Wiki MASTER THREAD: Actual FSD Beta downloads and experiences

[EVNow]

I used the fsd beta to get to work today. Mostly 50-55 mph roads with stoplights. Overall better than expected based on some other user reports.

I think the car struggles more in the mostly unmarked neighborhoods, rather than well marked secondary highways. Kind of makes sense - AP has been handling well marked secondary highways for a long time. Unmarked roads are new and also more difficult to plan for, since there are fewer clear-cut rules.

 

[qdeathstar]

Elon has stated that he runs the latest alpha build on his car, and that he uses FSD both ways on his daily commute.


There's no way that Tesla would assume the liability of releasing an alpha build state version to anyone outside of the company.

If you were an FSD dev, wouldn't you frost the cake? I don't think it's unreasonable to imagine that the devs make sure that Uncle Elon has a very nice drive both ways to work. And I don't think Elon is naive enough to not think that, too. He has stated that he takes the current private beta build out quite often and takes it through its paces on an extended basis.

Self driving cars has always been one of Elon's passions. He keeps a very close eye on his pet project, but yeah... he does have a tendency to guild the lily when it comes to its actual performance.

As public beta testers, we're testing software that is many iterations away from what Elon is driving on his car; most likely months behind.

What’s the difference between the liability of releasing an old beta vs a new alpha? They both do unpredictable things

 

[qdeathstar]

Beta didn’t handle Hollywood Blvd very well, lol. Too hesitant with all the traffic and pedestrians causing too many abrupt stops early at traffic lights forcing me to quickly hit the go pedal briefly (happened twice in a row), so I just look over. Nothing really dangerous (ie risk of death or injury) if I didn’t intervene beyond possibly causing a fender bender if someone behind me was too close and didn’t expect an early stop, but I don’t want property damage either.

The weird glitch where Beta wants to get out of the proper lane in the opposite direction of an upcoming turn continues. I really have no idea what the decision making process is for lane placement to have it return that result, but it’s fairly frequent.

I think it is thinking the cars in front of you or slowing for something else and wants to go around them, it’s a little to impatient in that regard. I made the same turn a few times and it will stay in the lane as long as there isn’t a car slowing down in front of me… both times there were cars in front it acted like it wanted to go around them.

 

[Phlier]

My take is that the first (autonomy) day was more theoretical and aspirational - after this they came up against some brick walls and had to change the approach to solve the problem, such as move to the vector space and vision only - for me AI day this year was much more polished and complete, with breakdowns of the individual nets, what they do and how they are trained.

You can see many of things from AI day in FSDBeta today.... the same cannot be said for Autonomy day in 2019.

Its like the jigsaw pieces are now in place, and it is now incremental improvements and training, lots of training..

Completely agree.

It's like they had their epiphany moment when they realized that this was going to be much, much harder than they thought, and they pretty much started over with a re-write that we're now testing as beta FSD 10.2.

However, I have a feeling (and boy, do I hope I'm wrong), that they only think that they've had their "wow, this is going to be so much harder than we thought!" moment, and that there is yet another brick wall waiting in the darkness ahead, leading to yet another epiphany of "oh my God... what have we done?" levels of "wow, this is soooooooo much harder than we thought!" Leading to yet another fundamental approach change and another re-write.

How many of those brick walls lie ahead? Is our current state of technology even capable of solving this problem? Is even Hardware 4.0 going to have the grunt to handle this?

It reminds me of Tony Stark's father, "I'm limited by the technology of my time..."

When I think about some of the things they still have to fix and/or implement, it's just mind blowing.

I'm hoping that they've worked on this long enough to really have a firm grasp of how hard it is, but I just can't help but have a little niggling feeling that says they still don't realize it.

IMO, Elon won't ever give up. He'll eventually deliver on his promises. But it's going to end up taking more time and work than what they think it will, even with everything they've learned so far.

What they've achieved so far is truly amazing, to be sure, but what FSD has really taught me is that this is a lot further off than what I'd previously thought.

It has also taught me that my wife is a much better driver than I'd given her credit for.

 

[Phlier]

What’s the difference between the liability of releasing an old beta vs a new alpha? They both do unpredictable things

Hopefully, likelihood of death.

 

[Phlier]

I think it is thinking the cars in front of you or slowing for something else and wants to go around them, it’s a little to impatient in that regard. I made the same turn a few times and it will stay in the lane as long as there isn’t a car slowing down in front of me… both times there were cars in front it acted like it wanted to go around them.

Re: When does FSD decide to go around cars?

This is something that I've been wanting to discuss, but was waiting until I had some more data points first.

For the most part, FSD has been pretty good about discerning parked cars from cars that are actual, active traffic.

But sometimes the logic fails pretty bad.

Example: There's an intersection that I've done quite a few times with FSD Beta. It's a right turn from a four way traffic light, and the right turn lane is usually 3 to 4 (or more) cars deep. It usually handles it quite well.

A few days ago, I was following a car that didn't have operating brake lights. We slowed to a stop for the right hand turn at the light. At this light, the right turn lane has quite a bit of separation from the straight lanes. About ten seconds after we'd come to a complete stop, FSD decides that the preceding car is no longer "active traffic," and starts to try to go around it. Again, the car we're following has no brake lights (or turn signal). Obviously, I disengaged.

It would appear that they are using tail lights, in at least some fashion, to help sort out "active traffic" from "parked car."

Has anyone else had something similar happen to them yet?

 

[EVNow]

What’s the difference between the liability of releasing an old beta vs a new alpha? They both do unpredictable things

Alpha is too bad for even Tesla to release

 

[rxlawdude]

We're testing embryonic software?

Life begins at the conception of the software. Or so some say.

 

[tpatana]

I wish I could set individual speed settings, like "don't drive faster than 15 mph on this street". Now on my home street I always need to use the wheel to dial down the speed, I think official limit is 25 so car goes ~29, but good/safe speed in here is ~10-15.

 

[qdeathstar]

I wish I could set individual speed settings, like "don't drive faster than 15 mph on this street". Now on my home street I always need to use the wheel to dial down the speed, I think official limit is 25 so car goes ~29, but good/safe speed in here is ~10-15.

Seconded. Location aware speed limit offsets!

 

[mhan00]

I think it is thinking the cars in front of you or slowing for something else and wants to go around them, it’s a little to impatient in that regard. I made the same turn a few times and it will stay in the lane as long as there isn’t a car slowing down in front of me… both times there were cars in front it acted like it wanted to go around them.

I thought that too, but yesterday it tried to change out of the lane I was in when there was no car in front of me.

 

[qdeathstar]

I thought that too, but yesterday it tried to change out of the lane I was in when there was no car in front of me.

Maybe it imagined the car!!!!

 

[meitschis]

Today it tried to attack an island, for no good reason :-(
I took over, but too late to avoid the curb completely.

https://imgur.com/a/2OARi0y

 

[jstjohnz]

Scored ticked to 100 Monday night. Beta downloaded Tuesday night.

 

[TIppy]

Re: When does FSD decide to go around cars?

This is something that I've been wanting to discuss, but was waiting until I had some more data points first.

For the most part, FSD has been pretty good about discerning parked cars from cars that are actual, active traffic.

But sometimes the logic fails pretty bad.

Example: There's an intersection that I've done quite a few times with FSD Beta. It's a right turn from a four way traffic light, and the right turn lane is usually 3 to 4 (or more) cars deep. It usually handles it quite well.

A few days ago, I was following a car that didn't have operating brake lights. We slowed to a stop for the right hand turn at the light. At this light, the right turn lane has quite a bit of separation from the straight lanes. About ten seconds after we'd come to a complete stop, FSD decides that the preceding car is no longer "active traffic," and starts to try to go around it. Again, the car we're following has no brake lights (or turn signal). Obviously, I disengaged.

It would appear that they are using tail lights, in at least some fashion, to help sort out "active traffic" from "parked car."

Has anyone else had something similar happen to them yet?

I find that frequently at left turn red lights, Beta wants to get out of the left turn lane after I've stopped to pass the cars ahead of me on the right. Their brake lights are fine.

I've also had it refuse to enter the left turn lane when there is a line of cars waiting at the red light. It chooses a right turn only lane instead.

I haven't been behind a car with no brake lights, yet.

 

[mackgriggs]

Initial thoughts:

Fascinating software....the difficulty of the task is huge, but they are certainly solving it. I got my masters in computer science with image processing emphasis so I know the level of tech here is ridiculous. Reminds me back in the day with dragon naturally speaking software....which was another computer science problem trying to be solved (voice recognition). Expensive software that still barely worked, and had to be trained by having users read paragraphs of text. People didn't think it could ever really be done. Now....any smart device has an uncanny level of voice recognition. I imagine this will be the same.

Anyway, if you treat the beta as an old autopilot replacement and only use it where you would have previously, then this is a large upgrade.

- Can go faster than 5 mph over the speed limit
- No more green light confirmations
- More confident when lanes merge
- Won't slow down for "blank" traffic lights anymore

What I mean by that....I have lots of these crosswalks ->

that are almost never on. Same with lights like in front of a fire station. With old autopilot, it would always slow down 10-15 miles an hour unless you manually push on the accelerator. Annoying. Now it treats them properly!


For the new autopilot abilities.....it basically drives like a teenager with a learners permit. I should know...as a parent that is now teaching a teenager. I'm talking about neighborhood driving, getting into turn lanes, unprotected left/right turns, etc. Its handling protected left/right turns pretty well so far, though...maybe like a teenager that just get their license.

I'm really excited to see this software progress, and still think its an amazing technological feat so far. I'm very hopeful for the future, and will continue to use it whenever I can.

 

[EVNow]

Today it tried to attack an island, for no good reason :-(
I took over, but too late to avoid the curb completely.

https://imgur.com/a/2OARi0y

That curb is badly placed too - it juts into the lane a bit. Anyway, clearly the car doesn't recognize curbs properly, all the time.

 

[EVNow]

Anyway, if you treat the beta and use it everywhere that you would have with the old autopilot, then this is a large upgrade.

- Can go faster than 5 mph over the speed limit
- No more green light confirmations
- Won't slow down for "blank" traffic lights anymore

There are some upgrades and some regressions
- FCWs / Sudden slowdowns
- Stops way behind at traffic signals

 

[Dan D.]

That curb is badly placed too - it juts into the lane a bit. Anyway, clearly the car doesn't recognize curbs properly, all the time.

It turned towards the curb/island.

 

[ConSaab]

As I was driving in my neighborhood, I was thinking of a setting that allowed users to set speed limit on roads... personally in my neighborhood I drive 15-20 mph however posted speed is 30 but thats way too fast... specially since beta loves high speed turns.. it is definitely uncomfortable feeling