I remember one of your earlier videos where Carol followed you in her car and tried to guess if you were driving or if FSD was driving. Would love to see an updated version of that! But no cheating this time... lol
@@AIDRIVR So it works for slow moiving car and slow moving target in the open on a clear sunny day! LOL Lets see you test it with your kid running at night with no street lights and rain! Theres a reason ti will never work! SONAR is not suitable for this!
I'm impressed how brave you were with this video. That tire could seriously damage your car and you really waited to give it all the chances in the world. Edit: lol I commented this before the last test!
I would expect the boxes and tire lying flat get interpreted as pot holes or road variations instead of debris. As best I can tell, FSD is not yet modeling low lying features and imperfections in the road surface. That probably a very hard problem.
@@MusicalMemeology No, a radar world not help. From a radar's perspective, the unmoving objects are the same as the road. That's why the early versions of Autoilot didn't prevent collisions with stopped vehicles on a freeway or trailers moving horizontally across a road. Vehicular millimeter wave radars only track movement in the direction of travel. Anything not moving is the same as the road. My Corolla has a similar radar. In testing it's lane keeping assist, it doesn't see things like concrete barriers, which are stationary. All current vehicle radars have this limitation. Vision is a de facto better solution.
@@MusicalMemeology As Ed said, radar has serious issues with stationary objects or objects moving horizontally in front of the vehicle. If you've ever driven a Tesla or other vehicle using radar, you'll notice how very late it brakes for stopped vehicles and how it doesn't care in the slightest about road debris. By comparison, Tesla's vision approach is vastly superior at this task.
@@MusicalMemeology Also, keep in mind, we're seeing the results of Tesla building and testing this software. It has lots of room for improvement and many are expected to come before it's ready for general release. Road debris probably needs more attention from the auto labeling system during training. But the resources to handle that are still being built, namely Dojo. I would expect reactions to potholes and random debris to improve after Dojo has been online for a while.
@@Tanstaaflitis Not at all accurate. LiDAR can see objects, moving or not, very well. Software has to be intelligent enough to identify them, but the information is there and more easily interpreted than visualization.
This is awesome. As a tesla driver I often find myself wondering what kinds of objects the car will/will not notice and plan around, and this gives me a much better sense: some things, better with people, but not super reliable in any of the cases -- at least for the current release. Thanks for doing this.
ROFL!!! “Good, it deserved that…” Thank you AiDriver and Carol! I’m glad Carol didn’t fall when rolling the wheel. ;) Carol is brave! Incredibly Awesome video
This is your funniest video so far! :). Keep up the good work and congrats to your fantastic wife. The support of your loved ones is important for reaching the next level! ❤
I had the issue of FSD avoiding puddles in the road. It kept changing lanes to avoid them. These puddles were bigger than potholes and a human would know they could safely be driven through (not a flooded road, just puddles in some lanes)
@@phyzygy these were obviously not potholes. They caused the car to weave from lane to lane in moderate traffic. Cars in front were passing thru them without dropping so the AI should have deleted the difference.
On the 2nd attempt (StarLink box sideways), the box upclose looks exactly like the rest of the shadows on the ground. I think as it got closer it "forgot" there was something there (AFAIK current FSD has no object permanence) and just saw it as shadow.
@@superkikim4916 FSD beta doesn't use the radar even on cars that have it. I don't know if radar sees cardboard or not, it definitely doesn't see some materials. Radar is mainly for metal (other cars). For this case vision is definitely better.
its amazing how a video like this can make me laugh more than a tv-series episode designed by a large number of professional comedians to make me laugh.
It's not a height problem but a memory problem, everything above 15cm is recognized as an object by the cameras but as soon as the object is outside the camera angle, the software forgets that there was an object and decides to drive over it.
That's what I'm thinking as well, as soon as it leaves the camera vision it leaves the neural networks understanding of the world. The car happily reroutes to a straight forward line through the object because as far as it is concerned the object no longer exists.
I think when the AI sees an upright car wheel in an orientation that is perpendicular to the direction the AI is driving it thinks it’s part of a car backing out into the street which is why is doesn’t want to go around and instead chooses to wait.
Deke - It seems like it is taking the "Training" reasonably seriously! 🤔🧐 in watching the Repeats, it also seemed to me it was figuring out that it was not good to hit it, so the next try, moved over to avoid the Obstacle! Cats & Squirrels? Low Height "Objects" - But, the are not always "Moving!" 😳🤔🧐
My theory in the beginnin: As soon as the car can't see the object anymore it also changes paths. But then the flat wheel should've also not been picked up. Another test with a cardboard box pulled by a string across the road would be really nice. Then you can "safely" slam into it instread of a wheel. As always a very interesting video and I hope FSD will come to Europe soon!
Very nice video! But I think you missed the brief moment at 2:22 where FSD did visualize the cardboard. I hope this system gets more control in the future and isn't overridden at the last second. Having gotten rid of the ultrasonic sensors, Tesla seems to want to develop the occupancy network much further. After that, the test should be a little more positive, since the close range is probably the problem.
Thanks to your wife for assisting in the testing. Tesla FSD Beta handles moving objects much bettie as the algorithms are focused on moving objects in the road. It is the stationary ones the technology struggles with. 1) What will FSD do when there is a massive pothole in the road? 2) What will FSD do when the object in the road blends into the road surface? 3) What will FSD do if a person wearing black had a green light on their head? I have seen people investing time to discover how they can trick FSD into seeing something and making wrong decisions based on what it 'sees'. Great work.
Your channel literally gave me interest into tesla's autopilot, something that I didn't really care about before That just shows the quality of the content you create, so thank you!
Really nice benchmark. Maybe to time consuming, but if you manage to do this for all new releases or every other, it would surely show us when they release those avoidance updates.
I think the false positive prevention threshold is the reason it simply doesnt react. Its safer to let the beta team annotate some data and swirve manually, rather than the car slamming on the brakes one too many times. If this is the case, the system should become much better over time.
This. They have to balance between making the drive safe and comfortable. Most probably on the roads there would be many more false positives then not recognised real cases. In other words, there are fewer cases of having a box in the middle, than having a kid running through the road, or a motorbike braking, etc, etc.
@@SzabolcsSzekacs the vast majority of difficulties in software come from crossing the 99.999% threshold. Tesla can release a unified stack with parking and madmax california level merging right now, but it'd only work like 95% of the time and be offensively unsafe. capability != ability
It's like you and a million other folks are working for TESLA for free. Brilliantly designed system to collect data and user generated content at the same time saving massive amounts of money on marketing, advertising, data collection, research and development, quality content creation (good job guys and gals for producing such a good video) and super valuable consumer feedback. Buying some stock Monday as it seams to be the bargain of the century.
It doesn't seem to be temporal at all. As soon as the box is not visualized for one frame, the route planner instantly routes through where the box was. As humans we of course know that box is still there whether we see it or not. Boxes don't just vanish. I'm sure they're working on this - the box should not be visualized for many frames before the car assumes it blew away or whatever
It's a balancing act. You might falsely detect an obstacle from far away. If you remember things too long, as you get close, you might no longer see it but still remember you thought you saw something earlier. Of course, a human would either decide they can obviously see the road is clear as they get closer, or they would gently decelerate because they're not sure. "Gently decelerate" is generally the best response, but the issue might be that FSD gets false positives too often, and it would be constantly slowing and unsure of itself. So I guess if they can make the obstacle detection more accurate, they can turn up the temporal aspects so it remembers for longer.
If you have a small not too dangerous object on the road the safest way might be to hit it rather than swerving or braking hard. It might not have chosen the most optimal path, but arguably the safest. Of course a lot depends on the particular situation.
Your nuts to brave these tests at such high speed. I would recommend 15 or slower. But hey no accidents so we are good! I can’t wait for my FSD computer to be installed.
Your wife is a real trooper! If that had been me and my wife, she would have said, "Yeah, no way. How about I drive and you walk out in front of the car?"
Do you think it would make a difference if it had a camera on the front bumper to help detect lower objects? IMO: I think the closer you get it the car does not see the object since the cameras are on top. As you can see it makes corrections the further you are but as soon as you can close to it FSD like kind of forget the object is there.
i doubt it. its judging the object as it approaches and it definitely knows its there. id say the difference is in what its detecting it as. it might even take into account the drivers eyesight. IE. FSD beta watches your eyes, if it knows your looking forward and can see it and you are ignoring it, then FSD can probably ignore it since you would take over if it was something it shouldnt hit. This isnt something they would want to rely on in the final version, but for training,... you never know.
Thank you for doing this! It's super, super interesting. I've been riding in Cruise AVs a lot in SF (so these scenarios you're testing for actually happen sometimes lol especially the truck ones), and the main difference I see is that Cruise seems to react much quicker and more confidently in situations like obstacles and pedestrians jumping into the street and further ahead with stationary obstacles than FSD does. I wonder if it has something to do with the power of the computers on board? The whole trunk of a Cruise is full of computers and you can hear the fans blowing away. I think when Tesla figures this out they'll be the leader because their system is more flexible, but as of now Cruise and Waymo seem to have quicker and more confident decision making.
I'm hoping Tesla has this kind of testing internally as well and that it's high up on the list of things to fix... Elon said many time the prime directive is "Do not crash" so I'm expecting much better results.
@@ericwiese7479 It's a very nice technology. Just compare what it's aware of to what humans are comprehending. This could make the real time decisions way more complex then what humans would consider in the moment. If I stop hard then I will cause a 5 car pileup behind me since 2 of the cars are not likely to have auto braking and none have decent following distance. If I slow to 10mph I'm 87% likely not to cause permanent damage to the adult human running in front of me, And 93% likely to avoid the 5 care pileup behind me. With a 40% chance of loosing a 5 million dollar lawsuit for not stopping...
@@cifey definitely in its infancy. The real world data the FSD beta testers are supplying Tesla’s Dojo supercomputer will help a lot. Plenty of data points
I don't think it's learning so much as choosing a different route each time. When it's planning its going through it's options many times each frame. Eventually it settles on a plan and executes that, so I think your seeing it choose a poor plan the first time and then the second time, with the situation slightly different, it chose a better plan.
I thought the same. I didn't think it did any sort of learning whilst driving (just executed the model it already had) but the tests did seem to indicate it might be remembering that the first path was a mistake.
@@clonkex It appears that way but I don't think that's how the neural net works. Its an illusion brought about by minute differences in each attempt forcing a different planning outcome. They train at home base then deploy to all. That way they can retain consistency across the fleet. The FSD computer architecture doesn't even have much on board data storage that I'm aware of. The memory would be flooded with data too quickly to be useful for a purpose like learning on the fly. If he had attempted this then gone home and received a new update awhile later then attempted again that's where the learning could occur and you may see actual improvement in understanding and executing the maneuver.
Makes sense actually, from far away the top of the box would likely appear above the road, making it a clear 3D object. As the car approaches the angle between the camera, the box, and the ground grows. This perspective probably causes the car to think that these are just road markings. Putting a camera as low as possible in the front of the vehicle would help resolve this.
I heard that tesla wants to erase ultrasonic sensors called parking sensors. Some low and small obstacles are below the hood of the car, so the cameras do not cover it.
If the car is in motion, it should be able to see such objects over time and model their behavior. When parked, I expect an updated Sentry Mode can model the changing world around the vehicle to prevent accidents when returned to driving condition.
Great video, very insightful! I think it makes sense to have a vertical threshold. When the car is far away the uncertainty about the height is likely higher so it starts planning for it but then as it sees it's under some vertical threshold it ignores it. As they improve the occupancy network they should be able to lower that threshold without introducing false positives that could lead to "phantom braking".
In the cardboard one where the Tesla ran over it the first and then went around it the second time. It could be possible that the Tesla is some how learning about it and fixing it’s mistakes.
Great test! Looking forward to more targeted testing like this. I suspect the detection system might be dialed to not be overly sensitive in regards to potential objects in the road since the occupancy network is new and false positives would lead to a lot of phantom braking. I expect them to tweak the sensitivity over time as the system gets more confident in its detection.
Good job on that video to both of you :) I have yet to let the car handle things when there are crossways involved, here in norway pedestrians have the right of way if there are crossways without traffic lights, and I have yet to see that FSD accknowleges that.
The whole thing proved how inteligent it is. It didn't stop for some cardboard cause it literally knew its cardboard. It can detect hollow objects. See the tire (layed) was same height as cardboard and it detected that it can go thru cardboard but not thru tire. That's a good thing cause imagine if it did full emergency braking for plastic bag on the road. It's kinda crazy how smart it is. And for the fast tire welllll it would stop eventually.
Excellent assessment and real world testing w/ FSD and reaction times. What’s also interesting, is that FSD seems to be LEARNING as you move through similar iterations of obstacles. I could be wrong here, but I really think the R&D engineers purposely have the BETA algorithms feed data to the Tesla servers when multiple mimic obstacles are presented within under minutes of each other. Even so within a 24 hr period. This tells me that Tesla and Elon has a learning software that gets BETTER at dodging obstacles on that GPS side street. Could it know the obstacles iterations within the GPS locations?! Perhaps. But clearly the computer and software IS LEARNING! Thanks again for your assessment. Thank you for sharing with consumers. I’m here for technology and science to help make the world a more safer place, all the while enjoying creature comforts. Big oil industries days are limited - as more and more consumers are introduced to EV technology. EV and AI taxi 🚕 cars are coming. Lastly, I have two FSD vehicles in my arsenal. I build up my confidence utilizing FSD on non surface streets all the time. From my subdivision/community mailbox, to my home. FSD is brilliant in knowing how to switch lanes- AFTER I PURPOSELY park next to mailbox, in the opposite incoming lane. I start up the FSD-SUMMON feature (without anyone in the car), and every single time- it pulls out from the opposite lane, and merge to the correct lane, and travels to my house-(while I walk along side monitoring) with neighbors mouth wide open! They simply can’t believe this technology is here. One neighbor tried to call the police and threaten me that I shouldn’t trust the technology because the car could accelerate into their house!!?? 😂 😂. I assured them- that it would not! They still recorded me, and sent the video to the local sheriffs office. I’m still awaiting my citation. Anyway, thanks for educating the public in EV, AI, and FSD TESLA cars. Tesla is far more advanced than any other car/technology company when it comes to transportation.
It seems like obstacles that it can’t see after a certain distance it ignores. If you look at what it hit (or got close to) they were mostly objects that barely were tall enough to hit the bumper of the car.
Great video so far. I am 3:27 seconds in. Idk if u do this later in the video but if u don’t next time car u try with 2 cars u behind the other car and we can see how fast Tesla reacts to it. Cuz 75% of the time we are trying to avoid a accident/debris
It's a pretty normal occurrence for a piece of cardboard to blow onto a highway, in those cases, it's actually important that the car to keep driving and go over it. So I'm not sure whether it's doing anything wrong here. I guess it depends on whether it can actually tell it's a box. I'd love to see a test done with a short object that isn't a cardboard box (or is, but has been made to look like something else)
It's not recommended to drive over cardboard. You never know what's inside. When you can see it from a distance, the best response is to go around, changing lanes if necessary and safe. In this case the car had ample time to safely go around, but it kept ignoring it once it got close.
This one reminds me of the British „tesla driver“ and his girl doing some testing in the backyard some two years ago. They were even funnier but not so scientific and fsd was just a wet dream of a few…
