Thanks from Germany. We have a really old FSD-Version here (and it didnt updated like yours) and i like to see how it "grow" over time in the US, hopefully it will come fast to us in europe.
@@pop_3310 Seems like their English isn't so great so that's probably what they meant. FSD, the software feature option you can buy, is available in Europe. But the FSD Beta program is not.
I am in Italy and purchased FSD with my Tesla. I can confirm that it doesn't exist yet. Both the highway and city driving are only Adaptive cruise + autosteer (same stack as on highway with FSD prior to v11)
I am continuously amazed by your ability to keep pushing your videos. This one had me on the edge of my seat the whole time, absolutely terrified by what you were gonna attempt next. An absolute emotional rollercoaster, but also quite educational.
Wow you are brave!! And Carol is very patient! What a good sport. Thanks for sharing all the data from your experiments. It’s nice to see the progress. Now if only I would get v11! My wife got it last Sunday!
Another stellar video and thank you Carol for helping!! Looks like some great improvements. I love that you attempted to approach this with some scientific rigor 😊
So if anyone is unimpressed by this version of FSD beta, I would like to know what they’re smoking or drinking, because this is terrific stuff here, and before I go further, my hats off to you and your wife for doing these videos, because it really shows the progress that Tesla continues to make, and this demonstrates FSD beta’s capability. So what will it be capable of in the very near future, because I don’t think most people truly understand any of this stuff and as you said, I don’t understand it myself, but I do know this, no other car manufacturer that has this level of capability without all of that external apparatus that you see some of these other so called self driving car companies like, CRUISE, LYFT, WAYMO, Uber and others who have equipped their leased vehicles, none of which they manufactures with, LiDAR, extra cameras, Geo fenced routes, and all that kind of stuff, which adds an additional amount of expense to the vehicle’s cost, so when Tesla make this work, just think of the advantage that they will have, because no one wants to buy a car that looks like some kind of weird looking Frankenstein monster. So it just doesn’t look like something that I would wanna buy, and so let them do their thing and Tesla will certainly continue to pursue their version of FSD beta to perfection, because they understand just how valuable this technology truly is, and I think the other guys know it too. So when I think about the fact that the AI is thinking and figuring things out, that’s what AI is about, and so I think too many people don’t truly appreciate just how incredible this stuff is. So I don’t have the update yet, and can’t wait to get it, but I can say this though, it would not surprise me that by the end of this year, FSD beta will be fully capable of driving itself. Cheers 🥂
I’m using 11 too. Yesterday, I noticed it recognised speed bumps with painted chevrons on the 25 mph street. It didn’t slow quite as much as I would have but it dropped 10-12 mph - reasonably comfortable. I don’t know if it recognises unmarked speed bumps… Thank you for the video and the time and effort you give!
It’s not even my car and my butthole clenches when your car gets close. Love these torture tests. Seeing the new visualizations for gates auto opening and closing made me realize the car really can detect objects in the way and the road, but still has difficulty making a decision on how to behave
This is the reason why Tesla is pushing for the decision and planning system to also use neural networks in their core instead of only partially. Because as much as we think the road is logical, it is anything but that. You can't simulate a driver with if-else statements. You can make a lane following algorithm with it, you can make an auto park with it, and you can make emergency breaking assists with it. But you cannot make a unconstrained level 3-5 system with simple statements. With a neural network you can train it to account for nuance and probable outcomes (as opposed to definitive ones) but if item detection already needed and extensive dataset, actually making a planning AI for the entirety of the road's situations will require a level of data never before seen.
You are great, and Carol is an awesome wife and partner in crime in these amazing videos. You are helping all of us really feel the evolution of FSD, and I'm sure Tesla is watching since you are doing things that are way outside the norm, but crucial to the allowing the Beta to become more reliable for the real world. Many thanks.
You are the real star of this show! Extremely useful information that I am very grateful for. I am 75 and feel like eventually my wife and I will need all the driving help we can get, hence the Tesla MY purchase. I haven’t subscribed to FSD Beta yet, waiting for the V11 stack plus $200/month is a lot for us that needs to be justified by increased safety. My big question now is how well FSD with HW3 will perform at speed, say 35-70 mph.
You probably shouldn’t. It is kind of more dangerous than driving yourself IF you aren’t paying attention. In your case, this can give you a false sens of confidence that is more dangerous. You should pay as much attention with FSB than without it. It’s far from perfect, and multiple models got recalled in the last couple of month. It is not safer, for the moment it is more of a gadget. Please drive safe! ❤
OMG he owns an ICE truck! 😲 How impure. 😄 Nice tests that almost no one else is doing for YT. I'm upset you'd risk a Model S like this, I mean wives are replaceable but...
This was really great. When it comes to oncoming car turning into your lane, I wonder if turn signals (on the oncoming car) would have made a difference.
Peter Gazdik: That's an interesting question. Clearly, humans (decent drivers anyway) would try to gauge intent and whether the other driver is clueless or not. The Tesla is still pretty OBLIVIOUS even when it's of actual concern, like the other driver being over the center line. And it doesn't help if a bad driver has a turn signal on and runs into you.
@@rogergeyer9851 Yes, I've also seen this behavior in previous videos on regular streets. If a colliding vehicle was moving slowly enough, FSD seemed to have considered it a stationary object and completely ignored it. IMO next visualization improvement could be an indicator of where or not FSD thinks a vehicle is moving.
I think it's just a combination of speed of motion depending on the object and being significantly in the way. Cars likely will stop if heading into your lane so they are more so ignored but a pedestrian would never be.
Awesome video. Thanks AIDRVR and Carol for all of your hard work. We truly appreciate the great content. We all learn from your tests and get entertained at the same time. :)
Thank you for doing these tests, I do not know of anyone else who does it and I think it is a fantastic way to get some insight into what's going on "under the hood". But like you, I have no idea what FSD is thinking in some of these situations. Looking forward to the next ones!
I actually thought that it was very impressive that beta caught the wheel in the road, particularly when it was in the shade through so much of the testing. It would have been easier to detect had it been in the sun, making a higher contrast image.
Not having any idea why it's making decisions it does is an inherent feature to how AI/neural networks currently work. Not even the people who develop them really have a good idea of why it makes certain decisions. This is also why self-driving cars are still terrifying, even though on average they're extremely safe
Benjamin: Yes. It's like we're literally growing "alien intelligence" when advanced AI work results in things like Alpha-Go or Alpha-Star, etc. This is a significant issue even for games of skill. Imagining it for something as complex and serious as a general driving solution is just mind boggling.
@@w0ttheh3ll: But we do have a good general concept for the way human minds work, and we're adding to that all the time. We might have no real idea AT ALL of the way AI works, given lots of parameters. It's something VERY different, IMO.
@@w0ttheh3ll I don't think so, because I think the premise flawed. People might not understand why they make certain decisions in life, but the basic goals of driving are usually pretty straightforward: Arrive at your destination, and do so safely and on time. The rules of the road are vastly simpler than the rules of life. Additionally, I think you're conflating the meta-knowledge I'm referring to. Generally, we can assume that any drivers that see a pedestrian on a road will come to the determination that they are in fact a pedestrian. Since we don't understand how an AI system "thinks", we have no way to intuit how likely that system is to identify that same pedestrian correctly. Humans typically have a huge portion of their cognitive abilities dedicated to understanding and predicting the behaviors of their fellow humans. We have built that up over tens or hundreds of thousands of years. We have no corresponding facilities for understand an AI's behavior, and what's more, we will be very unlikely to develop such abilities as long as AI remains an inscrutable black box. All that being said, please note that I deliberately talked about whether human drivers who *see* a pedestrian come to determine that it's a human; an AI system is much more likely to perceive a pedestrian than a human driver, which is certainly a massive point in its favor. On the other hand, if the AI system makes an incorrect decision about how to react to the pedestrian, it may well decide to speed up, for some unknowable reason due to its AI. It's incredibly unlikely that a human driver would react in the same way; the risks with a human driver almost exclusively lie with the lack of perception. So, when I say it's terrifying, I suppose I'm saying that I'm scared about the changing risk structure of existing around multi-ton missiles that are controlled by robots, the decisions of which I have a reduced risk of predicting as I attempt to navigate myself around them.
Great testing. The Sissys complaining about the wife testing in the prior video didn't take into account your multi-factor safety features. You not only wore a full helmet, but you also took out life insurance on her.
Your videos never fail to amaze me, and this one was no exception! The production value and attention to detail were truly incredible, and it's clear that your content is always improving. I'm eagerly looking forward to the upcoming updates and can't wait to see what you have in store for us. Keep up the great work!
I've been watching al your videos since my last comment like 10 videos ago, and im glad that you have managed to deal with your stress. al videos are top notch and are so enjoying to watch!! 😊
Jonathan Mui: Yes, because with Chuck Cook in recent months on the no-lines road he tests on, it repeatedly ignores branches in the road (even significant ones) that are hanging from trees. Lots and lots of cases to work on.
Great work! Very interesting test results! On the truck turning towards your lane test, actually I'm kind of okay with that change. I actually think in my experience the current v10 build freaks out and slows down *too* easily IMO. Like if someone is coming up to turn out of a parking lot or driveway, the car will slow down quite a bit when it really shouldn't. The default assumption appears to be (and should be) that even if it's moving towards your lane in that moment, they're not just going to go straight into your lane like a crazy person. Not saying it doesn't happen, but I think waiting until it actually enters your path is appropriate and more natural IMO.
Andrew: Yes. It's difficult. If it is too cautious, then you get a bad passenger experience. I suspect the "lots of data" and TIME is going to really matter for things like this. (Or I hope so, because having to hard code solutions is somewhere between bad and risky). Maybe the full blown occupancy network, coupled with training and Dojo can really make a difference for issues like this.
Unfortuntely in my experience v11 is just like v10 in this respect. I don't know why it acted like this in the video but in actual driving for me it still freaks out and damn near slams the b rakes whenever a car even thinks about entering my lane.
IMO, if you're going to have someone roll/push an obstacle onto the road, they need to be HIDING behind something because with FSD detecting the person, it's going to default to the best way to avoid hitting the human.
HOLY CRAP that first test with the truck made me yell out loud.I was terrified how close your car was. I do not have the cohones for that, more power to you!
I find the turn in test fascinating because in my experience with the beta, anytime another car even hints at entering the road, or hell is even just sitting in the media pointing at the road, FSD Beta for me loves to slam on the brakes.
I love these obstacle tests, and damn do i respect you for putting *both* your cars at risk for them. Thanks AIDriver, and (respectfully) thanks Carol!😉
9:18 Guzzler didn’t want to enter a road, as it‘s correctly interpreted by the path planner, so no intention to stop or going slower. The guzzler is „just“ standing in the road. So, if this would be an official entrance of road, then it would slow down, I guess.
Insta360 is sponsoring you because your use of their camera is off the charts great. I would not be surprised to your video clips in their sales material.
You and Carol are amazing for doing these tests. It's so cool to see how FSD reacts and I was pleasantly surprised how it did. I almost wonder if the other vehicle needs to be perpendicular (or nearly) to be considered. You'd expect how much they're in your lane to matter. Hopefully with the next major version.
Hillary Thornton: Could be time re a benchmark. But there's no real world danger beta gets "too good" in truly complex situations any time soon. Simple situations with good weather and pretty good 99ish percent of the time, sure, we're there. But that's not nearly good enough for robo-taxis.
Fantastic video and keep them coming. Really appreciate your narration and use cases showcased during your videos too. Very informative. + It’s time to take wide out to a nice dinner for her support. 😊💪☕️🤠🚀🚀
Don't compliment her too much bcz I'm watching you 😂😂😂 I think the Beta reacts to a vehicle crossing over the lines into your lane IF said vehicle also is on a collision course with the Beta vehicle. If it's just in your lane with room for you to pass and it's path is not heading to cross yours Beta will continue as normal. By the way. Great vid and thanks for all the effort. Your vids are always fun!
The truck tests are interesting. It would be really easy to train FSD just for a car turning into your lane , but what makes it hard is that in vast majority of cases cars don’t do that. Driving is crazy, everyone is going 50MPH in opposite directions with like 1-2 feet space in-between. We all accept this and expect oncoming traffic to behave. It’s not acceptable for a Tesla to suddenly slam its breaks because an oncomer decide to swerve a little in their own lane. So to a human this seems easy… but it’s actually extremely difficult. You have to have essentially 0 false positives, and 0 false negatives. (where a positive is where it thinks there’s a car coming into your lane, a false positive is where it thinks a car is coming into your lane but it actually isn’t) Right now it is virtually 0 false positives and a few false negatives. Since normal case is that a car doesn’t suddenly turn into your lane, it makes sense Tesla is approaching getting the 0:0 false ratio from the direction of 0 false positives.
Carol knows what we want. 👍👍👍 Carol! And you, for your courage! I don't understand people who say "FSD will be sooo safe". I have avoided a couple of times really tiny* objects on the road, that were maybe only 2 inches high but very dangerous like metal rods or sheet metal parts or nails on parking lots. Avoiding quite small but extremly deep pot holes also prevented a lot of damage to my undercarriage. And if not dangerous I go around even very small animals like toads, squirrels or small cats. Who am I that I could decide that small animals are worth nothing? I safed my life also once by avoiding a very, very unexpected object falling from very high in front of me, I could because I saw it starting to roll. That all needs understanding the whole situations and what these objects really are and what they will do seconds later. FSD will never calculate how a small animal typically will behave or if construction objects can start rolling or not, FSD will not be able to distiguishe between the behaviour of a tractor or digger etc. etc.
an example: the second biker predicts something, FSD would not be able to predict it because it has no clue what a straw bale is. This situation at high speed and a much bigger trailer happened to me on a country road and the bale came from very high where FSD anyway ignores all objects (except traffic lights). ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-eB3rrvfvf2w.html
9:45 I think there must be some sort of inflection point involved… Like it allows the other driver to mess about within their own lane, under the assumption that the other driver will typically choose to self-correct and/or not leave their lane. But as soon as they cross over the Tesla is ready to react
Wow! I will be too scared to do what you did. Thank you, thank you! BTW, I really like your steering wheel cover that covers the steering wheel pretty completely making the steering wheel a bit fatter. Please tell me where you get it from. Since my fingers are suffering from arthritis, I can’t really wrap my fingers around the steering wheel on my Model Y.
Thanks for all the testing. I believe the software is set to recognize the blue truck icon once it goes across to the wrong side of the road/street. Also, you probably haven’t tried to test the head on ability while driving both vehicles at a similar speed towards each other.
On the truck entering lane scenario, I wonder if it was going slow enough that FSD expected the truck would not hit. Anticipating that the driver going so slow would see and stop.
I assume that just objects that potentially cross your own lane are displayed in blue. So if a vehicle protrudes slightly over the lane line and there is enough remaining width, but the lanes don't cross, then it stays white and FSD uses the remaining lane width to create enough clearance.
Zedus: But when it's OVER the line, ignoring it is just wrong. There needs to be awareness and balance, re managing risk from bad drivers. It's something humans (who pay attention) do well.
@@rogergeyer9851 Just because it's not blue doesn't mean it doesn't get attention. I referred only to when it is displayed in blue. The coloring is primarily for the driver, who then also sees that this vehicle is crossing his line. If a vehicle is slightly above the center line because it might want to turn but does not want to block the entire lane for following traffic, then a human driver will also drive past it, assuming that the vehicle will stop there. At least that's my assumption based on the videos I've seen so far.
The video I've been waiting to see! I still think the issue is the need for more compute for training the self labeling tasks. I'm sure FSD is recognizing "something in the way" and measuring position and velocity. But it's not recognizing "what" that something is and what to do about it. Hopefully, Dojo improves these results when it comes online.
@Ed Wells FSD doesn't have to know what it is. Just where it is, and where it's going. Actually, I don't think these tests are especially challenging. FSD does many harder things very well. I suppose these things simply hasn't been a priority yet. There are a lot of YT videos with wheels on the loose. So eventually FSD will learn to avoid them, and cars trying to crash into you too. Sometime in the future we will have demolition rallys with only Tesla FSD. And all cars are coming out of it unscratched!
@@larsnystrom6698 Shouldn't it react differently to a paper bag blown around compared to a box, a balloon compared to a baby, etc? I suppose, I agree that if it does not know what it is, it should be cautious. But then we potentially get a lot of manuevers for leaves, snow, coloured roads, etc.
@@robertwhite3503 That's my unspoken point. There are things FSD should train to avoid or take action and things it should not. The list is virtually endless, so it'll take a huge effort to label everything and train the reactions. That includes all the various lighting and weather conditions (e.g. something that affects the tests with the box in this video). I don't believe this is impossible, just a matter of computing time and priority.
Ed Wells: Hopefully the same sort of thing can lead to it actually being able to READ many road signs. Because to handle things like construction or unusual situations well, that's a rather critical skill for it to approach practical wide network robo-taxi level.
@@rogergeyer9851 However, I do not believe Tesla has any plans for OCR. It recognises a 20mph sign and a 30mph sign in the same way you recognise a cat and a dog. Providing there is some standardisation of signs that may be enough. Also as a human we recognise a construction zone before we can read the signs. Hopefully no hand-written ones.