It's not that FSD doesn't do a good job. But he just doesn't do it. So this is probably expected. In many situations, especially with a fallible system (because it still relies on programming), it is better not to intervene, and even less to turn the wheel. (what a human often does, when faced with a dog, only to get hit by a car in the next lane) Everything is detected (not always displayed), but not everything is planned yet.
@@RK-wz4cc On the highway ? Statistics are the priority for Tesla, not impact avoidance ! This system is not capable as well as a human of distinguishing between a stuffed toy, a plastic bag and a real animal. Even man has difficulty with this, sometimes even man takes the risk of rolling over a bag (which might not be empty). It's a compromise for the car's drivability at the moment, but it can't be turned the other way unless the FSD stops the car at the slightest doubt. Same for holes in the road, slowing down the car can be frustrating or even be reported by the driver as an error when the FSD thought it recognized a hole, when it is, I don't know, a puddle. So, it's way more difficult than it looks, from a human perspective. And for Tesla, the main thing is that it remains user-friendly for the moment that the FSD is not 99.99% sure of what it can doo. It is also the desire of Tesla fanboys to want a system which takes more risks and which does not slow down too much at the slightest doubt. I would call it the "fanboys paradox". The release of FSD v12 will change these things a lot because it will drive more carefully without making statistics between on fanboy action and an other. More naturally human like experience, than the resemblance of an empty plastic bag and a plastic bag which looks like it contains something (to put it simply). Something you can't program unless you have to know every possible case in the world.
@@theplouf5533 AutoPilot is only for highways but FSD has more use in cities. It even didn’t have highway support until recently. It should be able to stop in front of an object if there is no cars around it and if it’s unsure what the object/obstacle is.
5:04 - This is a rather difficult situation. On one hand, you don't want to run over/kill wild animals if you can avoid it, and you also don't want to run over sharp objects like pieces of metal. On the other hand, how do you differentiate between something harmless like leaves or a crumpled up piece of paper? It's difficult even for humans to identify these things from a distance, and you may only have a few seconds to decide what to do!
Totally agree. Unless there's a better way to feed data to FSD in real time other than its cameras... Either build a smart street that would supplement FSD with real time data or FSD needs to be AGI... Pick your poison.
That would be the obvious litmus test. What would a human do. If it would be difficult for a human to differentiate those objects, then I would accept the car to do “just as bad” but not WORSE
the fact that the road might be to narrow to safely go around might effect the judgement.Going in the ditch to avoid a squirrel or even a dog might not be good driving, a gravel road with uncertain edges might be different than a paved road with curbs.
Honestly a bit disappointing for how well I would have expected it to perform, but hopefully future versions will vastly improve. At the moment squirrels are not in safe company 🐿️
I can see the headlines: "Robotaxi hits kid". Most of us would not drive around the kid, we would stop and wait for the kid to go clear across and be on the sidewalk on the other side. We have a lot of deer here especially in the fall. They browse along the edge of the road, look up when I drive up, stare at me like ("Hey, dummy, I'm eating here!") and go back to munching. I drive by at something like 1MPH, no FSDb.
For a rolling ball, it really should prioritize stopping, or going in front of the ball’s path, not behind. As humans, we know (or should) that a rolling ball is often followed by a running child.
Great stuff. Adding in footage from the dash/sidecam feeds for those close call scenarios will def make it clearer as to how close/far the car was. Next would be to replicate it in dusk low light conditions
Either it works better, because there are less brightness changes (branches throwing shadows) or it performs worse because of noise. I guess the latter because noise would intervene even more with the pixels of small objects
Service menu lets you bring up all the cameras. If you can use it while driving that would be good for these tests to see what the car sees. And / or GoPro on the side of the car so we can see how close it's getting.
for the case "a young child in the middle of the road, not moving" there would be only one correct programming "hey @Elon": stop the car in a protecting manner in front of the child and a computer voice would ask the driver or passengers "do you want to assist the child or continue?" because that's what humans would take into account
Camera vs lidar: We saw how "not so good" lidar was at detecting a person with a bicycle, crossing a multi lane street at night... I think, the spokes of the bicycle wheels makes it hard for lidar to detect. And so, some objects in a color close to the road color are hard to detect by cameras.
This was really helpful. When using FSD on a freeway, it's very easy to get lulled into a false feeling that the car is smarter than it is. Definitely need to keep attentive. A good reminder.
I kinda like this more than the normal drives. Drives through Ann Arbor are fairly routine now since they haven’t added many new features. These safety critical tests are important for people to understand how the safety systems work, for better or for worse. What can we expect our vehicles to assist with?
My 20 with FSD Beta stopped for a 2 year old running in the street a few months ago. I would have stopped myself in time but it was nice to see that FSD did before I tried.
I have a 2023 Model Y with hardware 3. One day I saw a paper bag in the middle of my lane. I thought I'd just run over it, but while in FSD, the car surprisingly moved into the left lane and went around it.
In Quebec you legally can’t stop for an animal. A few years ago a woman stopped for ducks and 2 motorcyclists died by hitting her car and she went to jail
The front camera and rear camera are great for the driver to see like if a child walked into the street or if you're pulling up to a parking space, you can see how close you are to the object
I dunno if it is universal, but the driver's guide I studied back in the day actually told you NOT to brake for squirrels, rabbits, or other small animals.
Those are super concerning results in my opinion. A total lack of concern for static objects and minimal caution around the child figure are both critical failures.
I wouldn’t want my car to hit any small object like the box. If the box had something heavy in it or a small boulder was in the road, serious battery damage/failure is a possibility. Regarding the mannequin, I wonder if the car detects motion of the legs of a real person so a mannequin slowly dragged across the road doesn’t register the way a kid running out in the road would. Somebody posted a video where the car emergency braked for a kid running across.
its useful to summarise that these tests equal more pedestrian issues a driver would face in 10-20 years. Also other drivers in trying to avoid these cases could have an accident with another car etc in trying to avoid the impact - where Tesla did not. The 2nd key is most drivers would miss one or some of these encounters and be distracted by a conversation or the phone.
RSymons has some very, very nice test with Tesla vision failing miserably. In a parking lot, it hits a big red bin, a human, a huge logg. Some of these problems are also visible in your test.
It would be really interesting to see how other car brands react to a child near the edge of the road. In Europe these city stop radars are mandatory on new cars, and in euroncap tests they stop.
Yeah, my cupra did an emergency break a few weeks ago while reversing on the street to park next to the sidewalk, when my nephew jumped out of our garden onto said sidewalk. With a distance of around 4-5 meters. Was a little shocking in the first moment but definitely appreciated for real emergencies
Good Video! Well done! I've had FSD emergency brake for wild turkeys. I would like to see chunks of truck tires, a common piece of debris that should have good contrast. A few days ago, I disengaged for a 10 ft step ladder lying on the highway, that obviously fell off a craftsman's truck. Again it had high contrast but I didn't feel any braking or get any alarm.
This just goes to show that, while things are greatly improved from the previous iterations, that the system still needs to learn more avoidance behaviors. Every improvement is a positive net gain, but it def shows that real world situations are hard to quantify for a computer program. Maybe a few more years, and it will work perfectly, or as perfectly as we can ever manage anyway.
Hey Chris. Great work doing these videos. What about putting a high visibility vest on the mannequin and using a red painted box. I have a feeling that the background being busy with foliage isn’t helping FSD. I think obstacle avoidance would be better if it could see the horizon better than a wall of trees. Maybe do these tests in a more open area to see if it makes a difference.
Wouldn’t Tesla’s own testing show the same things? Why wouldn’t they be adapting their camera placement? Everybody who thinks Tesla is just going to flip a switch and the whole fleet is ready for Robotaxi is not being rational. The camera placement has so many blind spots all around that are obvious to anyone paying attention
I would love to see footage from top-down angle, or from above and head-on to the car next time! You could get this shot from a stationary drone in the air recording. Also maybe comparing the results to previous tests of older versions of FSD
I just got the latest update night before last and was excited to see if my major issue with unnecessary lane changes was corrected. I had a bit of strangeness before turning FSD on where it was like the lane management logic wasn't working properly such that the car wanted to pull to the left and 'twitched' one way or the other. Then an error popped up saying the Lane Management had been disabled due to an issue but might work again after the car was shut down. Good news - it did. Now on to the bad news. Once I got to the place where I usually turn FSD on I did so, and the instant I turned it on it put the left turn signal and wanted to move to the left lane for no good reason as the closest car in front of me was over a quarter mile away. Waiting for version 12.........
It would be great to have a Tesla driving simulator to allow experiencing unexpected events, e.g. vehicles swerving into our lane, boxes of kittens appearing in the road, etc. Would be nice to know what might happen and to be more ready to react while car is in self-driving mode.
“I would just hit it” sounds harsh but i had people like 4 feet behind me once going 45 and i took out 2 pigeons and the front dash cam vid showed me a lot😭 i felt so bad.
I’m writing this at the 4:40 mark in the video, so you may have corrected this later, but can you for future videos like this place your forward looking camera so that it fully shows your foot pedals and complete steering wheel? I don’t know if that’s possible, but if it is, it’d sure go a long way towards shutting down any doubters from commenting that you moved the steering wheel or pressed on the accelerator or brake pedals at any time during the testing. By the way, the clarity is fantastic! Definitely infinitely better than that Dan O’Dowd guy.
@@DirtyTesla , yeah, While I know that the system would turn off, I’m referring to more of what I’m hoping, at least part of your target audience would be, those that don’t know that, or even worse, those Dan O’Dowd types that will use anything they can to discredit your great work. There are just too many out there like Dan who could, and would very easily use the fact that not seeing the whole steering wheel or the pedals to convince the less knowledgeable of whatever lies they want to make about your testing. We know this from the Dan O’Dowd tests themselves. Love your channel and keep doing what you’re doing!
I think the dirt road is a large part of the problem here. Because the car is more willing to go around an object and use the whole road. I think in an actual lane on a road it may do better at completely stopping
Wow, finally we get a counter to Dan O'Dowd's vid showing "bad Tesla" running of "innocent" children. You know, sometimes humans have a tough call when something unexpected appears in their path. FSD is in a no-win situation because nobody wins in this case. Should you slam on the brakes to avoid the deer or swerve. Nobody knows, it depends. In milliseconds, all kinds of bad things can happen. Microseconds. How to program that? Nice vid, well done
I love these kind of experiments. I'm curious if the chunk of wood or the large ball would have been seen better if it was not in the shade. Is it possible that the camera system doesn't react to exposure changes fast enough?
When doing data sciences, we are using a "baseline" to compare with AI. For example you can ask some people to drive the car, not telling what will happen, and see how they avoid a dog crossing at the last time. Of course, you need a lot of people. Or you can just look at statistics, how many humains hit animals, at what speed, etc.
Great video. Just an observation/guess: Our human central vision is better than the 4K camera, i.e. you could recognize the squirrel before the system (at least we believe that, unless the system is opting not to overwhelm us with too much information on the situation graphic). Higher pixel count cameras will help solve this problem even possibly without having to increase processing power of the vision system too much. That is such a system can track multiple distant obstacles without doing what we do which is shift our gaze. Pretty cool to think system could avoid what might distract us, e.g. squirrel running out in road, attracting our attention, missing fact that a child is/was chasing it. In this case system will opt to kill the squirrel or at least put it between the tires based on trajectory at known location.
Wow, FSD is still years away. Also, not using a camera, radar, or ultrasonic sensors that can see immediately in front of the bumper was a horrible idea. Unless Tesla knows how to bend light, they have a critical blind spot.
From my experience, the car is good at knowing if people are about to cross the road. It has been trained on videos of people going into the road and knows what people look like when they’re about to step into the road. The mannequin is standing so even if it’s moving, it doesn’t resemble the real world examples that the car has been trained on.
I wonder how much the gravel road could have scewed the tests vs a clearly marked paved road. Room for improvement for sure. No doubt it's going to be achieved in future versions.
During the intro, the boss wife is sitting in the car staring into the distance, lost in thoughts not paying attention to her husband blabbering away for a couple minutes into a camera 😜
Nice to see the progress of FSD Beta as the price creeps up. There's no way I'm going to let FSD drive me without supervising it. It will still drive into huge and small potholes, drive over / into road debris. Maybe with the new AI-based version, it will do far better in these circumstances. Thanks for providing these updates.
7:32 As a human you often don't see things but you manage because you remember where they are. It not, a human would not be able to pass that dog since the hood is not transparent.
You should show your rear cams and Repeaters to show how much room it actually has to the road edge when it misses an object by inches. My guess is feet of room which it has little confidence in detecting
Overall, I don’t want the car to swerve and do crazy things. Have you seen all the videos where people overcorrect and launch themselves? Cars are meant to protect you in the event of a crash. I’d MUCH rather hit something than go cranking the wheel to avoid things at all costs, only to do 5X the damage.
Can you put some white paper or something down to pretend there is a crosswalk with the dummy to the side and see if the car will stop to let the dummy cross?
There is a edge case that I am not cretin that may not have been addressed. Think of driving on a bridge at speed and coming up on section that has fallen in and is missing or a possible wash out. Will the car recognized the missing span or road section missing and stop. There is another suggestion I have a 2023 Y long range and I am 73 and my vision is not what it used to be and my neck is not as flexible and it is difficult to turn my head far enough to comfortably to double check the blind spot . My suggestion is when the turn signal is on I would love to get the side camera image full screen there is no reason to have the map or speedometer when changing lanes. I bought this car for the safety features as I see this may be the last car I will buy. I have to take my eyes off the road too long to scan the small blind spot that is only 2.5 x 5.65" image on the 15" screen to be of any use. I do not think the government will ever allow cameras instead of mirrors until this is addressed I would also think that if your signal is on and a car is in your blind spot that the Tesla should sound an alarm that could be adjusted for volume. I am a firm be-leaver that as we age the more technology to help in our activities the better. I know that Elon may view this and I hope that something can be done for those of us that are older and those with physical challenges can be addressed.
Thanks for another informative video. When using FSD in my 2022 MX and navigating in my neighborhood, there is a large, ornate, metal mailbox painted white that shows briefly as a person walking on the screen parallel to the direction I am driving. This "ghost" appears every time. Just a quirk of the system.
Thanks for doing this! At 19:13 the screen was visualizing Stephanie standing next to the car talking to you, but it also showed another person standing on the shoulder of the road... was this the child manikin?!!
In a few years when FSD EV's are mainstream it is inevitable that law enforcement will have access to these FSD networks. This means two things: 1.) Law Enforcement will be able to pull you over and/or disable your vehicle whenever they have need and 2.) Hacking these networks will be a thing. As the vehicles evolve and have no steering or way for humans to control them you will be at the mercy of these individuals.
Think about the training data. How many examples does it have of things like this happening in collected data? Most people would likely come to a full stop if they drove up on a small kid in the road who wasn't with an adult. FSD doesn't seem to have this concept covered yet. "Well, I could keep going, but my goal needs to change to protecting this human, rather than getting to my destination." Maybe Optimus will be able to teach it a few tricks. 🙂
Is hardware 3 similar with the latest updates? I think the blind spots are a problem for pedestrian detection and the reason for there being no top down 360 view like other cars have. All this needs to work regardless of mode!
I think for next time, we would like to see the kid get hit... Great experimenting and it's definitely going to be interesting for the Tesla team watching your crazy antics!
I find this quite shocking! This is definitely not something that should be used on public roads! Driving at 10-12 miles per hour close to a child who obviously has problems crossing the road? Not acceptable!
One extra safety concern with the test. There is some risk of the car catching the rope if it drives over it which could be very bad for someone whose body, hands or fingers get tangled in the rope. Maybe a small risk but unlikely things happen with enough tests. Not sure how to make it less dangerous, maybe a light weight string that would break?
Sponsored by DeleteMe. Take back control of your personal information: visit JoinDeleteMe.com/dirtytesla and use code DIRTYTESLA for 20% off!
Do a test as though a human got hit by another car and is now in the path of the Tesla. To simulate the cruise incident.
@@cybertruck4988 Grim
Such is my experience. FSD doesn't do a good job avoiding smaller obstacles that are still big enough to cause significant damage
It's not that FSD doesn't do a good job. But he just doesn't do it. So this is probably expected. In many situations, especially with a fallible system (because it still relies on programming), it is better not to intervene, and even less to turn the wheel. (what a human often does, when faced with a dog, only to get hit by a car in the next lane)
Everything is detected (not always displayed), but not everything is planned yet.
@@theplouf5533 the car could have just braked and stopped IF it had detected it.
@@RK-wz4cc On the highway ? Statistics are the priority for Tesla, not impact avoidance ! This system is not capable as well as a human of distinguishing between a stuffed toy, a plastic bag and a real animal. Even man has difficulty with this, sometimes even man takes the risk of rolling over a bag (which might not be empty).
It's a compromise for the car's drivability at the moment, but it can't be turned the other way unless the FSD stops the car at the slightest doubt.
Same for holes in the road, slowing down the car can be frustrating or even be reported by the driver as an error when the FSD thought it recognized a hole, when it is, I don't know, a puddle.
So, it's way more difficult than it looks, from a human perspective. And for Tesla, the main thing is that it remains user-friendly for the moment that the FSD is not 99.99% sure of what it can doo.
It is also the desire of Tesla fanboys to want a system which takes more risks and which does not slow down too much at the slightest doubt. I would call it the "fanboys paradox".
The release of FSD v12 will change these things a lot because it will drive more carefully without making statistics between on fanboy action and an other. More naturally human like experience, than the resemblance of an empty plastic bag and a plastic bag which looks like it contains something (to put it simply). Something you can't program unless you have to know every possible case in the world.
@@theplouf5533 AutoPilot is only for highways but FSD has more use in cities. It even didn’t have highway support until recently. It should be able to stop in front of an object if there is no cars around it and if it’s unsure what the object/obstacle is.
Wow you found every shitty excuse to praise the Teslas "FSD"
Great honest test, been around since you had less than 3k, and you've never failed to deliver good content. Have a good day man 👍👍👍
Thank you! Hoping to see it behave even safer around this child mannequin next time.
Great content to have your wife test if the car will hit her 😂
I love videos like this because it helps to reduce FUD (Fear, Uncertainty, and Doubt). Thanks for the testing!
5:04 - This is a rather difficult situation. On one hand, you don't want to run over/kill wild animals if you can avoid it, and you also don't want to run over sharp objects like pieces of metal. On the other hand, how do you differentiate between something harmless like leaves or a crumpled up piece of paper? It's difficult even for humans to identify these things from a distance, and you may only have a few seconds to decide what to do!
Totally agree. Unless there's a better way to feed data to FSD in real time other than its cameras... Either build a smart street that would supplement FSD with real time data or FSD needs to be AGI... Pick your poison.
That would be the obvious litmus test. What would a human do. If it would be difficult for a human to differentiate those objects, then I would accept the car to do “just as bad” but not WORSE
Exactly it would cause a lot of phantom braking
My genius consultant says, one should drive at the speed which is appropriate for the conditions at present and at this location.
It's very reassuring, it brakes when the kid is standing but not if the kid fell over
the fact that the road might be to narrow to safely go around might effect the judgement.Going in the ditch to avoid a squirrel or even a dog might not be good driving, a gravel road with uncertain edges might be different than a paved road with curbs.
Stroller : 3:29
Plushie : 4:59
Dog : 6:23
Box : 8:51
Cedar post : 9:51
Ball : 11:17
Electric scooter : 13:12
Child : 15:18
Honestly a bit disappointing for how well I would have expected it to perform, but hopefully future versions will vastly improve. At the moment squirrels are not in safe company 🐿️
4:18 Stéphanie is ready for the Oscars 😂
I can see the headlines: "Robotaxi hits kid". Most of us would not drive around the kid, we would stop and wait for the kid to go clear across and be on the sidewalk on the other side. We have a lot of deer here especially in the fall. They browse along the edge of the road, look up when I drive up, stare at me like ("Hey, dummy, I'm eating here!") and go back to munching. I drive by at something like 1MPH, no FSDb.
For a rolling ball, it really should prioritize stopping, or going in front of the ball’s path, not behind. As humans, we know (or should) that a rolling ball is often followed by a running child.
Yeah, if uncertain it should always just start the hazard lights and stop immediately, especially if there are no cars directly behind it
Just blows my mind that Tesla is going all in with cameras but yet doesn't have a front bumper camera.
Forget about cameras they don't even have an actuator to close the frunk
Great stuff. Adding in footage from the dash/sidecam feeds for those close call scenarios will def make it clearer as to how close/far the car was. Next would be to replicate it in dusk low light conditions
Either it works better, because there are less brightness changes (branches throwing shadows) or it performs worse because of noise. I guess the latter because noise would intervene even more with the pixels of small objects
That squirrel test just ruined every vegan 😂😂😂
Service menu lets you bring up all the cameras. If you can use it while driving that would be good for these tests to see what the car sees. And / or GoPro on the side of the car so we can see how close it's getting.
ok, it can dodge balls, but what about wrenches!?
for the case "a young child in the middle of the road, not moving" there would be only one correct programming "hey @Elon":
stop the car in a protecting manner in front of the child and a computer voice would ask the driver or passengers "do you want to assist the child or continue?" because that's what humans would take into account
Wow, your wife is a KEEPER !! My wife would never let me do any of this with her ! Nice to see her more involved in the videos.
Camera vs lidar: We saw how "not so good" lidar was at detecting a person with a bicycle, crossing a multi lane street at night...
I think, the spokes of the bicycle wheels makes it hard for lidar to detect.
And so, some objects in a color close to the road color are hard to detect by cameras.
Good job! Please keep these tests going. You can be the Chuck Cook of UPLs for "Small Objects in the Road" (including speed bumps).
Tesla's marketing is next level. Users buy their software for over 7K-15K and volunteers free to do all tests and Tesla collects all data.
It's great that you live in an area where you can easily do this. Most especially with very minimal to no problems.
Great video, Y😎!
This is one of the best videos that I've seen about this issue. Thanks Regards.
This was really helpful. When using FSD on a freeway, it's very easy to get lulled into a false feeling that the car is smarter than it is. Definitely need to keep attentive. A good reminder.
I kinda like this more than the normal drives. Drives through Ann Arbor are fairly routine now since they haven’t added many new features. These safety critical tests are important for people to understand how the safety systems work, for better or for worse. What can we expect our vehicles to assist with?
My 20 with FSD Beta stopped for a 2 year old running in the street a few months ago. I would have stopped myself in time but it was nice to see that FSD did before I tried.
I have a 2023 Model Y with hardware 3. One day I saw a paper bag in the middle of my lane. I thought I'd just run over it, but while in FSD, the car surprisingly moved into the left lane and went around it.
Great video! 👏👏 thank you very much to both of you for your work 😉😉
rip doggy 2021-2023* ❤
I actually got it in 2021 and it's survived until now lol th-cam.com/video/-Oro7GUT3Wk/w-d-xo.html
In Quebec you legally can’t stop for an animal. A few years ago a woman stopped for ducks and 2 motorcyclists died by hitting her car and she went to jail
I just wish I could get across town with FSD without it doing something wrong/stupid!
The front camera and rear camera are great for the driver to see like if a child walked into the street or if you're pulling up to a parking space, you can see how close you are to the object
I dunno if it is universal, but the driver's guide I studied back in the day actually told you NOT to brake for squirrels, rabbits, or other small animals.
Those are super concerning results in my opinion. A total lack of concern for static objects and minimal caution around the child figure are both critical failures.
I wouldn’t want my car to hit any small object like the box. If the box had something heavy in it or a small boulder was in the road, serious battery damage/failure is a possibility. Regarding the mannequin, I wonder if the car detects motion of the legs of a real person so a mannequin slowly dragged across the road doesn’t register the way a kid running out in the road would. Somebody posted a video where the car emergency braked for a kid running across.
I continue using FSD beta although it tries to murder me every time I do, with the hopes that I’m helping improve it so someday it doesn’t…
its useful to summarise that these tests equal more pedestrian issues a driver would face in 10-20 years. Also other drivers in trying to avoid these cases could have an accident with another car etc in trying to avoid the impact - where Tesla did not. The 2nd key is most drivers would miss one or some of these encounters and be distracted by a conversation or the phone.
RSymons has some very, very nice test with Tesla vision failing miserably. In a parking lot, it hits a big red bin, a human, a huge logg. Some of these problems are also visible in your test.
Great test. Thank you for doing this.
That entire add for delete me was soooo smooth.
I already see the next headlines of WSJ: Tesla kills Scooter on Backroad!
It would be really interesting to see how other car brands react to a child near the edge of the road. In Europe these city stop radars are mandatory on new cars, and in euroncap tests they stop.
Yeah, my cupra did an emergency break a few weeks ago while reversing on the street to park next to the sidewalk, when my nephew jumped out of our garden onto said sidewalk. With a distance of around 4-5 meters. Was a little shocking in the first moment but definitely appreciated for real emergencies
For the past four years, I have been requesting to toss a Hot Wheels toy car onto the road and observe tesla reaction...:(
Good Video! Well done! I've had FSD emergency brake for wild turkeys. I would like to see chunks of truck tires, a common piece of debris that should have good contrast. A few days ago, I disengaged for a 10 ft step ladder lying on the highway, that obviously fell off a craftsman's truck. Again it had high contrast but I didn't feel any braking or get any alarm.
This just goes to show that, while things are greatly improved from the previous iterations, that the system still needs to learn more avoidance behaviors. Every improvement is a positive net gain, but it def shows that real world situations are hard to quantify for a computer program. Maybe a few more years, and it will work perfectly, or as perfectly as we can ever manage anyway.
awesome video man!! There is work to do with FSD, but it is already amazing.
Very cool tests. Tesla must hire you my friend for all these scenarios. Awesome Job.
"If you didn't know me you'd think I was insane"
"Yeah, you'd get a dirty look from me."
Stephanie is on brand. ☺
yeah robottaxi feels a bit far away i think
Hey Chris. Great work doing these videos. What about putting a high visibility vest on the mannequin and using a red painted box.
I have a feeling that the background being busy with foliage isn’t helping FSD. I think obstacle avoidance would be better if it could see the horizon better than a wall of trees.
Maybe do these tests in a more open area to see if it makes a difference.
And always remember to request all objects intending to be in your path to be properly colored and highlighted. 🤯
1:06 * If you have FSD on your car in North America. My screen says "Full Self Driving" in the options but we have almost none of this stuff. 😉
Wouldn’t Tesla’s own testing show the same things? Why wouldn’t they be adapting their camera placement? Everybody who thinks Tesla is just going to flip a switch and the whole fleet is ready for Robotaxi is not being rational. The camera placement has so many blind spots all around that are obvious to anyone paying attention
I would love to see footage from top-down angle, or from above and head-on to the car next time!
You could get this shot from a stationary drone in the air recording.
Also maybe comparing the results to previous tests of older versions of FSD
This is great, kudos for doing this!!
I just got the latest update night before last and was excited to see if my major issue with unnecessary lane changes was corrected. I had a bit of strangeness before turning FSD on where it was like the lane management logic wasn't working properly such that the car wanted to pull to the left and 'twitched' one way or the other. Then an error popped up saying the Lane Management had been disabled due to an issue but might work again after the car was shut down. Good news - it did. Now on to the bad news. Once I got to the place where I usually turn FSD on I did so, and the instant I turned it on it put the left turn signal and wanted to move to the left lane for no good reason as the closest car in front of me was over a quarter mile away. Waiting for version 12.........
0:46 Her questioning her life choices in the car 😆
She really trusts him
It would be great to have a Tesla driving simulator to allow experiencing unexpected events, e.g. vehicles swerving into our lane, boxes of kittens appearing in the road, etc. Would be nice to know what might happen and to be more ready to react while car is in self-driving mode.
“I would just hit it” sounds harsh but i had people like 4 feet behind me once going 45 and i took out 2 pigeons and the front dash cam vid showed me a lot😭 i felt so bad.
Your neighbor probably thought you had lost it. 😂
I’m writing this at the 4:40 mark in the video, so you may have corrected this later, but can you for future videos like this place your forward looking camera so that it fully shows your foot pedals and complete steering wheel? I don’t know if that’s possible, but if it is, it’d sure go a long way towards shutting down any doubters from commenting that you moved the steering wheel or pressed on the accelerator or brake pedals at any time during the testing. By the way, the clarity is fantastic! Definitely infinitely better than that Dan O’Dowd guy.
Sure, it's just a bit dark down there. Of course I can't steer or brake without the system shutting off.
@@DirtyTesla , yeah, While I know that the system would turn off, I’m referring to more of what I’m hoping, at least part of your target audience would be, those that don’t know that, or even worse, those Dan O’Dowd types that will use anything they can to discredit your great work. There are just too many out there like Dan who could, and would very easily use the fact that not seeing the whole steering wheel or the pedals to convince the less knowledgeable of whatever lies they want to make about your testing. We know this from the Dan O’Dowd tests themselves.
Love your channel and keep doing what you’re doing!
I think the dirt road is a large part of the problem here. Because the car is more willing to go around an object and use the whole road. I think in an actual lane on a road it may do better at completely stopping
12:50 "The ball did ok, kind of middle of the road" 😉
Thanks for another great review.
12:35 What if the ball rolling into the road was followed by kid running after it? I think it should have stopped and not gone toward the kid.
Wow, finally we get a counter to Dan O'Dowd's vid showing "bad Tesla" running of "innocent" children. You know, sometimes humans have a tough call when something unexpected appears in their path. FSD is in a no-win situation because nobody wins in this case. Should you slam on the brakes to avoid the deer or swerve. Nobody knows, it depends. In milliseconds, all kinds of bad things can happen. Microseconds. How to program that?
Nice vid, well done
I love these kind of experiments. I'm curious if the chunk of wood or the large ball would have been seen better if it was not in the shade. Is it possible that the camera system doesn't react to exposure changes fast enough?
Well done! Great info. Thanks
Thank you for always making such great videos. I'm sure this will be good too.
Addition: It was great
😅12:38 ‘Dodge’ ball…. Nice…😜. Is that a new game for Teslas?
Fun and educational. Well done!
When doing data sciences, we are using a "baseline" to compare with AI. For example you can ask some people to drive the car, not telling what will happen, and see how they avoid a dog crossing at the last time. Of course, you need a lot of people. Or you can just look at statistics, how many humains hit animals, at what speed, etc.
By "a dog", I do not mean a real one, of course !
Great video. Just an observation/guess: Our human central vision is better than the 4K camera, i.e. you could recognize the squirrel before the system (at least we believe that, unless the system is opting not to overwhelm us with too much information on the situation graphic). Higher pixel count cameras will help solve this problem even possibly without having to increase processing power of the vision system too much. That is such a system can track multiple distant obstacles without doing what we do which is shift our gaze. Pretty cool to think system could avoid what might distract us, e.g. squirrel running out in road, attracting our attention, missing fact that a child is/was chasing it. In this case system will opt to kill the squirrel or at least put it between the tires based on trajectory at known location.
It does a great job identifying ppl crossing the street. It sees them way before I ever do.
Wow, FSD is still years away. Also, not using a camera, radar, or ultrasonic sensors that can see immediately in front of the bumper was a horrible idea. Unless Tesla knows how to bend light, they have a critical blind spot.
From my experience, the car is good at knowing if people are about to cross the road. It has been trained on videos of people going into the road and knows what people look like when they’re about to step into the road. The mannequin is standing so even if it’s moving, it doesn’t resemble the real world examples that the car has been trained on.
I do agree with this
I wonder how much the gravel road could have scewed the tests vs a clearly marked paved road. Room for improvement for sure. No doubt it's going to be achieved in future versions.
During the intro, the boss wife is sitting in the car staring into the distance, lost in thoughts not paying attention to her husband blabbering away for a couple minutes into a camera 😜
Nice to see the progress of FSD Beta as the price creeps up. There's no way I'm going to let FSD drive me without supervising it. It will still drive into huge and small potholes, drive over / into road debris.
Maybe with the new AI-based version, it will do far better in these circumstances.
Thanks for providing these updates.
Would’ve been more interesting if you’d had a outside cam or the cars cameras shown too.
7:32 As a human you often don't see things but you manage because you remember where they are. It not, a human would not be able to pass that dog since the hood is not transparent.
Kudos to Stephanie for helping the Tesla FSD nerds community :D
I was on a 25mph road, a gaggle of Canada geese was in the road, and FSDB has 0 plans on stopping.
Mind did not see turkey vultures in the road. Was angry that I crossed the double yellow to avoid
You should show your rear cams and Repeaters to show how much room it actually has to the road edge when it misses an object by inches.
My guess is feet of room which it has little confidence in detecting
How dare you anticipate my thoughts of version 12 coming and how it may do in comparison! 😂
Overall, I don’t want the car to swerve and do crazy things. Have you seen all the videos where people overcorrect and launch themselves?
Cars are meant to protect you in the event of a crash. I’d MUCH rather hit something than go cranking the wheel to avoid things at all costs, only to do 5X the damage.
Can you put some white paper or something down to pretend there is a crosswalk with the dummy to the side and see if the car will stop to let the dummy cross?
There is a edge case that I am not cretin that may not have been addressed. Think of driving on a bridge at speed and coming up on section that has fallen in and is missing or a possible wash out. Will the car recognized the missing span or road section missing and stop. There is another suggestion I have a 2023 Y long range and I am 73 and my vision is not what it used to be and my neck is not as flexible and it is difficult to turn my head far enough to comfortably to double check the blind spot . My suggestion is when the turn signal is on I would love to get the side camera image full screen there is no reason to have the map or speedometer when changing lanes. I bought this car for the safety features as I see this may be the last car I will buy. I have to take my eyes off the road too long to scan the small blind spot that is only 2.5 x 5.65" image on the 15" screen to be of any use. I do not think the government will ever allow cameras instead of mirrors until this is addressed I would also think that if your signal is on and a car is in your blind spot that the Tesla should sound an alarm that could be adjusted for volume. I am a firm be-leaver that as we age the more technology to help in our activities the better. I know that Elon may view this and I hope that something can be done for those of us that are older and those with physical challenges can be addressed.
Thanks for another informative video. When using FSD in my 2022 MX and navigating in my neighborhood, there is a large, ornate, metal mailbox painted white that shows briefly as a person walking on the screen parallel to the direction I am driving. This "ghost" appears every time. Just a quirk of the system.
Thanks for doing this! At 19:13 the screen was visualizing Stephanie standing next to the car talking to you, but it also showed another person standing on the shoulder of the road... was this the child manikin?!!
Probably
Hahah wify earned a laugh for acting the crazy squirrel dog toy
In a few years when FSD EV's are mainstream it is inevitable that law enforcement will have access to these FSD networks. This means two things: 1.) Law Enforcement will be able to pull you over and/or disable your vehicle whenever they have need and 2.) Hacking these networks will be a thing. As the vehicles evolve and have no steering or way for humans to control them you will be at the mercy of these individuals.
Think about the training data. How many examples does it have of things like this happening in collected data? Most people would likely come to a full stop if they drove up on a small kid in the road who wasn't with an adult. FSD doesn't seem to have this concept covered yet. "Well, I could keep going, but my goal needs to change to protecting this human, rather than getting to my destination." Maybe Optimus will be able to teach it a few tricks. 🙂
Is hardware 3 similar with the latest updates? I think the blind spots are a problem for pedestrian detection and the reason for there being no top down 360 view like other cars have. All this needs to work regardless of mode!
This is why HW3 needs an upgrade with either a front bumper camera or radar. I believe HW3 still had the ability to do radar...
I am relieved to know that slow moving zombie children are safe around Teslas.
I think for next time, we would like to see the kid get hit... Great experimenting and it's definitely going to be interesting for the Tesla team watching your crazy antics!
I find this quite shocking! This is definitely not something that should be used on public roads!
Driving at 10-12 miles per hour close to a child who obviously has problems crossing the road? Not acceptable!
Good testing, IMHO not safe yet, but improving and a good level 2 drive assist.
One extra safety concern with the test. There is some risk of the car catching the rope if it drives over it which could be very bad for someone whose body, hands or fingers get tangled in the rope. Maybe a small risk but unlikely things happen with enough tests. Not sure how to make it less dangerous, maybe a light weight string that would break?
Yep, she was letting go of the rope before it was run over, but thanks
8:00 Since the dog fell over it probably lost track of it.