"But the news said Tesla will run over children!"-my mother yesterday. I asked if her car would avoid running over children, she said, "yes, because I am driving." I pointed out that the Tesla had a driver who forced the car to run over the "children". She was confused. "Why?" Because the guy who made the video owns a company that competes with Tesla FSD. "That asshole should be locked up for that." Gotta love her!
Wouldn't it be nice if we could train our neural nets to avoid collisions with fear based memory barriers like the one created after watching that news broadcast?
@@captainchaos3667 How do you know? The entire enterprise is a farce from the 1960's vid quality so you can't see the messages on screen to the fact the it's produced by a conflicted Co. (provides 'competing' tho vastly inferior software) to blatant deception as a matter of habit. O'Doud is a know fraud.
@@captainchaos3667 the interior (well two of the three anyway) shots were very blurry but it did look like the car tried saying that it wouldn't brake due to the driver pressing the accelerator in one of the shots. also due in part to that blurryness (also due to camera angle) there is no way to confirm that the driver wasn't interfering
This has tremendous implications for the Optimus robot; imagine imaging a chair this way as opposed to a bounding box! You can't sit in a bounding box!
Karpathy did an amazing job, but the baton has been passed. Much of this nomenclature feels like it has evolved beyond what TESLA has previously been communicating. FSD is in good hands and accelerating.
Ever since Elon began talking about vector space I started believing FSD was not only possible with vision only, but could only be achieved with vision based neural nets. Lidar truly is a crutch. The way I explain it to people is: "Do you think your computer could theoretically drive a car in GTA perfectly, given the fact that the game knows everything that's going on in the road? Because Tesla is more or less recreating that game world in real time inside of a Tesla, and then letting the computer drive the car within that game." What a time to be alive!
I am continually amazed about the features/functions the software is capable of with the hardware. When they get bored, auto seat belts for general safety and as a restraint for unruly occupants in the vehicle would be weird but useful; I wouldn't want to have to write that explanation for a manual or legal filing.
The person sitting on the ground in the road (18:30) is so serious. This happened to me earlier this year, driving down a residential road, saw some debris next to the car that would be safe to run over. And then the late realization that the debris was actually legs belonging to a person who was working under their car. Fortunately for him and my sanity, I avoided running over that guy's legs. (Great coverage btw! Thank you!)
Super job explaining Occupancy Networks Dr Know It All. In my humble opinion, Occupancy Networks produce a sort of vision for the car similar to the vision that we humans have.
Thank you for your work Dr know it all. I watched Ashoks presentation yesterday and sort of understood it. I knew I would need your translation to fully process it. I tried to watch the presentations from the other speakers at that conference but their topics were even more esoteric and difficult to make sense of to me as a non-software educated layman. It would be interesting to see what cruise and waymo were talking about in their presentations to see just how far off they are from what Tesla Autopilot is doing.
They win on a technicality and it will always be what it is. HD maps can't scale, and LIDAR needs HD maps if the vision system paired with it isn't good enough (and if the vision system is good enough it in turn does not need lidar).
Took a graduate course on neural networks as an undergraduate in 1988. My topic was motion vector fields. But there was just no basis for doing anything useful back then. Born out of sync!
Me to. I did Machine code on a 8080. Made a traffic light intersection. Worked the first run (no bugs). Others in the group did the hardware. The good ol days. 1984.
If there is doubt about camera optics vs lidar. Food sorters have been around for years. They use camera optics with usually a blue background belt conveyor. Once tracking a snapped bad shape, or color it spits it out by a small air puff solenoid rail. Quite amazing. We call them $500,000 dollar 6 ft conveyors. They are 2D. The hard part is getting the object to remain steady on the belt while tracking. They use much computing power. Ambient light has to be removed. Thanks to optical cameras for helping make our food safer.
Geometry and generalization are important, but ontology is still very relevant when making decisions at high travel speeds where braking can cause fiery pile ups. eg. tumbling styrofoam and hovering plastic bags on the highway.
Am I right in thinking that, over time, these materials can be characterised as lightweight via computations like slowing in air, density via gravity versus air slowdown, recognised to be lightweight, and thus less of a problem compared to , say, bricks falling off a lorry. Humans ( I presume) either through inattention, or insufficient time, will also make mistakes...maybe, if the object is bundles of dollars, the Tesla will say " You are advised to pick up as much of the object as possible..."
Nice video for us nerds :) Nothing wrong with a bit of tech talk in the afternoons. I think that the occupancy network will be very useful for Optimus. Low speed, plenty of latency to process the where am I, whats around me, what position are my limbs and head in. Jogging at 5mph down the path, avoiding people and phone poles etc. doesn't seem like too far of a stretch with that model grinding away.
As I watched, I was struck by the thought that the problem they are trying to solve is very similar to the NC machining problem. NC Machining Problem: Compute a safe “tool path” through a valley of interfering 3D obstructions Autopilot Problem: Compute a safe “vehicle path” through a valley of interfering 3D obstructions (with the additional challenge of doing it in real time)! Back some 30+ years ago, I worked with a research team to explore different software approaches to solving the NC machining “interference problem”, (that problem being our computed “Tool Path” kept cutting through adjacent parts of the 3D model because the “tool path” was unaware of the many 3D obstructions surrounding it.) Part of the solution to that general NC problem, which we came up with, was to capture every possible “obstruction” in the adjacent area and build a kind-of 3D “protective bubble” around each item. Then compute the tool path and check it didn’t intersect any bubble, and correct the path if required.
the problem I have had with FSD is that Tesla did not preserve speed limit map data . This means that until it sees a road sign it assumes the road is 25 mph for many roads. Roads that have had the sames speeds for nearly two decades and oddly my car used to know these speeds. So when you cross an intersection and the road name changes the car will adopt a 25mph as the default speed which can be very jarring and require immediate correction
You've now graduated from "Dr know it all to Professor know it all"! Tesla's approach to solving autonomy is the best and only way even though it might seem insurmountable now, I really believe they will get there eventually as the team keep working on it! FSD level 5 might even be closer than we think! I also believe if there is a way to fuse Camera and Radar into one device, we reap the advantages of both as one unit! If this could be done, it will really simplify data collection and computation by the NN!
Also been concerned if HW3 has sufficient compute. Glad to see they don’t have to update every HW2.5 computer. People with HW2.5 may need to upgrade to get the safety features.
What do you think about Teslas being able to honk their own? Good or Bad Idear? In my opinion, this is a feature that is by far the most needed. If it becomes a dangerous situation, the Tesla should signal this not only to the driver but also to its surroundings, just as a normal car driver would do. Even better would be a honking net that not only honks in really really dangerous situations, but also when other participants make mistakes that would force the Tesla to take evasive action, i.e., alerting the others to the mistake with a short one-time honk even before its own evasive maneuver.
I have always predicted that Tesla will enable something like Collision Avoidance to all the customers signing up for Tesla Insurance regardless of Autopilot or FSD purchase.. Can you imagine 100's of thousands (or millions) of people paying for Tesla Insurance and Tesla not needing to pay anything back (because no crashes)..
Occupancy network is claimed to run in 10ms, but the demo shown at 7:44 still has the same lag as all previous FSD versions, i.e. ~330ms, so it seems there is still a fairly long way to go to get the total lag down to the sub 50ms range needed for a real-time system.
With voxels, you can arbitrarily increase the resolution close to the vehicle (or any point of interest) without a great penalty to the whole scene because of the octree representation usually used.
Thanks for the summary, even if I'm still struggling with the basic AI technology - have to go back and re-view your primers. Good comments in here regarding the collision threat from behind, especially if being tailgated by a human. Does the occlusion modelling include objects approaching from behind? That's a non-trivial concern if travelling in dense traffic.
I don't know but on basic Autopilot on curvy highway driving it gets uncomfortably close to the shoulders, specially on 2 lane highways. I have to either lower the speed to the speed limit or disengage. I hope this new architecture helps.
You mentioned that the occupancy network isn't going to produce voxels that are accurate in the centimeter scale. They're going to try and do just that with the removal of USS. What are your thoughts?
Very good explanation - thank you very much for that. This is another brick in the wall.. would be great to have this collision avoidance in every car. Bad perspectives for breaking through police roadblocks with getaway cars 😁 Is this already implemented in 10.69?
I find the collision software technology intriguing. Though I don't think we will see you level 5 autonomy for a number of years if ever. I believe this collision avoidance technology is very doable in the next few years. Even if they can only reduce the accidents caused by the driver by 50% it would be an incredible achievement. Think of how the value of Tesla cars will increase. Tesla insurance be able to reduce prices dramatically. And then other car companies will be forced to compete and create their own systems. Who would want to buy an unsafe car? The effect on the cost of driving and lives saved would be tremendous.
Is this 'no need to recognize (and label) what occupies the road' what George Hotz was saying all along and predicted Tesla will also switch to this approach? BTW, get the man on your channel! :)
Second comment about the last section, while it does a good job of keeping the car from having a collision when the driver presses the accelerator by accident shouldn't it also try to obey traffic signals? Doesn't matter if you are not going to hit something if you are just blowing through stop signs and lights.
I think the best solution to satisfy everyone would be a big red "override safety" switch under a cover, so safety can be ditched for freedom at the user's risk. Of course, libertarians won't be happy with any vehicle that can be hacked or stopped by an EMP weapon, but it'd be enough for street racers, bank robbers etc.
I would think once we get to level 4/5, a requirement would be some kind of signage that alerts other drivers that this car potentially could be autonomous. Sort of like the 'Student Driver' signs on cars today.
I was wondering if the grey scale on the voxels was like a signed distance field, so it gives more detail by providing the distance of the true surface from the center of the voxel. Notice that sign posts are darker at the bottom as is the road.
Collision Avoidance: It would be nice to have the horn sound or some type of LOUD recorded sound be played i.e. WAKE UP to actually warn other drivers that they are actually making a DUMB move so that they can take appropriate measures to keep from hitting you. Recently I actually had to cross a double yellow line and drive my car off the street and almost into a tree to avoid a car that exited a parking lot on my right side and pulled directly into my lane of traffic with very limited notice. Luckily, no one was injured, but it was VERY close. If I had not been alert and no oncoming traffic was present I would have broadsided the offending driver. It was a lucky day for both me and the offending driver. I am glad Tesla is working on Collision Avoidance as I am sure it will save millions in vehicle damage and millions of lives. Thank you Elon and your intelligent employees for working on such a worthy project and providing the necessary solutions. You are the man!!!!
I like it detects a Motorcycle slowing down way ahead of me on a fast traveling highway. They are smaller, lighter, and break way faster. Driving my M3P I witnessed this in Autopilot Saturday. I would have not known otherwise until much closer. Likely with a hard human brake unless designated attention was placed on it by my eyes not looking at anything else.
Excellent video. Would it be possible for you to create a details playlist with all the sub net that go into Tesla FSB Beta as one video and how it all works together. Then each sub net type then get broken down and explained in technical details as a sub video on it own. Then maybe turn this into a training course at some point on udemy or something like that if you like. Or maybe you could then just go on to teach this as your academic institution as a course or even put it on coursera or edx. Just ideas for you as you are doing an amazing job.
I don't recall the exact numbers, but I think Tesla's annual report/??? talks about 30% or some such number of Tesla's accidents are rear enders.... I wonder if the non FSD vehicles will start to try to avoid these as well? Or, if non FSD vehicles will overide steering/gas/brakes from user to avoid front/side impacts that are non Tesla driver caused..... We can call these idiot accident avoidance scenarios.... In other words, FSD is about the ego car driving and getting to a destination. Active accident avoidance is about minimizing other vehicles from hitting a Tesla..... Is this what you are saying that is coming as a standard feature?
I wonder what percentage rear-ender accidents are for all accidents with all the "distracted" driving going on out there. It seems very high everywhere
Does Tesla have any plans underway to inform the fleet of “invisible” accidents in the highway ahead? This could include whiteouts, upcoming crash scenes, flash floods, mudslides, etc. In addition, this general danger could be provided to following ICE cars by activating the Tesla’s flashers!
Runs the raw image through a ragnet and a bi-fpn. Oooh is that all. I'm very thankful for this community that understands the implications of all these AI advancements. I do not. I understand hardware more than most as a dealership tech but AI coding escapes me completely. It's ironic that the big T will help me retire as OEM's and my people fade out of relevancy.
You mean something like this: "Ford laying off 3000 white-collar workers ... as part of a broader restructuring to sharpen focus on electric vehicles and batteries"
Hardware can replace software vice versa. The hardware used for this is dedicated toward its application. This frees up computational software code. For example a GPU is dedicated toward graphics.
What if you need to hit a person? Ex: law enforcement need to run over a suspect shooting...can that still happen? Is there an off switch, or does it take over?
His name is pronounced a-shóke. Long O, accent on the 2nd syllable. A shocking revelation, I know. Just listen to the first seconds of his Occupancy Network talk where he introduces himself
So, $15,000 for FSD and $60,000 for the car. So not affordable to the average person. Thanks Elon! Thought Tesla hired smart people yet they couldn't even spell the word doesn't, it was shown as "Does't" on the slide. lol
Implying the cameras have a 360 degree field of view is not accurate. Combined they can produce a 360 degree view. The Angular field of view and the Tesla cameras vary from 35 to 150 from what details I could find.
Geometry over ontology... I suppose Teslas will no longer run at full speed into stationary or overturned trucks blocking the highway? it's not a panacea either, imagine innocuous light-weight objects that wind can blow onto the road, like plastic bags - the car would now break or swerve to avoid them...
Tesla crash is totally not a threat for my 100% portfolio of Tesla stock because I am so angry that I have to wait two years to get my model S in Taiwan, so please withdraw your order and let me get immediately
"But the news said Tesla will run over children!"-my mother yesterday. I asked if her car would avoid running over children, she said, "yes, because I am driving." I pointed out that the Tesla had a driver who forced the car to run over the "children". She was confused. "Why?" Because the guy who made the video owns a company that competes with Tesla FSD. "That asshole should be locked up for that." Gotta love her!
Wouldn't it be nice if we could train our neural nets to avoid collisions with fear based memory barriers like the one created after watching that news broadcast?
Only deformed or Woke children.
That's not true though. The driver didn't "force" the car to do anything, he didn't touch the accelerator or the brake pedal.
@@captainchaos3667 How do you know? The entire enterprise is a farce from the 1960's vid quality so you can't see the messages on screen to the fact the it's produced by a conflicted Co. (provides 'competing' tho vastly inferior software) to blatant deception as a matter of habit. O'Doud is a know fraud.
@@captainchaos3667 the interior (well two of the three anyway) shots were very blurry but it did look like the car tried saying that it wouldn't brake due to the driver pressing the accelerator in one of the shots. also due in part to that blurryness (also due to camera angle) there is no way to confirm that the driver wasn't interfering
This was highly useful. Thanks so much for presenting this so eloquently and understandably.
This has tremendous implications for the Optimus robot; imagine imaging a chair this way as opposed to a bounding box! You can't sit in a bounding box!
That robot can't shit in a box
Karpathy did an amazing job, but the baton has been passed. Much of this nomenclature feels like it has evolved beyond what TESLA has previously been communicating. FSD is in good hands and accelerating.
Agreed
Technological apex, at the height of elegance.
(What a beautiful thing.)
Ever since Elon began talking about vector space I started believing FSD was not only possible with vision only, but could only be achieved with vision based neural nets. Lidar truly is a crutch.
The way I explain it to people is:
"Do you think your computer could theoretically drive a car in GTA perfectly, given the fact that the game knows everything that's going on in the road? Because Tesla is more or less recreating that game world in real time inside of a Tesla, and then letting the computer drive the car within that game."
What a time to be alive!
Earlier this year I was driving ~
I am continually amazed about the features/functions the software is capable of with the hardware. When they get bored, auto seat belts for general safety and as a restraint for unruly occupants in the vehicle would be weird but useful; I wouldn't want to have to write that explanation for a manual or legal filing.
amazing, no other car company does this, this is insane levels of expertise
The person sitting on the ground in the road (18:30) is so serious. This happened to me earlier this year, driving down a residential road, saw some debris next to the car that would be safe to run over.
And then the late realization that the debris was actually legs belonging to a person who was working under their car. Fortunately for him and my sanity, I avoided running over that guy's legs.
(Great coverage btw! Thank you!)
Wow - just Wow!
Super job explaining Occupancy Networks Dr Know It All. In my humble opinion, Occupancy Networks produce a sort of vision for the car similar to the vision that we humans have.
You break this down very well. You must be quite a good teacher.
Good explainer for the average person. Thanks Dr. Know It All!
John. Great summary of Ashok's keynote. You made this so fascinating, I need to watch the entire video.
Thank you for your work Dr know it all. I watched Ashoks presentation yesterday and sort of understood it. I knew I would need your translation to fully process it. I tried to watch the presentations from the other speakers at that conference but their topics were even more esoteric and difficult to make sense of to me as a non-software educated layman. It would be interesting to see what cruise and waymo were talking about in their presentations to see just how far off they are from what Tesla Autopilot is doing.
Thank you Doc. My head is about to explode...in a good way.
Great explanation. THANKS!
Tesla FSD is the only scalable solution for autonomy, even when others are claiming to be further ahead with their maps and Lidar.
They win on a technicality and it will always be what it is. HD maps can't scale, and LIDAR needs HD maps if the vision system paired with it isn't good enough (and if the vision system is good enough it in turn does not need lidar).
Took a graduate course on neural networks as an undergraduate in 1988. My topic was motion vector fields. But there was just no basis for doing anything useful back then. Born out of sync!
Me to. I did Machine code on a 8080. Made a traffic light intersection. Worked the first run (no bugs). Others in the group did the hardware. The good ol days. 1984.
Love these deep dives!
Nice explanation, very informative.
I'm not smart enough to understand any of this but I enjoy listening
this release has solidified my confidence that fsd will happen soon-ish. the short sightedness of wallstreet is astounding.
They like to see "proof" via increases in quarterly earnings - and it seems with the latest release, that may be soon...
This is so so so helpful, thank your my teacher!
Amazing developments! Thank you.
If there is doubt about camera optics vs lidar. Food sorters have been around for years. They use camera optics with usually a blue background belt conveyor. Once tracking a snapped bad shape, or color it spits it out by a small air puff solenoid rail. Quite amazing. We call them $500,000 dollar 6 ft conveyors. They are 2D. The hard part is getting the object to remain steady on the belt while tracking. They use much computing power. Ambient light has to be removed. Thanks to optical cameras for helping make our food safer.
Geometry and generalization are important, but ontology is still very relevant when making decisions at high travel speeds where braking can cause fiery pile ups. eg. tumbling styrofoam and hovering plastic bags on the highway.
Am I right in thinking that, over time, these materials can be characterised as lightweight via computations like slowing in air, density via gravity versus air slowdown, recognised to be lightweight, and thus less of a problem compared to , say, bricks falling off a lorry. Humans ( I presume) either through inattention, or insufficient time, will also make mistakes...maybe, if the object is bundles of dollars, the Tesla will say " You are advised to pick up as much of the object as possible..."
Cool Romulan Warbird 👍
More, please! Great job.
Nice video for us nerds :) Nothing wrong with a bit of tech talk in the afternoons. I think that the occupancy network will be very useful for Optimus. Low speed, plenty of latency to process the where am I, whats around me, what position are my limbs and head in. Jogging at 5mph down the path, avoiding people and phone poles etc. doesn't seem like too far of a stretch with that model grinding away.
Fantastic deep-dive in to the continued amazing FSD, and another great video, thanks :)
How do you think this would handle the trolley problem?
As I watched, I was struck by the thought that the problem they are trying to solve is very similar to the NC machining problem.
NC Machining Problem: Compute a safe “tool path” through a valley of interfering 3D obstructions
Autopilot Problem: Compute a safe “vehicle path” through a valley of interfering 3D obstructions (with the additional challenge of doing it in real time)!
Back some 30+ years ago, I worked with a research team to explore different software approaches to solving the NC machining “interference problem”, (that problem being our computed “Tool Path” kept cutting through adjacent parts of the 3D model because the “tool path” was unaware of the many 3D obstructions surrounding it.)
Part of the solution to that general NC problem, which we came up with, was to capture every possible “obstruction” in the adjacent area and build a kind-of 3D “protective bubble” around each item. Then compute the tool path and check it didn’t intersect any bubble, and correct the path if required.
Brillian explanation. You do a great job. Now i think i understad more why EM is sure about fsd comming. Thank you.
I'm now convinced that FSD success is not a matter of if, but when.
Thank you for getting into the weeds about this DKA!
More geeky content! This was a great review and explanation set! Thanks for taking the time to make this
Thank you! Cool stuff reminds me of encryption!
the problem I have had with FSD is that Tesla did not preserve speed limit map data . This means that until it sees a road sign it assumes the road is 25 mph for many roads. Roads that have had the sames speeds for nearly two decades and oddly my car used to know these speeds. So when you cross an intersection and the road name changes the car will adopt a 25mph as the default speed which can be very jarring and require immediate correction
You've now graduated from "Dr know it all to Professor know it all"!
Tesla's approach to solving autonomy is the best and only way even though it might seem insurmountable now, I really believe they will get there eventually as the team keep working on it!
FSD level 5 might even be closer than we think!
I also believe if there is a way to fuse Camera and Radar into one device, we reap the advantages of both as one unit! If this could be done, it will really simplify data collection and computation by the NN!
🤗👍 I AM PREPARED TO BE AMAZED 😎💚💚💚
🤗🤩🤩 AND I AMAZED THANK YOU DR. FOR SHARING THIS WITH US LAYPEOPLE UNDERSTANDABLY 🤔💚💚💚
Potholes voxel are missing.
Still Wow, really really nice solutions
Also been concerned if HW3 has sufficient compute. Glad to see they don’t have to update every HW2.5 computer. People with HW2.5 may need to upgrade to get the safety features.
What do you think about Teslas being able to honk their own? Good or Bad Idear? In my opinion, this is a feature that is by far the most needed. If it becomes a dangerous situation, the Tesla should signal this not only to the driver but also to its surroundings, just as a normal car driver would do. Even better would be a honking net that not only honks in really really dangerous situations, but also when other participants make mistakes that would force the Tesla to take evasive action, i.e., alerting the others to the mistake with a short one-time honk even before its own evasive maneuver.
Awesome video. Super informative. You should send it to Chuck (unprotected left turn Chuck).
I have always predicted that Tesla will enable something like Collision Avoidance to all the customers signing up for Tesla Insurance regardless of Autopilot or FSD purchase..
Can you imagine 100's of thousands (or millions) of people paying for Tesla Insurance and Tesla not needing to pay anything back (because no crashes)..
You are a clever robot 🤖🤗👍😎💚💚💚
@@budgetaudiophilelife-long5461 Dang, what gave it away?
@@revaddict your initials are the same as mine 🤗😁💚💚💚
Once again GREAT VIDEO
"LIDAR is a fools errand". -Elon Musk in 2019.
Occupancy network is claimed to run in 10ms, but the demo shown at 7:44 still has the same lag as all previous FSD versions, i.e. ~330ms, so it seems there is still a fairly long way to go to get the total lag down to the sub 50ms range needed for a real-time system.
With voxels, you can arbitrarily increase the resolution close to the vehicle (or any point of interest) without a great penalty to the whole scene because of the octree representation usually used.
I don't see the link to the presentation anywhere?
Thanks for the summary, even if I'm still struggling with the basic AI technology - have to go back and re-view your primers. Good comments in here regarding the collision threat from behind, especially if being tailgated by a human. Does the occlusion modelling include objects approaching from behind? That's a non-trivial concern if travelling in dense traffic.
I don't know but on basic Autopilot on curvy highway driving it gets uncomfortably close to the shoulders, specially on 2 lane highways. I have to either lower the speed to the speed limit or disengage. I hope this new architecture helps.
You mentioned that the occupancy network isn't going to produce voxels that are accurate in the centimeter scale. They're going to try and do just that with the removal of USS. What are your thoughts?
Another great video!
Amazing!
Very good explanation - thank you very much for that. This is another brick in the wall.. would be great to have this collision avoidance in every car.
Bad perspectives for breaking through police roadblocks with getaway cars 😁
Is this already implemented in 10.69?
18:20
You'd think its a pile of garbage, but its actually a human being
I'm sorry but this whole section was just unintentionally hilarious
I find the collision software technology intriguing. Though I don't think we will see you level 5 autonomy for a number of years if ever. I believe this collision avoidance technology is very doable in the next few years. Even if they can only reduce the accidents caused by the driver by 50% it would be an incredible achievement. Think of how the value of Tesla cars will increase. Tesla insurance be able to reduce prices dramatically. And then other car companies will be forced to compete and create their own systems. Who would want to buy an unsafe car? The effect on the cost of driving and lives saved would be tremendous.
Is this 'no need to recognize (and label) what occupies the road' what George Hotz was saying all along and predicted Tesla will also switch to this approach?
BTW, get the man on your channel! :)
Second comment about the last section, while it does a good job of keeping the car from having a collision when the driver presses the accelerator by accident shouldn't it also try to obey traffic signals? Doesn't matter if you are not going to hit something if you are just blowing through stop signs and lights.
I think the best solution to satisfy everyone would be a big red "override safety" switch under a cover, so safety can be ditched for freedom at the user's risk. Of course, libertarians won't be happy with any vehicle that can be hacked or stopped by an EMP weapon, but it'd be enough for street racers, bank robbers etc.
$15k is just so out of reach for so many people.
loved it!!!!!!
If it was released on 8/20
And presented on 6/20
When was it written? 🤔
I would think once we get to level 4/5, a requirement would be some kind of signage that alerts other drivers that this car potentially could be autonomous. Sort of like the 'Student Driver' signs on cars today.
I was wondering if the grey scale on the voxels was like a signed distance field, so it gives more detail by providing the distance of the true surface from the center of the voxel. Notice that sign posts are darker at the bottom as is the road.
Collision Avoidance: It would be nice to have the horn sound or some type of LOUD recorded sound be played i.e. WAKE UP to actually warn other drivers that they are actually making a DUMB move so that they can take appropriate measures to keep from hitting you.
Recently I actually had to cross a double yellow line and drive my car off the street and almost into a tree to avoid a car that exited a parking lot on my right side and pulled directly into my lane of traffic with very limited notice. Luckily, no one was injured, but it was VERY close. If I had not been alert and no oncoming traffic was present I would have broadsided the offending driver. It was a lucky day for both me and the offending driver. I am glad Tesla is working on Collision Avoidance as I am sure it will save millions in vehicle damage and millions of lives. Thank you Elon and your intelligent employees for working on such a worthy project and providing the necessary solutions. You are the man!!!!
I like it detects a Motorcycle slowing down way ahead of me on a fast traveling highway. They are smaller, lighter, and break way faster. Driving my M3P I witnessed this in Autopilot Saturday. I would have not known otherwise until much closer. Likely with a hard human brake unless designated attention was placed on it by my eyes not looking at anything else.
I think they could use HW3 for the Tesla Bot - so if they are upgraded to HW4 there is no need to throw them away. :-)
Excellent video. Would it be possible for you to create a details playlist with all the sub net that go into Tesla FSB Beta as one video and how it all works together. Then each sub net type then get broken down and explained in technical details as a sub video on it own. Then maybe turn this into a training course at some point on udemy or something like that if you like. Or maybe you could then just go on to teach this as your academic institution as a course or even put it on coursera or edx. Just ideas for you as you are doing an amazing job.
very cool
The audio in the left ear has more gain than the right ear. Noticed this in a previous video.
It is not very noticeable.
Why can't it park itself when at a supercharger?
Would be nice if our cars would stop before hitting something when we're not accidentally backing up too 🙄
8:09 PM
8/29/2022
I love your #Tesla Videos
Wow....Another person is already hired to be the Director. Good thing for Tesla.
Ashok has been the director of FSD for a few years, Karparty was the AI director.
Is there anyone know how to defeat the detection
Thanks John, I'm all super-geek'd up! Thank the Universe, Gaia & all the Gods for giving us clever engineers & nutty Professors ;)
MLP = "maximum likelihood predictor" ?
Did an adult review this video prior to release? We've not seen one this meandering or discombobulated! Best of luck!
I don't recall the exact numbers, but I think Tesla's annual report/??? talks about 30% or some such number of Tesla's accidents are rear enders.... I wonder if the non FSD vehicles will start to try to avoid these as well? Or, if non FSD vehicles will overide steering/gas/brakes from user to avoid front/side impacts that are non Tesla driver caused..... We can call these idiot accident avoidance scenarios....
In other words, FSD is about the ego car driving and getting to a destination. Active accident avoidance is about minimizing other vehicles from hitting a Tesla..... Is this what you are saying that is coming as a standard feature?
I wonder what percentage rear-ender accidents are for all accidents with all the "distracted" driving going on out there. It seems very high everywhere
This explains why Karpathy left. Completely different approach than labeling.
Does Tesla have any plans underway to inform the fleet of “invisible” accidents in the highway ahead? This could include whiteouts, upcoming crash scenes, flash floods, mudslides, etc. In addition, this general danger could be provided to following ICE cars by activating the Tesla’s flashers!
Runs the raw image through a ragnet and a bi-fpn. Oooh is that all. I'm very thankful for this community that understands the implications of all these AI advancements. I do not. I understand hardware more than most as a dealership tech but AI coding escapes me completely. It's ironic that the big T will help me retire as OEM's and my people fade out of relevancy.
You mean something like this: "Ford laying off 3000 white-collar workers ... as part of a broader restructuring to sharpen focus on electric vehicles and batteries"
Hardware can replace software vice versa. The hardware used for this is dedicated toward its application. This frees up computational software code. For example a GPU is dedicated toward graphics.
❤️❤️❤️TESLA❤️❤️❤️
"But Tesla is just a car company..."
voxel = 3d pixel
It’s (Ah-Shoke)
My wife stomped on the accelerator instead of the brake two weeks ago.
Hit the car in front of her. :(
Many human drivers don't know where their car is within a meter of other objects.
That's tragedy
What if you need to hit a person? Ex: law enforcement need to run over a suspect shooting...can that still happen? Is there an off switch, or does it take over?
His name is pronounced a-shóke. Long O, accent on the 2nd syllable. A shocking revelation, I know.
Just listen to the first seconds of his Occupancy Network talk where he introduces himself
So, $15,000 for FSD and $60,000 for the car. So not affordable to the average person. Thanks Elon!
Thought Tesla hired smart people yet they couldn't even spell the word doesn't, it was shown as "Does't" on the slide. lol
Misinformation was cleaning up stairs? Weird household you have. We use a Vorwerk vacuum cleaner.
Implying the cameras have a 360 degree field of view is not accurate. Combined they can produce a 360 degree view. The Angular field of view and the Tesla cameras vary from 35 to 150 from what details I could find.
Has anyone ever told you that you have way too much advertising content? I sure think you do!!
Artificial multiple Insect Eyes.
Geometry over ontology... I suppose Teslas will no longer run at full speed into stationary or overturned trucks blocking the highway? it's not a panacea either, imagine innocuous light-weight objects that wind can blow onto the road, like plastic bags - the car would now break or swerve to avoid them...
Tesla crash is totally not a threat for my 100% portfolio of Tesla stock because I am so angry that I have to wait two years to get my model S in Taiwan, so please withdraw your order and let me get immediately