so many drivers operate under the SUOSTS system (Speed Up On Seeing Turn Signal), it can be the most most nuanced, artful, and stressful part of driving.

Not necessarily the sort of thing one would be comfortable inviting a stranger to drive ... in your car. Make the stranger a tad awkward, hesitant, indecisive, and non-assertive, and the discomfort level rises fast.

As excited as many of us are to watch autonomous driving technology begin to stretch its wings, the truth is that most of us just aren't happy turning the scariest part of driving over to a rookie ... who doesn't drive the way we like to drive ... probably doesn't drive as well as we drive ... stumbles, hems, and haws at just the wrong moments ... while we sit in the driver's seat waiting to pick up the mess and avoid disaster.

All of that said, so long as you don't put it into crazy-tough situations (SB 405, attempting to exit Valley Vista ... with zero-gap merging traffic from NB 101 ... and then crossing two lanes of inbound traffic from SB 101 ... all within half a mile), it actually does a pretty responsible job.

I've deliberately (and on high alert) "confirmed" with cars next to me or in my blind spot, and it's always waited for them to clear. Biggest gripe, as others have noted is that its go-to move is the slowdown-and-slink-behind, even when a speed-up-and-slide-ahead would be far safer and less disruptive to those behind.

On the plus side, unlike a human driver, it can check on traffic coming up from behind in the new lane without interrupting tracking of the car ahead in the current lane. This is the one area where multi-GHz computer + multiple cameras can multi-sense better than a human.

On the still-learning side, I can say that the Nav part of NOA whiffed for me twice, once trying to move me left out of the proper NB 405 --> WB 118 ramp lane and another time not wanting to get right enough for the EB 118 --> SB 405 transition.


TLDR: Early days. And, we may have overestimated our readiness to invite a different-styled stranger to share the driver's seat with us during the hairiest moments of our drives.

 

[insaneoctane]

I have to admit that I am not really sure what to think of NOA in its current configuration. I have been using it every chance that I can over the last few days because it's the new toy. I think what surprises me the most is that I don't t think it's value is growing on me (like much of the tech in my Tesla) but rather it is feeling like something that I probably won't choose to use. I will try it a few times on my trip to work (110 miles round-trip) to confirm this feeling. I think it simply comes down to what is a net benefit (and what isn't). For EAP in general, for my usage, it's clearly a net benefit. I let it do its thing and I know what to watch for when it comes to weaknesses. I have found a great value, especially in traffic, with EAP and can find much stress relief using it. But when I add NOA into the mix, I really have to monitor it extra vigilantly. I am constantly looking for the screen to display a suggested maneuver and double check it's suggestions (lane changes) for safety. I find it anything but relaxing actually. These new lane changes have gotten pretty dicey. It often slows down when I wouldn't have. It seems very timid and aborts surprisingly often. I find it is too conservative at wanting to move early into position for upcoming events that would cost an annoying loss of time compared to what I would naturally do. It seems like it might be more useful on empty roads and interchanges (like a cross country demonstration?) than on busy roads during morning and evening commutes. But, that's what I drive. I'm a little saddened that I am not over the moon with this newest feature yet, but I am trying to be honest. I sure hope it gets better by learning quickly.... Makes FSD seem so far away!

 

[leisurelee]

I think it is fascinating reading other people's experiences with this technology.

I have been taking a different approach when testing the AP and NOA functionality, and that is to put the car in "Sunday driving" conditions, which it seemingly excels at.

Here's what i mean: I live in SoCal, so all of the freeways are multi-lane (3+ lanes). With that in mind, if I drive in the second lane from the right and stick to or very close to the speed limit, and if I behave like an attentive passenger and just stick to my lane, my car is a happy car. Basically, I pretend that I'm already in a self-driving car and ignore my personal impulses to speed or weave around traffic to get there sooner.

I'm planning a much longer trip next week, which will give me a much better chance to test things out. I'm hoping my strategy pans out over a 700 mile trip along coast 101.

 

[Matias]

This is what happened to me today. What do you guys think? I don't want to be an unpaid Beta tester anymore.
.

You are not unpaid tester, you are paying