Thoughts from a first time FSD Beta User: HW4

TLDR Summary: The FSD Beta Stack (11.4.4) on the highway is much smoother and more comfortable than the current production Navigate on Autopilot. (2023.26.11). Self driving on local roads is legitimately for beta testing and feedback, and not comfortable at intersections. NHTSA is doing a major disservice by forcing 'letter of law' behavior at stop signs.

I've been following FSD Beta since it went public a few years ago. I have seen many you tube videos from the usual suspects (Dirty Tesla, Chuck Cook, etc.). Still, until you experience it, you don't really know what it's like. As someone who has only owned a Tesla for a couple months and just recently had the Beta available (Model S with HW4), I thought I'd try and explain my experience on trying it out for the first time.

Background: I have used the production Navigate on Autopilot on a 2000 mile (round trip) road trip, so I have fair experience using that. Today, I enabled beta and took it for a test drive on a combination of suburban and rural roads and on an interstate.

Highway: Very impressed with the smoothness and comfort level with lane changes compared to the production Navigate on Autopilot stack. I can easily see in the not so distant future this becoming the production Navigate on Autopilot and getting Level 3 automation status to boot. In my 20 miles or so on the highway, it was a noticeable improvement. In fact, once my Free FSD trial is over...if I were to go on another highway road trip, I would subscribe to the beta for a month just to have it for highway driving.

Local Roads: As I anticiapted, I needed to intervene a few times where I was not at all comfortable with what I thought the car was doing. Moreover, there are some things that I thought it would be better at...that it is not. For example, going along a straightaway and needing to make an unprotected left turn....several times even though there were no vehicles in sight, the "smoothness" just was't there. It would slow down too late and too drastically, then "ease" too slowly through the turn, and abruptly speed up. Several times if the road I was on was curving one way (to the left), and there was a separate road intersecting to the right, the car would seemingly "be confused", slow drastically and continue forward.

Stop signs are a mess. I blame this squarely on NHTSA vs. Tesla. Most of you are probably aware that NHTSA had Tesla "recall" the beta to change behavior at stop signs. FSD must come to a full stop *at the stop sign* and then proceed to go through the intersection "safely." The problem is, especially in suburbs, stop signs are commonly well before the driver (or car) is able to actually see if it is clear. So the experience is like this:
1) Stop at the sign. (COMPLETE stop and pause).
2) Proceed slowly (Creep) because you can't see oncoming traffic
3) Then stop again once you can see and determine if it is safe.

You've seen this time and time again on FSD videos...but you don't appreciate how annoying this is until you experience it firsthand.

What NHTSA needs to understand, is that this is not "safe". Nor is it how everyone actually drives. What should happen (IMO), is that the driver / car (recognizing the stop sign) should slow down to a creep (not necessarily stop) "at" the stop sign, but continue to creep until you reach area where you can actually see the cross traffic. If NHTSA insists on a "full stop" (vs. "rolling stop")...that's fine...but THAT is where the stop should occur: where you can determine if cross traffic is clear...not necessarily at the location o the stop sign.

Not only is the current behavior annoying to myself as a driver...but will of course also be annoying to other drivers.

At the end of the day, I enjoy driving myself on suburban and rural roads, so I don't envision ever really using FSD Beta to drive locally. I am confident that if anyone can solve self-driving with a scalable solution, Tesla has the only shot based on their approach. I do hope that regulatory agencies can see the forest though the trees and understand that driving in the real world is not necessarily consistent with the letter of the law, and for good reason.

I'll continue to play with FSD beta over the coming month (until my trial subscription is up), and afterward will either subscribe temporarily for a month to test out any major new advancements...or when going on a highway road trip. Again, the FSD beta stack for highway driving is much improved (IMO) over the current production Navigate on Autopilot.

"Tesla Hardware 2 cameras, Sunnyvale" by mliu92 is licensed under CC BY-SA 2.0.
Admin note: Image added for Blog Feed Thumbnail

 

[heymull]

I agree with you about the stop sign behavior with the new beta FSD in my 2-month old HW4 MYLR, it is very annoying and confusing to other drivers, and they often can't be bothered to wait for your car (you! the Driver who is visible to the other drivers) to figure out what it is doing, so they just shoot out in front and the Tesla then has to slow down and then restart the accel phase.

I also fully agree that this new beta FSD is dozens of times better on the highway, than anything that I could configure before (because beta FSD was being withheld for HW4 till this update, and the other options were jaggy/not responsive/over responsive). So now, I can imagine paying the monthly subscription for long distance trips after my free-FSD time expires in a month.

 

[Ed243]

The moving over to the left lane when there is no one in front of me is perplexing. Still, it's impressive tech.

 

[BigNick]

How does it handle situations where a stop sign is present but hidden behind foliage or other obstacles (such as a big-ass truck or van parked there)? Does it have a "map" that knows the sign exists or does it just blow straight on through?

This is just one of the real-world situations that the software will have to overcome before it's ready for prime time.

 

[Joey D]

How does it handle situations where a stop sign is present but hidden behind foliage or other obstacles (such as a big-ass truck or van parked there)? Does it have a "map" that knows the sign exists or does it just blow straight on through?

This is just one of the real-world situations that the software will have to overcome before it's ready for prime time.

Good question...I believe (but cannot 100% confirm) that it uses a combination of known (mapped) stop sign / traffic signal locations, in addition to using vision to supplement. As far as I know, Tesla uses Google Map services as a baseline for much information, and then possibly supplements it with it's own data gathered from the fleet.

 

[Dewg]

How does it handle situations where a stop sign is present but hidden behind foliage or other obstacles (such as a big-ass truck or van parked there)? Does it have a "map" that knows the sign exists or does it just blow straight on through?

This is just one of the real-world situations that the software will have to overcome before it's ready for prime time.

How would a human handle that situation if they've never been there before?

 

[GBMaryland]

For me, the one thing FSDb does that's going to get you killed is the auto-change into another lane on local roads. It continues to attempt to change lanes into turn lanes... (lanes that have dashed lines for turning). This all because Tesla has turned off the ability to prevent FSDb from changing lanes AT ALL TIMES. The current minimize lane change is not NO lane changes without human intervention. This causes serious issues in non-highway driving.

 

[enemji]

For me, the one thing FSDb does that's going to get you killed is the auto-change into another lane on local roads. It continues to attempt to change lanes into turn lanes... (lanes that have dashed lines for turning). This all because Tesla has turned off the ability to prevent FSDb from changing lanes AT ALL TIMES. The current minimize lane change is not NO lane changes without human intervention. This causes serious issues in non-highway driving.

What are you talking about?

I can use the turn signal and get the FSDb to switch lanes as needed.

 

[uscbucsfan]

How does it handle situations where a stop sign is present but hidden behind foliage or other obstacles (such as a big-ass truck or van parked there)? Does it have a "map" that knows the sign exists or does it just blow straight on through?

This is just one of the real-world situations that the software will have to overcome before it's ready for prime time.

Right now...FSD will act like there's no stop sign, even if the sign/map has been there for 20+ years. We've seen this. The car relies more on vision than map data.

 

[uscbucsfan]

What are you talking about?

I can use the turn signal and get the FSDb to switch lanes as needed.

FSD will enter short turn lanes that split from a main lane at a high rate of speed ever since 11.4 was released...it's dangerous and can happen at a split second.

 

[BigNick]

How would a human handle that situation if they've never been there before?

Every human driver is different, but among the possibilities are:

* Observe the behavior of drivers ahead of you. If a "local" driver stops, assume there is a sign they know about but you cannot see.
* Check to see if cross traffic has stop signs or just goes right through the intersection.
* Stop anyway. Might get you a horn and a finger but that's better than a t-bone wreck.
* Blindly forge ahead and pray for the best.

 

[Dewg]

Every human driver is different, but among the possibilities are:

* Observe the behavior of drivers ahead of you. If a "local" driver stops, assume there is a sign they know about but you cannot see.
* Check to see if cross traffic has stop signs or just goes right through the intersection.
* Stop anyway. Might get you a horn and a finger but that's better than a t-bone wreck.
* Blindly forge ahead and pray for the best.

And therein lies the problem. Which one should FSD Beta do? If there is no sign visible (behind overgrown foliage), faded lines on the road, and no lead car - what does Beta do? Does the map data always take authority on driving, even when cameras disagree? Do cameras take authority, even when map data disagrees? Does it slow down, traveling through the intersection slowly and carefully, but therefore running a stop sign which is illegal? Does it slow down and stop, even without a stop sign showing?

I'm not saying I have the answer, just that it's a tough question to answer.

 

[rlsd]

And therein lies the problem. Which one should FSD Beta do? If there is no sign visible (behind overgrown foliage), faded lines on the road, and no lead car - what does Beta do? Does the map data always take authority on driving, even when cameras disagree? Do cameras take authority, even when map data disagrees? Does it slow down, traveling through the intersection slowly and carefully, but therefore running a stop sign which is illegal? Does it slow down and stop, even without a stop sign showing?

I'm not saying I have the answer, just that it's a tough question to answer.

Answer: Show message "FSD cannot handle this situation. Please take over."

 

[Dewg]

Answer: Show message "FSD cannot handle this situation. Please take over."

Interesting - hadn't thought of the red-wheel alarm method.

 

[rlsd]

Interesting - hadn't thought of the red-wheel alarm method.

Tesla needs to teach FSD to be sincere.

 

[JB47394]

I'm not saying I have the answer, just that it's a tough question to answer.

If the car cannot establish that it has right of way then it should assume that it does not, and it should slow with the assumption that it will have to stop. If it learns that it has right of way at any point, it resumes. I can imagine a variety of heuristics for figuring it out, but a trained system would ideally figure things out on its own. For example, I may not be able to see a stop sign for the intersection, but if the opposite side of the intersection has a stop sign, I can be sure that I have to stop.

Alternately, an accurate map database would be nice.

 

[zoomer0056]

How would a human handle that situation if they've never been there before?

With caution.

 

[Dewg]

If the car cannot establish that it has right of way then it should assume that it does not, and it should slow with the assumption that it will have to stop. If it learns that it has right of way at any point, it resumes. I can imagine a variety of heuristics for figuring it out, but a trained system would ideally figure things out on its own. For example, I may not be able to see a stop sign for the intersection, but if the opposite side of the intersection has a stop sign, I can be sure that I have to stop.

Alternately, an accurate map database would be nice.

In that case, accurate map data is important, and I think map data must take authority in the driving task. If the lane lines have faded (poor city management), and the stop sign is obscured, FSD Beta should still come to a stop. Absent accurate map data, all Beta can process is what it visually sees, which would be an unmarked intersection with no stop sign - like a residential street.

 

[clydeiii]

Stop sign behavior is variable. Sometimes it behaves relatively humanlike and other times it's too cautious.

 

[Dan D.]

Stop signs are a mess. ... The problem is, especially in suburbs, stop signs are commonly well before the driver (or car) is able to actually see if it is clear. So the experience is like this:
1) Stop at the sign. (COMPLETE stop and pause).
2) Proceed slowly (Creep) because you can't see oncoming traffic
3) Then stop again once you can see and determine if it is safe.

I am perfectly capable of fully stopping at a stop sign, wherever it is located and then proceeding to the point where I may need to look for traffic - without getting anybody annoyed. This is accomplished by doing it all quickly.

That is where Tesla needs to improve, if this process is bothering people. Come to a stop quickly, pause the requisite half-second or whatever. Proceed quickly to the edge of traffic and quickly scan and determine if you can go. If you can't go, then you wait until you can go. None of this should delay anybody. If FSD is unable to do this quickly then that's FSD's fault. Humans can do it quickly, especially there is no need to do step 3 (stop) if it is safe to go, you roll and go, no stopping required. You have to decide quickly.

Slowing to a stop, creeping, proceeding slowly, etc. are all slow behaviours. Humans don't need to be slow, you get it done. I'll bet most people who accomplish your step #1 (stop at the stop sign), will actually gun it for a brief instant to get to the edge of visibility, and gun it through step 3 without stopping if they can, or stop reasonably hard if they can't go at step 3.

FSD can greatly improve stop sign behaviour, by doing every step quicker. If it can't do it quicker because it's not good enough - that's a different issue, but there's no safety reason you can't stop fully at the stop sign line, and get going safely into traffic without taking tons of time. That's just inefficient programming or poor sensing/decision-making. There's no need to blame laws if FSD can't get it done.