In South Africa, Speed Bumps usually have white "triangles" painted on them and usually have red chevrons (red reflectors on a pole about 3 foot high) on either side of the road where the speed bumps are, so can be seen reasonably easily. I assume this is a UK / European thing as South Africa follows those types of road signage.
12:34 there is no issue with not stopping. You have 0.5 seconds after the light turns yellow to be in the intersection. No laws were broken or in serious risk of being broken
I'd be happy at the end of the year to still be at Level 2, just no nag wheel and fixes for easy things like moving into the wrong lanes and smoother steering.
Just drive through mountain road at night 2 days ago and fsd did not able to stop the road closed barrier at 50 mph , almost get crashed if i haven't press the break evevn the emergency break didnt kick in .however the car was able to react and stop at 45 mph but not 50 or 55 more . It was very scared, please fix this tesla !
11.4.4 has improved since getting it on mine. Initially, there was this spot that it would "[stop] for traffic control" that has never been there in the nearly 15 years I've used that road. After a week or so, it no longer tries to stop there; it just "[changes] lanes to follow route." The only thing is that it didn't stop for that in the 11.3 branch...
I’ve experienced changes between updates. I don’t know if it’s from navigation updates or hidden feature toggles, or different weather. Sometimes it’s better, sometimes it’s worse.
The car itself does not learn anything from you except your driving speed as it pertains to navigation time estimates. The car frequently uploads files from events when parked & on wifi.
A neural net shouldn't be able to change unless they retrain it with new weights. But I have also noticed that it feels like things have changed on a single version. My belief is their patching things with Tesla map data until they can solve it with the neural net.
BTW at the end, be careful pointing your hand at some car that had to wait too long for you, there might be some aggressive dude inside, who will turn around.
Biggest issue for me continues to be car moving into lanes it thinks are driving lanes but are for only turning or going around turning vehicles. Also roundabouts continue to be a dealbreaker.
15:00 It isn't just about how much space you technically have. It is how it creeps forward that matters as well. Even if Tesla could guarantee 100% perfect detection of road edge, most people can see traffic better than the Tesla repeater cameras. This means that oncoming drivers are going to be on edge wondering why you pulled up so far, when you could have stopped sooner and seen them coming.
2 other mistakes that you failed to mention; 1 it drives over the solid white line, going on to the shoulder when making a right turn. 2 it is illegal to use the center local left turning lane (marked with yellow lines) to get to the next street left turning lane (marked with white).
In the real world, with a MASSIVE robo-taxi network, there will be LOTS of "unusal" situations like construction, people parked weirdly, the unexpected, bad weather, etc. etc, that must be dealt with. And the fare paying taxi rider is NOT going to want to sit in the driver's seat and act as a backup driver and be legally responsible if anything happens, re an accident. So it's making progress, incrementally, over time, which is GREAT. But we're a LONG away (like at last several years) from reliable massive practical robo-taxi networks collecting massive amounts of fares daily -- until PROVEN otherwise. And great beta testers like Chris and many others in lots of areas testing lots of things, clearly demonstrate that -- again, until proven otherwise. I like Chris, think robo-taxis will happen, but Elon constantly saying non-credible things isn't helping anyone, as it hasn't been since 2016 or so.
It is strange that my M3 2020 model does so much worse with its simplest self driving package. The car is updated. Alot of phantom braking and unnecessary warnings. Have you thought about the differenses in the packages and the performances?
4.4 is still pretty much poop! 4.3 probably 3 times better. It does dangerous things almost every day I drive it...also likes to try to randomly exit a highway for no reason, sometimes doesn't get in the highway entrance lane, and roundabouts are still horrible (always have been). It's about time for Tesla to raise the price, lol.
Chris, if Tesla can get it's cars to obey traffic driving laws/signs etc 100% plus the human intuition for driving situations they are surely there? - do you have a definitive list of goals Tesla must master before level 5 autonomy is reached in your opinion or an official list?
I'll give you a few things I see as currently missing from FSD that are required to be actually autonomous. 1. Correctly reacting to emergency vehicles with lights on (as Chris mentions in the video). 2. Correctly reacting to a school bus stopped with flashing lights (this may vary by state, but in most places you are not allowed to pass the bus). 3. Correctly reacting to road closed signs (currently FSD just wants to drive around the sign / barrier if there is room). 4. Correctly dealing with roundabouts (sometimes it does well, but too often it still stops at the yield sign when it shouldn't). 5. Correctly merging when getting onto the freeway. Currently it waits until it is forced to merge, which may be correct when traffic is backed up and moving slowly but is not correct or safe when traffic is flowing at the speed limit. 6. Probably the hardest one to get right, it needs to be able to correctly interact with a person directing traffic. 7. Bonus item, in Texas it is the law on the freeway that you must slow way down or change lanes out of the right lane if there is a police vehicle on the shoulder with its lights on. This doesn't apply to all states, but it's the law in Texas. Elon can say it will be ready by the end of the year all he wants to, but if all of these items are not currently on the list of things to train FSD for in the next 6 months then FSD will not actually be ready by the end of the year. None of these things are edge cases, there are still actual edge cases that will need to be dealt with also. Elon has said for years that it will be ready at the end of the year, but obviously they have not even started training FSD for several of these situations so that was not ever a realistic possibility in years past. If they don't have all of these items on the todo list for this year, then it's also not possible for FSD to be ready this year (I don't know if they're on the TODO list for this year or not, I'm just saying that they were clearly NOT on the TODO list for years past despite him making the prediction in years past).
I agree with all your examples where FSD is currently not able to complete these tasks effectively - these will need to be trained and solved, it won't be ready at level 5 for maybe 2yrs imo, that said recent advances in AI has potentially boosted the decision making progress of FSD together with the ever increasing fleet (for data) and the Dojo supercomputer Neural Net to process the algorithms to solve these fundamental issues.
@@Deviloc1I think a lot of this stuff was likely on their to do list in years past, but that they just never got to it. They have mentioned they kept hitting the local maximum of where each version of the software would actually improve to, so I think they never got the chance to start addressing more specific issues/details like they have been able to more recently
12:30 stop for the yellow. When I make that decision I peek at the rear view mirror, if a car is behind me it raise the chance I''ll go thru. Speed and distance between me and the car behind play into the calculation too. Was there a car behind you?
Car and Tesla. By the way, I have difficulty communicating because I had a stroke in Broca’s area, the part of the brain that controls speech. 2/8/2021 but I lived again. (My wife helped me compose this.)
Tesla must have tons of data on where the car never needs intervention and could probably force FSD to take routes which have a super high success rate if it wanted.
No. This is a horrible idea. FSD needs to drive on roads it normally doesn't so it can learn. If everyone only used it on roads it's flawless with then it will never improve.
No need for it to be 5x safer. In theory, 1.01x safer is all we need, but maybe we aim for a minimum of say 1.5x safer initially so that we have a margin for error until we have 10s of billions of km to really prove that it's safer. That said, even today the failure modes are mostly not safety related. P.S. Things like emergency vehicles are "easy" (in the context of the brilliant minds of Tesla) to solve. The issue is that it takes compute time away from general purpose training. I suspect Dojo and/or additional Nvidia compute capacity will allow them to solve this. Same with windshield wiper issues, auto-brights, and auto-parking. I'm betting that they generally know what to do but are compute constrained.
On I-287 in NJ this evening, I was behind a white Model Y with large lettering on the back that said this was an Autonomous Vehicle and to keep your distance. There was nobody in the vehicle and it was moving along with traffic at a good clip, between 75-80 in a 65 mph zone. A new phase in testing?
Do you have it ever need to go left but then it tries to get in the right line or middle lane? The Driving directions voice announces it correctly, but the FSD doesn't go to the correct lane. Had that happen 2x Sunday. That car was on 11.3.6 I think tho. Probably decently better in 11.4.4 or newer? It was generally pretty good in 11.3.6!
"I'm not complaining". You kind of are, and you kind of should more. The repeated failures, if nothing else, should have improved his terrible ability to make predictions. But it doesn't, which makes me think it's wilful.
Just keep speaking the truth. The problem with the “community” is there’s a huge amount of kool-aid drinkers. They say everything is an edge case. They get delusional and defensive really fast. Its scary actually.
the car needs the full set of sensors, GPS and lidar in addition to the HD cameras if it wants to go FSD quickly..Once the car AI has learned how to drive independently the redundant sensors and lidar can be removed and the functionality of these systems transferred to the cameras
New DC drives are ultra efficient at speeds and the batteries are ridiculously the weakest part of this equation, New zero point drives are on their way unfortunately the Energy companies are looking at owning everything including the free Zero point energy bonus waiting in the wings I have seen these fission drives and they can be built around a harmonic containment vessel with unlimited power at virtually no costs as the materials are both cheap and abundant since silica and enclosures can be readily manufactured and a 50 Kilowatt power station is smaller than a microwave and can literally run on urea right from the driver as a source of fuel. The Chinese are looking at cobalt high cap batteries and the silica based charged slag for roadside charging and flowed energy is molten sand with a oxide embedded allowing the direct vamping of energy from the slag actually we already perfected these drives and it was always in the works since before 1994. EV tax grabs and taxpayer welfare needs stopped and Biden, Trudeau need to get the hell out of car manufacturing as they are stealing tax dollars in the name of carbon sequestering while buying options with their trust funds. These Globalists need stopped and the free world should demand these Globalists be arrested and charged with War Crimes while they have their patents stripped and licensing revoked as they have misled the public the theft of public funds while lying about the content of vehicles. The car industry has lied about everything including the actual costs of cars and trucks while pricing customers out of driving while padding their pockets they are war criminals as they have stolen future vehicles to reflect no energy on the grids while using subsidies from government and criminal actors in the power distribution business.
of course it's all speculation, but I agree with you. I don't think the Tesla thought the caution crosswalk signs were stop signs because the screen visual didn't show stop signs. It does, however, show a crosswalk visualization
The legalities are the problem with FSD...eg. who is responsible if someone is killed and the Tesla is at fault. Who is responsible for traffic violations.
One thing FSD is doing better on highway is keeping me center mostly while steering right. It also corrected my very old habit of driving to much on the right edge of the road. It did scare me at the begging mostly on construction roads with jersey’s.
I live in Chapel Hill NC and there are three different intersections that 11.4.4 has tried to run red lights at for me. In general, they are intersections where the road curves in a sweeping way approaching the light so the light is off to the side and not straight in front of the car as the car is approaching the intersection. Don't know if you've gotten recommendations for looking for that kind of a stoplight or not. Clearly a safety issue.
I think you should mostly be making videos where FSD fails or does badly. Otherwise, how can it improve? Just stress what you are doing and state most of the time it is fine. And all FSD drivers need to report all issues. Flying over speed humps and ignoring the signs, hitting pot holes, school zone issues, double stop signs, obscured stop signs....
90+ percent of my disengagements are for "courtesy" and "comfort" reasons. Rarely are they for a serious issue. There are of course the occasional disengagements because it doesn't understand a sign or road marking. I really wish it had some audible indicator of when it is about to turn left. Let's hope that Dojo will bring in some exponential learning soon.
Look at the route planner at 4:48 when it goes through the intersection. Looks like it was considering driving into the oncoming traffic and making some kind of weave maneuver. Lot of phantom cones appeared in the visualization as well
A friend recently got a model Y with HW4. He purchased FSD Beta, but apparently it is not available on HW4. It's a huge problem for people who got a vehicle with HW4 and purchased FSD beta!
Every once in a while there is a giant discovery in machine learning that changes the landscape. The most recent one in the news being the transformer architecture. I think we need 1 or 2 more big discoveries like that to get FSD without supervision.
Today I saw a video about "liquid NNs". (ua-cam.com/video/0FNkrjVIcuk/v-deo.html) It seems like a huge step as well. It would be interesting to see FSD using this type of neural net. Although it would probably mean training it from scratch...
@@Ejckej The training part isn’t hard. It takes probably a week or so. It’s the designing, validating, hyperparameter optimization, data labeling, etc. that takes forever
I had a very troubling experience yesterday following 2 motorcycles. The closest to me once in my FSD view was swirling so my MYP was doing the same even going over a full yellow line even if a dump truck was coming on the opposite way! 😮 I tested it 3 times. It was on a secondary road with relatively fresh lines. Besides that, I drove a blend of highway, backroads and city with a few phantom breaks and over cautious with 4 corners. I find it very slow to continue after a stop. Why take 3-5 seconds when no one left or right and mostly while an anxious driver is behind? In Canada, we call these quick stop and go behaviours “an American stop” 😊 so Elon stop being a grandpa😂
Of course, speed limit signs still need a lot of work. Short 25MPH (the car thinks) in the middle of a 45 speed zone. Happens very frequently here in Mississippi. Another thing that causes me to take over is the creeping forward on narrow roads which freaks out cross traffic, so I often have to brake myself. I also get inconsistency in behavior (like your blind lane to the right) in various scenarios. Errant & tardy turn signals, although not serious are annoying to other drivers as well as the driver. Overall I'm very pleased with the progress of FSD Beta.
I also have FSD think speed limit is 25 in a 40 zone. Seems a lot more frequent with recent software. Yet another thing to fix "by the end of the year" to get the car to drive the entire trip.
Perhaps it is now, becasue the system fails too often, but due to automation complacency, the system + human is not safer than system alone if it gets better (say 100-1000 miles per DE).. Google "Disengagements: Wrong Metric for AV Testing"
I agree. What is great about it is that the car does so well, there wouldn't be much to watch without listening to Chris' thoughts. I just wish FSD would work as well for me, and to handle more of the unusual cases properly.
I just made an interesting observation: the FSD visualisation is not based on just the current environment, modelled by the car "looking around". Just check the environment and its visulatization at 0:59. The upper left corner of the screen is showing a road which is totally blocked by the building on the corner. I assume some "network" effect here, probably another cars have already explored that part of the road and this Tesla somehow benefits from that shared knowledge.
The visualization is based on the car looking around and map data; the car knew there is a street behind the building cuz of the map data. Teslas don’t talk to each other yet, from our knowledge but would be a cool improvement
@@nick0780they work in unexplored territory, but even a human needs some context from a map to drive smoothly and confidently in a new area. I can only guess that at some point the cars will “remember” their frequent routes, as you or I would, but they will also use the fleet of cars to continually get accurate coarse map data for pretty much everywhere. Would be pretty interesting if Tesla some day sold a public map that was kept continually up to date using the cars
Map data even overrides the vision system. I have two intersections where the map data says you can't go. The visualization clearly shows a path but the car still won't take it.
One huge problem I see with driverless cars (without a steering wheel!) are certain edge cases. For example, I think everyone knows the videos where a gang just stops a car and tries to rob them or whatever. In most of these videos, the drivers just set back a bit, go full throttle through that situation. Just imagine your sitting in your Tesla without a steering wheel, it of course stops for these people, and you're ABSOLUTELY trapped there - no way to escape.
Hey Chris! I just drove my Model Y from California to Michigan last week. I will say the car did struggle more in Michigan on FSD compared to California. But Michigan has AWFUL roads. But my dad might see you at the chargers at Meijers in Ann Arbor keep an eye out for a matte white model y with California plates lol
I don’t believe it can truly reverse. I think there have been cases of the car stopping and sort of rolling back with momentum, but nothing with it actually reversing using the motors
@@SyntheticSpy ua-cam.com/video/c1SpE18GO4o/v-deo.html this video seems to show it reversing at an intersection. The driver says that the road was slightly downhill so it couldn't be just rolling back.
Local regulation aside, do you think the system is going to be able to handle road systems in other countries? e.g. in the UK we don't generally have grid based cities that were built with vehicles in mind which leads to a lot of twisty, narrow streets and mixed lane traffic with a preference for roundabouts instead of intersections. Always feels like that when Tesla is talking about FSD progress it's really just the US market. I guess the exception would be highway driving which is pretty similar everywhere and works very well. Loving the content and seeing how the system is improving.
I was a person who sent some to you, but you didn't really go the way I suggested (no big deal). I suggested going up 7th from stadium and then turning left on Huron. In your drive, had you wanted to turn right from liberty onto south seventh rather than turning left the bike lane that shares the right causes the issue. You mentioned it but the route went left so it was not an issue for you. And then I believe you turned right on Huron rather than left, which is no big deal for the car. Also you turned left from beakes onto 4th I think instead of main and so forth, so of the 8 or so spots that I have had issues with I think you only tried 3 (the double red light and the flashing yellows and the passenger crossing bollards), but it was still nice to see.
You could try finding one of those crosswalk posts for which is stops or really slows down and then cover the mini-stop with white tape and try again to see if it makes any diffenrece. Then, you should remove the tape. 😉
I still see it as difficult as in South Carolina you have to yield on a huge amount of RIGHT turns with random inconsistent signage that generally does warn you in time
Thanks for looking for and testing more edge type situations these are the challenges facing FSd and until most are fixed; it can't be closer to level 5; if there isnt' attention to them; why woud they get fixed? So this is better than running on the same roads. Thanks
As someone buying tesla stock every single week i hope they take another few years for FSD to be complete lol. Stock price is going to be insane when ever you can sleep in your car and it takes you to work
The yellow light at 12:32 had a car in the opposing left turn lane waiting to turn. I would have stopped at the yellow not only because it had been yellow for awhile but also the car waiting to turn was committed and out in the cross traffic lane so ego should have stopped to let that car get out of that situation.
You did not notice that the Tesla creeped into the bicycle lane at the light, which had there been a trike or group of bikers behind you, they would have been pushed into the curb. Also, real world situations include inclement weather, construction lanes which will not be handled by Elon's prediction of Level 4. I'm guess his interpretation is Level 4 when things are ideal! Ha! The random wrong turn signal is a accident waiting to happen, I get it 10% of my drives. Getting better, but these unique situations will always make me handle FSD like a student driver, especially with my wife in the passenger seat gripping the side door handle and gritting her teeth.
I agree for the bike lane, and I have the same experience. At red light beta tends to creep into the bike lane, making it unsafe for bikes coming behind. I can't say how it does if a bike was already visible, but going into the bike lane and stopping there is not something I like (and a potential for road rage incident).
Locally, i've had serious issues with .4.4 which weren't there or were "corrected" in .3. Right now, the car always tries to moves/pass to the right. The bus stop deviation is back, it is back into bike lanes and right-only lanes which are obviously and clearly identified. Highway lane change is less smooth than .3 as well. such a weird release.
I agree with you it is an obsession with going right. on the interstate, It wants to take almost every exit it can find. It really seems like it wants to go right, unless of course it's sitting in the passing lane and it refuses to get out of it, unless they are on my bumper. In town It's wonderful though.
I'm curious: When the car does a roll-back, is it because it's on an incline, or is it because the CPU realized a little too late that it was stopping too far forward, so it reversed voltage to the motor for a slight (safe) roll-back?
I've noticed that odd rollback a few times, too. Mine usually occur after moderate to heavy stops, no incline, and well before a stop line. It probably only rolls back 2-3 inches at most, but it feels like 2-3 FEET. :)
8:35 the average driver has no idea what to do when there's an emergency vehicle. Some drivers stick to their route and ignore it. Some change lanes in front of them. Some put themselves right behind them to take advantage of the higher speed through dense traffic. And if parked on the side, some crash into them and some drive just very close by. So I think they may just trying to figure out what's the best approach here.
What I would really like to see is for Tesla to merge the AP and FSD forks so that AP uses the FSD code. They can keep AP features exactly as they are, only do lane keeping, not honor lights or stop signs, not make turns, etc. but just use the FSD code to do the lane keeping for AP. This would correct a lot of improper lane maneuvers, phantom breaking, etc. related to the old AP code.
@@lowmax4431they did it for people who have FSD Beta, but not for the standard autopilot that comes with every car. I think they are probably waiting until highway FSD is pretty much flawless before replacing autopilot for everyone, maybe even allowing for level 3 or higher on freeways when they do so
I still don’t understand why so many of us are still on FSD 11.3. I’ve been a FSD beta user since ver 9 and honestly I’m disappointed by the state of FSD. It works great in most situations but then it does something that makes absolutely no sense.
Apparently there is some background data related to safety, and 11.3.6 has a higher safety profile vs the 11.4 branch... so as soon as safety is above the acceptable threshold, mass updates will resume. Take it with a grain of salt.
I just want FSD to be there for me in an emergency like I'm impaired (medically or otherwise) and I need it to drive me home or to the nearest emergency room. I will always be paying attention as much as possible and hope it anticipates better and isn't distracted like the average driver. I would love for FSD to allow us to practice and pre-plan such routes like "take me to the grocery store" or "take me to my doctor's office", etc - maybe 10 pre-planned routes including it parking itself by looking in a pre-planned parking lot. That said I just don't see how Tesla will ever be able to avoid a huge pothole because right now I don't see it seeing those and even I sometimes have trouble distinguishing between a shadow, or a recently fixed pothole and a real pothole. Have you shown any pothole navigation in Michigan where they organically grow their own pothole monsters?
It seems like Tesla is prioritizing general safety and decision making before addressing smoothness things like speed bumps or potholes, but of course none of us have inside info. One thing that we have to remember though is that what is shown on the visualization is a separate thing to what the car sees. The car may be detecting potholes, maybe in a sort of practice mode, and just not acting on them as far as we know
Compare FSD to average driver, drivers who are usually impaired in some way, vision impaired, distracted drivers looking at smart phones, drug / alcohol impaired drivers, old drivers who get confused or range of motion impaired, young drivers high and speeding. so consider that FSD is a lot better than the "average " driver, with so many potential impairments. Thanks Great Video !
is Tesla using driving scenes from normal drivers to train its software? Since that would be a huge database to learn from (situations and driver reactions).
They do, but in a very smart way. They query the fleet to collect the requested situations in shadow mode. So this data is heavily filtered, but very high quality. But even with all the fleet doing heavy filtering, they have way too much data. They really do need Dojo (or equivalent) for the training.
I love to use FSD but the only issue is FSD don’t recognize road bumps!
Hard problem - difficult to distinguish from shadows or patches
In South Africa, Speed Bumps usually have white "triangles" painted on them and usually have red chevrons (red reflectors on a pole about 3 foot high) on either side of the road where the speed bumps are, so can be seen reasonably easily. I assume this is a UK / European thing as South Africa follows those types of road signage.
@@bartwaggoner2000I don't think it's hard, they just prioritize safety first and then comfort
It detects them most of the time for me. It seems to require a pretty high confidence threshold though.
The fact that you complain about speed bump it’s means tesla is getting better
You could tape over the small stop signs on the poles then try to drive past them.
12:34 there is no issue with not stopping. You have 0.5 seconds after the light turns yellow to be in the intersection. No laws were broken or in serious risk of being broken
Fsd don’t stop at double red lights in my city
Be interesting to see if a small STOP sign, like a bumper sticker, would affect FSD
15:00 It's fine. If their tires are no longer visible, it's getting close :P
I'd be happy at the end of the year to still be at Level 2, just no nag wheel and fixes for easy things like moving into the wrong lanes and smoother steering.
Just drive through mountain road at night 2 days ago and fsd did not able to stop the road closed barrier at 50 mph , almost get crashed if i haven't press the break evevn the emergency break didnt kick in .however the car was able to react and stop at 45 mph but not 50 or 55 more . It was very scared, please fix this tesla !
11.4.4 has improved since getting it on mine. Initially, there was this spot that it would "[stop] for traffic control" that has never been there in the nearly 15 years I've used that road. After a week or so, it no longer tries to stop there; it just "[changes] lanes to follow route." The only thing is that it didn't stop for that in the 11.3 branch...
Chris, what are your settings? Chill/avg/assert, etc...thx! Love the channel...
Impressive. Does the software improve in between updates?
I noticed this today as well
I’ve experienced changes between updates. I don’t know if it’s from navigation updates or hidden feature toggles, or different weather. Sometimes it’s better, sometimes it’s worse.
The car itself does not learn anything from you except your driving speed as it pertains to navigation time estimates. The car frequently uploads files from events when parked & on wifi.
It seems to... but hard to prove.
A neural net shouldn't be able to change unless they retrain it with new weights. But I have also noticed that it feels like things have changed on a single version. My belief is their patching things with Tesla map data until they can solve it with the neural net.
BTW at the end, be careful pointing your hand at some car that had to wait too long for you, there might be some aggressive dude inside, who will turn around.
Biggest issue for me continues to be car moving into lanes it thinks are driving lanes but are for only turning or going around turning vehicles. Also roundabouts continue to be a dealbreaker.
Yeah I complain about that a lot on Twitter lol
I've seen sharp reduction in bad lane shifts in two weeks, to almost none.
15:00 It isn't just about how much space you technically have. It is how it creeps forward that matters as well. Even if Tesla could guarantee 100% perfect detection of road edge, most people can see traffic better than the Tesla repeater cameras. This means that oncoming drivers are going to be on edge wondering why you pulled up so far, when you could have stopped sooner and seen them coming.
2 other mistakes that you failed to mention; 1 it drives over the solid white line, going on to the shoulder when making a right turn. 2 it is illegal to use the center local left turning lane (marked with yellow lines) to get to the next street left turning lane (marked with white).
In the real world, with a MASSIVE robo-taxi network, there will be LOTS of "unusal" situations like construction, people parked weirdly, the unexpected, bad weather, etc. etc, that must be dealt with. And the fare paying taxi rider is NOT going to want to sit in the driver's seat and act as a backup driver and be legally responsible if anything happens, re an accident.
So it's making progress, incrementally, over time, which is GREAT. But we're a LONG away (like at last several years) from reliable massive practical robo-taxi networks collecting massive amounts of fares daily -- until PROVEN otherwise.
And great beta testers like Chris and many others in lots of areas testing lots of things, clearly demonstrate that -- again, until proven otherwise.
I like Chris, think robo-taxis will happen, but Elon constantly saying non-credible things isn't helping anyone, as it hasn't been since 2016 or so.
It is strange that my M3 2020 model does so much worse with its simplest self driving package. The car is updated.
Alot of phantom braking and unnecessary warnings.
Have you thought about the differenses in the packages and the performances?
4.4 is still pretty much poop! 4.3 probably 3 times better. It does dangerous things almost every day I drive it...also likes to try to randomly exit a highway for no reason, sometimes doesn't get in the highway entrance lane, and roundabouts are still horrible (always have been). It's about time for Tesla to raise the price, lol.
Love your videos. Please Keep them coming.
Chris, if Tesla can get it's cars to obey traffic driving laws/signs etc 100% plus the human intuition for driving situations they are surely there? - do you have a definitive list of goals Tesla must master before level 5 autonomy is reached in your opinion or an official list?
I'll give you a few things I see as currently missing from FSD that are required to be actually autonomous. 1. Correctly reacting to emergency vehicles with lights on (as Chris mentions in the video). 2. Correctly reacting to a school bus stopped with flashing lights (this may vary by state, but in most places you are not allowed to pass the bus). 3. Correctly reacting to road closed signs (currently FSD just wants to drive around the sign / barrier if there is room). 4. Correctly dealing with roundabouts (sometimes it does well, but too often it still stops at the yield sign when it shouldn't). 5. Correctly merging when getting onto the freeway. Currently it waits until it is forced to merge, which may be correct when traffic is backed up and moving slowly but is not correct or safe when traffic is flowing at the speed limit. 6. Probably the hardest one to get right, it needs to be able to correctly interact with a person directing traffic. 7. Bonus item, in Texas it is the law on the freeway that you must slow way down or change lanes out of the right lane if there is a police vehicle on the shoulder with its lights on. This doesn't apply to all states, but it's the law in Texas.
Elon can say it will be ready by the end of the year all he wants to, but if all of these items are not currently on the list of things to train FSD for in the next 6 months then FSD will not actually be ready by the end of the year. None of these things are edge cases, there are still actual edge cases that will need to be dealt with also. Elon has said for years that it will be ready at the end of the year, but obviously they have not even started training FSD for several of these situations so that was not ever a realistic possibility in years past. If they don't have all of these items on the todo list for this year, then it's also not possible for FSD to be ready this year (I don't know if they're on the TODO list for this year or not, I'm just saying that they were clearly NOT on the TODO list for years past despite him making the prediction in years past).
I agree with all your examples where FSD is currently not able to complete these tasks effectively - these will need to be trained and solved, it won't be ready at level 5 for maybe 2yrs imo, that said recent advances in AI has potentially boosted the decision making progress of FSD together with the ever increasing fleet (for data) and the Dojo supercomputer Neural Net to process the algorithms to solve these fundamental issues.
@@Deviloc1I think a lot of this stuff was likely on their to do list in years past, but that they just never got to it. They have mentioned they kept hitting the local maximum of where each version of the software would actually improve to, so I think they never got the chance to start addressing more specific issues/details like they have been able to more recently
@@Deviloc1 Add to the list:
it needs to be able to drive onto and off of a ferry landing as directed by the ferry deck hands
How about they master level 3 first?
Fsd beta is awful, I would easily accept my money back. Mine is horrible. Unusable.
Can you tell me what year Elon didn’t say FSD end of this year? He been saying this every year for past 5 years i think.
12:30 stop for the yellow. When I make that decision I peek at the rear view mirror, if a car is behind me it raise the chance I''ll go thru. Speed and distance between me and the car behind play into the calculation too. Was there a car behind you?
Nope
This is a joke of a test, bring it to Boston in the winter time
Keep up the great work!
I think FSD is going to be a first class driver aid. Level 4 or 5? Not in my lifetime.
Car and Tesla. By the way, I have difficulty communicating because I had a stroke in Broca’s area, the part of the brain that controls speech. 2/8/2021 but I lived again. (My wife helped me compose this.)
Tesla must have tons of data on where the car never needs intervention and could probably force FSD to take routes which have a super high success rate if it wanted.
No. This is a horrible idea. FSD needs to drive on roads it normally doesn't so it can learn. If everyone only used it on roads it's flawless with then it will never improve.
@@HaikHakopyan who said everyone would do it? The route would be way longer
No need for it to be 5x safer. In theory, 1.01x safer is all we need, but maybe we aim for a minimum of say 1.5x safer initially so that we have a margin for error until we have 10s of billions of km to really prove that it's safer. That said, even today the failure modes are mostly not safety related.
P.S. Things like emergency vehicles are "easy" (in the context of the brilliant minds of Tesla) to solve. The issue is that it takes compute time away from general purpose training. I suspect Dojo and/or additional Nvidia compute capacity will allow them to solve this. Same with windshield wiper issues, auto-brights, and auto-parking. I'm betting that they generally know what to do but are compute constrained.
On I-287 in NJ this evening, I was behind a white Model Y with large lettering on the back that said this was an Autonomous Vehicle and to keep your distance. There was nobody in the vehicle and it was moving along with traffic at a good clip, between 75-80 in a 65 mph zone. A new phase in testing?
Do you have it ever need to go left but then it tries to get in the right line or middle lane?
The Driving directions voice announces it correctly, but the FSD doesn't go to the correct lane. Had that happen 2x Sunday. That car was on 11.3.6 I think tho. Probably decently better in 11.4.4 or newer? It was generally pretty good in 11.3.6!
"I'm not complaining". You kind of are, and you kind of should more. The repeated failures, if nothing else, should have improved his terrible ability to make predictions. But it doesn't, which makes me think it's wilful.
Version 11.4.4. It goes into bypass lanes 95% of the time and goes into turn lanes unexpectedly 80% of the time. Need to fix!!!
Just keep speaking the truth. The problem with the “community” is there’s a huge amount of kool-aid drinkers. They say everything is an edge case. They get delusional and defensive really fast. Its scary actually.
the car needs the full set of sensors, GPS and lidar in addition to the HD cameras if it wants to go FSD quickly..Once the car AI has learned how to drive independently the redundant sensors and lidar can be removed and the functionality of these systems transferred to the cameras
Flagstaff AZ and Durango Co - Google maps sends people in circles - try FSD there
started viewing the video 5 seconds after ported :D
New DC drives are ultra efficient at speeds and the batteries are ridiculously the weakest part of this equation, New zero point drives are on their way unfortunately the Energy companies are looking at owning everything including the free Zero point energy bonus waiting in the wings I have seen these fission drives and they can be built around a harmonic containment vessel with unlimited power at virtually no costs as the materials are both cheap and abundant since silica and enclosures can be readily manufactured and a 50 Kilowatt power station is smaller than a microwave and can literally run on urea right from the driver as a source of fuel.
The Chinese are looking at cobalt high cap batteries and the silica based charged slag for roadside charging and flowed energy is molten sand with a oxide embedded allowing the direct vamping of energy from the slag actually we already perfected these drives and it was always in the works since before 1994.
EV tax grabs and taxpayer welfare needs stopped and Biden, Trudeau need to get the hell out of car manufacturing as they are stealing tax dollars in the name of carbon sequestering while buying options with their trust funds. These Globalists need stopped and the free world should demand these Globalists be arrested and charged with War Crimes while they have their patents stripped and licensing revoked as they have misled the public the theft of public funds while lying about the content of vehicles. The car industry has lied about everything including the actual costs of cars and trucks while pricing customers out of driving while padding their pockets they are war criminals as they have stolen future vehicles to reflect no energy on the grids while using subsidies from government and criminal actors in the power distribution business.
of course it's all speculation, but I agree with you. I don't think the Tesla thought the caution crosswalk signs were stop signs because the screen visual didn't show stop signs. It does, however, show a crosswalk visualization
The legalities are the problem with FSD...eg. who is responsible if someone is killed and the Tesla is at fault. Who is responsible for traffic violations.
One thing FSD is doing better on highway is keeping me center mostly while steering right. It also corrected my very old habit of driving to much on the right edge of the road. It did scare me at the begging mostly on construction roads with jersey’s.
Musk said that they were thinking about it wrongly and the answer is much simpler. Plus dojo capabilities.
Can you try to park in area in which the car initially needs to reverse a few meters before starting the route?
Can we download some no-official ad-ons / plugins inside the FSD beta for better driving ?
I live in Chapel Hill NC and there are three different intersections that 11.4.4 has tried to run red lights at for me. In general, they are intersections where the road curves in a sweeping way approaching the light so the light is off to the side and not straight in front of the car as the car is approaching the intersection. Don't know if you've gotten recommendations for looking for that kind of a stoplight or not. Clearly a safety issue.
I think you should mostly be making videos where FSD fails or does badly. Otherwise, how can it improve? Just stress what you are doing and state most of the time it is fine. And all FSD drivers need to report all issues. Flying over speed humps and ignoring the signs, hitting pot holes, school zone issues, double stop signs, obscured stop signs....
90+ percent of my disengagements are for "courtesy" and "comfort" reasons. Rarely are they for a serious issue. There are of course the occasional disengagements because it doesn't understand a sign or road marking.
I really wish it had some audible indicator of when it is about to turn left.
Let's hope that Dojo will bring in some exponential learning soon.
Construction sites on the country side were it's spots of bad short and long gravel due to work on the road...or just good gravel roads.
Look at the route planner at 4:48 when it goes through the intersection. Looks like it was considering driving into the oncoming traffic and making some kind of weave maneuver. Lot of phantom cones appeared in the visualization as well
Biggest issue for me is when it makes a left turn and it cuts through the opposite lane.
but if those cones did say "stop" on them.. well.. can you blame it?
Thanks for the video! Hope the optimism on Dojo is warranted.
yo, idk about you guys but I have no idea where your supposed to stop at that double red light
Great editing BTW - I know it takes a lot of effort: thanks.
Yeah..
So much can change for someone to predict
Did you say "Ah-ventually"?
Do any of your Teslas use Tesla vision over the USS?
Dan will provide suggestions…
A friend recently got a model Y with HW4. He purchased FSD Beta, but apparently it is not available on HW4. It's a huge problem for people who got a vehicle with HW4 and purchased FSD beta!
In a year it'll be a problem for us with hardware. 3. When we aren't getting updates. I would rather be in their shoes.
Every once in a while there is a giant discovery in machine learning that changes the landscape. The most recent one in the news being the transformer architecture. I think we need 1 or 2 more big discoveries like that to get FSD without supervision.
Today I saw a video about "liquid NNs". (ua-cam.com/video/0FNkrjVIcuk/v-deo.html) It seems like a huge step as well. It would be interesting to see FSD using this type of neural net. Although it would probably mean training it from scratch...
@@Ejckej The training part isn’t hard. It takes probably a week or so. It’s the designing, validating, hyperparameter optimization, data labeling, etc. that takes forever
I had a very troubling experience yesterday following 2 motorcycles. The closest to me once in my FSD view was swirling so my MYP was doing the same even going over a full yellow line even if a dump truck was coming on the opposite way! 😮 I tested it 3 times. It was on a secondary road with relatively fresh lines. Besides that, I drove a blend of highway, backroads and city with a few phantom breaks and over cautious with 4 corners. I find it very slow to continue after a stop. Why take 3-5 seconds when no one left or right and mostly while an anxious driver is behind? In Canada, we call these quick stop and go behaviours “an American stop” 😊 so Elon stop being a grandpa😂
Of course, speed limit signs still need a lot of work. Short 25MPH (the car thinks) in the middle of a 45 speed zone. Happens very frequently here in Mississippi. Another thing that causes me to take over is the creeping forward on narrow roads which freaks out cross traffic, so I often have to brake myself. I also get inconsistency in behavior (like your blind lane to the right) in various scenarios. Errant & tardy turn signals, although not serious are annoying to other drivers as well as the driver. Overall I'm very pleased with the progress of FSD Beta.
I also have FSD think speed limit is 25 in a 40 zone. Seems a lot more frequent with recent software. Yet another thing to fix "by the end of the year" to get the car to drive the entire trip.
please just stop talking so much.
I live for the graph at the end. Data!
Chris, you can't claim FSDb plus human > human. That's false.
Perhaps it is now, becasue the system fails too often, but due to automation complacency, the system + human is not safer than system alone if it gets better (say 100-1000 miles per DE).. Google "Disengagements: Wrong Metric for AV Testing"
Just because FSD is capable of no supervision doesn't mean it will be approved to do so.
maybe by the time HW5 or HW6 is released FSD might be FSD, lol. All the millions of teslas on the road today will never be FSD.
Very Nice Bro! Get Good People!
also, report wrong turn signals.
Love what your videos are turning into! It’s less a FSD video now and more an opinion piece on FSD or Tesla in general. Keep it up please😀
I agree. What is great about it is that the car does so well, there wouldn't be much to watch without listening to Chris' thoughts. I just wish FSD would work as well for me, and to handle more of the unusual cases properly.
I just made an interesting observation: the FSD visualisation is not based on just the current environment, modelled by the car "looking around". Just check the environment and its visulatization at 0:59. The upper left corner of the screen is showing a road which is totally blocked by the building on the corner. I assume some "network" effect here, probably another cars have already explored that part of the road and this Tesla somehow benefits from that shared knowledge.
The visualization is based on the car looking around and map data; the car knew there is a street behind the building cuz of the map data. Teslas don’t talk to each other yet, from our knowledge but would be a cool improvement
I thought map data would not be used because the goal is to be operational in unexplored terrain.
@@nick0780they work in unexplored territory, but even a human needs some context from a map to drive smoothly and confidently in a new area. I can only guess that at some point the cars will “remember” their frequent routes, as you or I would, but they will also use the fleet of cars to continually get accurate coarse map data for pretty much everywhere. Would be pretty interesting if Tesla some day sold a public map that was kept continually up to date using the cars
@@nick0780 map data is used when available , but also FSD works in unmapped areas as seen in remote locations where map data is limited
Map data even overrides the vision system. I have two intersections where the map data says you can't go. The visualization clearly shows a path but the car still won't take it.
One huge problem I see with driverless cars (without a steering wheel!) are certain edge cases. For example, I think everyone knows the videos where a gang just stops a car and tries to rob them or whatever. In most of these videos, the drivers just set back a bit, go full throttle through that situation. Just imagine your sitting in your Tesla without a steering wheel, it of course stops for these people, and you're ABSOLUTELY trapped there - no way to escape.
move away from these people
Cars should always have steering wheels as a “just in case device”
Hey Chris! I just drove my Model Y from California to Michigan last week. I will say the car did struggle more in Michigan on FSD compared to California. But Michigan has AWFUL roads. But my dad might see you at the chargers at Meijers in Ann Arbor keep an eye out for a matte white model y with California plates lol
Currently, is it possible for FSD Beta to reverse? I saw one video that showed it reversing but none others. Has the functionality been implemented?
I don’t believe it can truly reverse. I think there have been cases of the car stopping and sort of rolling back with momentum, but nothing with it actually reversing using the motors
@@SyntheticSpy ua-cam.com/video/c1SpE18GO4o/v-deo.html this video seems to show it reversing at an intersection. The driver says that the road was slightly downhill so it couldn't be just rolling back.
Maybe 2024. Maybe. 😢
Local regulation aside, do you think the system is going to be able to handle road systems in other countries? e.g. in the UK we don't generally have grid based cities that were built with vehicles in mind which leads to a lot of twisty, narrow streets and mixed lane traffic with a preference for roundabouts instead of intersections.
Always feels like that when Tesla is talking about FSD progress it's really just the US market. I guess the exception would be highway driving which is pretty similar everywhere and works very well.
Loving the content and seeing how the system is improving.
I was a person who sent some to you, but you didn't really go the way I suggested (no big deal). I suggested going up 7th from stadium and then turning left on Huron. In your drive, had you wanted to turn right from liberty onto south seventh rather than turning left the bike lane that shares the right causes the issue. You mentioned it but the route went left so it was not an issue for you. And then I believe you turned right on Huron rather than left, which is no big deal for the car. Also you turned left from beakes onto 4th I think instead of main and so forth, so of the 8 or so spots that I have had issues with I think you only tried 3 (the double red light and the flashing yellows and the passenger crossing bollards), but it was still nice to see.
You could try finding one of those crosswalk posts for which is stops or really slows down and then cover the mini-stop with white tape and try again to see if it makes any diffenrece. Then, you should remove the tape. 😉
tips on how to stay out of FSD car? 😂😂 upvoted!!!
I still see it as difficult as in South Carolina you have to yield on a huge amount of RIGHT turns with random inconsistent signage that generally does warn you in time
Thanks for looking for and testing more edge type situations these are the challenges facing FSd and until most are fixed; it can't be closer to level 5; if there isnt' attention to them; why woud they get fixed? So this is better than running on the same roads. Thanks
8 Mile
As someone buying tesla stock every single week i hope they take another few years for FSD to be complete lol. Stock price is going to be insane when ever you can sleep in your car and it takes you to work
The yellow light at 12:32 had a car in the opposing left turn lane waiting to turn. I would have stopped at the yellow not only because it had been yellow for awhile but also the car waiting to turn was committed and out in the cross traffic lane so ego should have stopped to let that car get out of that situation.
Chris, sure appreciated your commentary on the status of FSD Beta. Thanks.
You did not notice that the Tesla creeped into the bicycle lane at the light, which had there been a trike or group of bikers behind you, they would have been pushed into the curb. Also, real world situations include inclement weather, construction lanes which will not be handled by Elon's prediction of Level 4. I'm guess his interpretation is Level 4 when things are ideal! Ha! The random wrong turn signal is a accident waiting to happen, I get it 10% of my drives. Getting better, but these unique situations will always make me handle FSD like a student driver, especially with my wife in the passenger seat gripping the side door handle and gritting her teeth.
I agree for the bike lane, and I have the same experience. At red light beta tends to creep into the bike lane, making it unsafe for bikes coming behind. I can't say how it does if a bike was already visible, but going into the bike lane and stopping there is not something I like (and a potential for road rage incident).
what does the button on the dashboard do ?
FSD seems to have completely fixed lane changes
Locally, i've had serious issues with .4.4 which weren't there or were "corrected" in .3. Right now, the car always tries to moves/pass to the right. The bus stop deviation is back, it is back into bike lanes and right-only lanes which are obviously and clearly identified. Highway lane change is less smooth than .3 as well. such a weird release.
I agree with you it is an obsession with going right. on the interstate, It wants to take almost every exit it can find. It really seems like it wants to go right, unless of course it's sitting in the passing lane and it refuses to get out of it, unless they are on my bumper. In town It's wonderful though.
Such a stunning city ❤
I'm curious: When the car does a roll-back, is it because it's on an incline, or is it because the CPU realized a little too late that it was stopping too far forward, so it reversed voltage to the motor for a slight (safe) roll-back?
I've noticed that odd rollback a few times, too. Mine usually occur after moderate to heavy stops, no incline, and well before a stop line. It probably only rolls back 2-3 inches at most, but it feels like 2-3 FEET. :)
8:35 the average driver has no idea what to do when there's an emergency vehicle. Some drivers stick to their route and ignore it. Some change lanes in front of them. Some put themselves right behind them to take advantage of the higher speed through dense traffic. And if parked on the side, some crash into them and some drive just very close by.
So I think they may just trying to figure out what's the best approach here.
What I would really like to see is for Tesla to merge the AP and FSD forks so that AP uses the FSD code. They can keep AP features exactly as they are, only do lane keeping, not honor lights or stop signs, not make turns, etc. but just use the FSD code to do the lane keeping for AP. This would correct a lot of improper lane maneuvers, phantom breaking, etc. related to the old AP code.
they already did this i think
@@lowmax4431they did it for people who have FSD Beta, but not for the standard autopilot that comes with every car. I think they are probably waiting until highway FSD is pretty much flawless before replacing autopilot for everyone, maybe even allowing for level 3 or higher on freeways when they do so
I still don’t understand why so many of us are still on FSD 11.3.
I’ve been a FSD beta user since ver 9 and honestly I’m disappointed by the state of FSD. It works great in most situations but then it does something that makes absolutely no sense.
Apparently there is some background data related to safety, and 11.3.6 has a higher safety profile vs the 11.4 branch... so as soon as safety is above the acceptable threshold, mass updates will resume. Take it with a grain of salt.
I just want FSD to be there for me in an emergency like I'm impaired (medically or otherwise) and I need it to drive me home or to the nearest emergency room. I will always be paying attention as much as possible and hope it anticipates better and isn't distracted like the average driver. I would love for FSD to allow us to practice and pre-plan such routes like "take me to the grocery store" or "take me to my doctor's office", etc - maybe 10 pre-planned routes including it parking itself by looking in a pre-planned parking lot. That said I just don't see how Tesla will ever be able to avoid a huge pothole because right now I don't see it seeing those and even I sometimes have trouble distinguishing between a shadow, or a recently fixed pothole and a real pothole. Have you shown any pothole navigation in Michigan where they organically grow their own pothole monsters?
It seems like Tesla is prioritizing general safety and decision making before addressing smoothness things like speed bumps or potholes, but of course none of us have inside info. One thing that we have to remember though is that what is shown on the visualization is a separate thing to what the car sees. The car may be detecting potholes, maybe in a sort of practice mode, and just not acting on them as far as we know
Seeing and hearing the condition of those roads makes me cringe. This video made me thankful for the quality of most of the roads I drive on in Texas.
Compare FSD to average driver, drivers who are usually impaired in some way, vision impaired, distracted drivers looking at smart phones, drug / alcohol impaired drivers, old drivers who get confused or range of motion impaired, young drivers high and speeding. so consider that FSD is a lot better than the "average " driver, with so many potential impairments. Thanks Great Video !
15:09 It's not stop signs, it even slows down too much with our Montreal sticks: ua-cam.com/video/S2kOiKf1nIQ/v-deo.html
Tesla theory: Tesla will always strive for a high turnover of used vehicles because ownership change increases chances for fsd purchases.
If i would have to follow a fsd Tesla it would makes me mad
is Tesla using driving scenes from normal drivers to train its software? Since that would be a huge database to learn from (situations and driver reactions).
They do, but in a very smart way.
They query the fleet to collect the requested situations in shadow mode. So this data is heavily filtered, but very high quality.
But even with all the fleet doing heavy filtering, they have way too much data. They really do need Dojo (or equivalent) for the training.