Great video! I have subscribed hoping for more content like this in the future. It would be awesome to have progress updates 2-4 times/year. Thank you for this video!
This is a great video. Good production values, interesting points of learning, and interesting subjects. ML is such a breakthrough in computing but also something that is almost foreign for older IT professionals. When I took a degree in computer science in the mid-to-late 1990is this didn't exist. I was thought religiously that the "out data" can never be better than the "in data". In layman terms: you can't take a digital photo and upscale it with a better resolution because you are limited to the indata. But with ML you can... With self-driving cars. How will the government exploit the camera data to map our lives or is a strong government a key to get safe self-driving? In China, they are making everything networked and data-linked to self-driving cars so the AI knows every school crossing, signal, and so on by data signals and not just a camera with AI. Well.. sorry for the old man rant at the end.
Great video! Just a quick note: LIDAR is NOT redundant for vision, as LIDAR cannot see. LIDAR is a "hardware diverse" method to simply detect the distance to a point in space. Hardware diversity and redundancy are completely different things. Only cameras will well-trained neural nets can give you "vision". I think of LIDAR like a blind person with a walking cane. It is a hardware diverse way of detecting the distance to an object if vision is unable to perform that function.
Thanks! I agree that it is not redundant for vision. However you do get a 3D map which should be good for driving with if you also have some other things like signs in HD maps. Traffic lights will be hard though... However, LIDAR actually does see a bit. Because in addition to measuring the "time of flight" which gets converted to a distance, you also get a value for the strength of the returning beam. This strength indicates how reflective the object is. So this can look a bit like a gray scale image for some objects.
Great content. I rarely subscribe to RUclips channels and I subscribed to yours. I’m really interested in learning about how viable Tesla’s approach is and if you think there are any fundamental flaws. I really appreciate also understanding more about MobilEye and Waymo.
Great video. Tesla may not be just uploading edge cases these days. We are in the fsd beta program and when we return home, it regularly uploads > 20 GB of data even when the drive has no disengagements (which are happening more frequently). It's quite possible entire videos of drives are being uploaded for automatic annotation and even possible mapping purposes (there are technical papers that show cameras can create similar datasets to lidar to give 3d maps of its environment). I suspect Tesla maybe quietly moving to replicate what Mobileye is proposing by using data collected from each car to develop high resolution maps. If each Tesla out of the 100K enrolled in the FSD beta program is uploading 20 GB each drive, it may provide a more rich dataset than what Mobileye is proposing. The dataset would be even more enhanced if the remaining Tesla fleet is also involved in data collection beyond just edge cases. It would be interesting to hear from other Tesla owners that are not presently involved with the FSD beta program are also experiencing increased data uploads after drives. I agree that once the differences in driving experience declines between the mass and rich data sets, it will be hard to justify continuing to rich data sets.
Thanks, very interesting comment! I also have this suspicion, somehow without "real intelligence", which humans have and computers do not, I think HD maps with at least lane lines and road edges might be what is needed to get enough robustness. Because the roads of the world are so wildly different. Humans use reasoning and logic to solve problems as they come, but cars might need HD maps... I think there might be diminishing returns of the HD map beyond lanes, road edges and voxels (mine-craft ich blocks of the surroundings). However as more compute is available offline maybe it is still better that the cars upload the raw image data and not the output. If Tesla gets HD maps, they will have "redundancy" for the world structure. Then maybe the question is if they can make vision only robust enough for cars/pedestrians/animals etc. Eventually maybe the companies will meet in the middle - Tesla adding a few cameras and increasing the camera resolution, Waymo perhaps removing some LIDARs. Question is who gets there faster. :) Will be interesting to follow.
Driving assistance systems and autonomous system are typically classified using SAE J3016. A few cars has autonomy in a limited operational domain. The Mercedes EQS and the S-class has a level 3 system a so called "eyes off" system (you can read a book or play a game on your phone) in stop and go traffic in speeds up to 60km/h. Honda has a similar system in Japan. All other systems are not autonomous, levels 0-2. In a Level 4 car you do not need to take over control ever, within the operational domain. It's like riding in a taxi. You are not driving. Such system are not legal anywhere in the world (for private operation), and you will not be able to buy such a car before 2030 likely,.
The real promise of self driving goes much further than safety. Subscribing to a robotaxi service promises to be way cheaper than owning your own vehicle. With such a fleet vehicles are utilized far more than private vehicles. They can also be sized to match the average occupancy rather than buying a large car for the rare occasion you might need it. Result is far fewer vehicles produced most of which are far smaller and more efficient. The real winner from this will be the environment, through quieter safer streets and vastly reduced emissions.
Good point, I agree - it will really reduce the cost for mobility. However, when it is very cheap to take a robotaxi, is there maybe a risk that the streets will not become quieter but instead more busy? I am thinking that people rather drive in their own private robotaxi than taking the subway etc.
@@lightinspace2963 I think it's definitely possible people will stop taking buses and trains, after all the will get door to door service. This is not as bad as it seems though because the load factor is so poor on buses especially that it's quite easy for a lightweight 2 person vehicle to emit less CO2 per passenger than a bus.
What few discuss with regard to the autonomous ride hailing approaches taken by Waymo, Mobileye, and Tesla is the economics of each. Nothing succeeds if it is not economically viable. For that reason, I think Waymo is driving down a blind alley. It's an over-engineered, expensive system that will never be profitable. Tesla's system is the most likely to succeed from an economics standpoint if the engineering issues can be solved, and it appears they will be. Mobileye will find a place, as well, but on a more limited scale. Mass data/rich data is not an accurate characterization of the two approaches. The data contained in Waymo's high definition maps is certainly voluminous, but unnecessarily so because vehicle localization can be achieved much more efficiently using GPS and vision. Tesla's occupancy networks serve the purpose of 3-D lidar maps for most purposes with more than adequate resolution. I like Mobileye's use of low resolution dynamic maps as a form of shared memory. I think Tesla would do well to implement this in their FSD system to deal with confusing lane structures, parking lots, driveways, speed bumps and potholes, road closures, etc. Thanks for the video.
Seems to me like the rich data approach is winning. The reliability has proven to be very difficult to improve. Every subsequent 9 after 99% seems to be exponentially more difficult to get. And there are a lot of 9 required to build public confidence. Unsurprisingly, it seems to be much easier to train networks on higher quality data. Still, it seems extremely hard. So I don't see mass data approach reaching the minimum threshold anytime soon. In fact the rich data companies will be in a better position to scale down their sensors once it starts working. They will have the ultimate baseline to compare these scaled down approaches against.
Elon said Level 5 autonomy was a "solved problem" in 2015, so his projections are totally meaningless. He's also been on record saying "next year" since 2016. As a note, Tesla is using HD maps. They may not have all of the detail that Waymo collects, but certainly they're much higher detail than their non-FSD beta maps. They've also said they're running their own maps collection program which seems to be similar to MobilEye's program as run on VW vehicles globally. Considering the industry history, given projects like NavLab, I see little evidence that there's going to be a large enough breakthrough to offer Level 4 on a majority of roads in a majority of the world. I'm highly skeptical we'll even have Level 4 on major highways in a majority of the world. Being a user of the FSD Beta, there's absolutely no merit to Elon's claims whatsoever.
Yeah, Elon has been wrong many times now so it gets harder and harder to believe his projections. However, he is a master at inspiring his own employees, investors and at attracting talent so now Tesla has an incredibly good team working very hard on this problem. Interesting with the HD maps, where did you hear about that? Would like to research that a bit more especially if it is also covering things such as lanes and road boundaries.
@@lightinspace2963 There's a limit to what inspiration can do, in the face of real technical challenges. "Vision" based self driving has been around since the 1908s with CMU's NavLab, so clearly the challenge isn't getting solved simply by a determined team of engineers. There are technical challenges to solve. As for HD maps, Tesla announced they were collecting them starting in parking lots, and continuing from there. People like Green The Only have pulled apart some of the map data, and we know that FSD specific maps are required for city streets autopilot.
@@TomCameron Tesla's "fsd" maps doesn't map the traffic lights to lanes and other stuff like that. And the drivable path is not on the centimeter level as a typical hd-map would be.
@@StefanNorberg These are all distinctions that don't make much difference, honestly. The fact of the matter is, Tesla is using an HD maps product from TomTom in their data tiles.
I don't see the point in comparing Teslas system to Waymo/Cruise/Mobileye. Waymo works today. You can get into a driverless Waymo taxi. Tesla FSD doesn't work without a lot of driver interventions. We can't compare them until Teslas system manages to drive around without a human driver.
Sure, but you can compare the technologies I think. I think no one has won this race until they deploy a system that works profitable in 100+ cities including connecting highways and larger roads.
Waymo is operating in the small area of Chandler. If that was all Tesla had tasked their system to do, then I'm sure they would be able to operate there too. They are trying to solve the much bigger problem of general ai. Which if they solve is a far more scalable solution.
@@daydreamer8373 Correct. LIDAR is NOT redundant for vision, as LIDAR cannot see. LIDAR is a "hardware diverse" method to simply detect the distance to a point in space. Hardware diversity and redundancy are completely different things. Only cameras will well-trained neural nets can give you "vision". I think of LIDAR like a blind person with a walking cane. It is a hardware diverse way of detecting the distance to an object if vision is unable to perform that function.
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
Excellent video, would be great if there’s an updated version for 2024/25. 👍
Thanks for watching! Let me know if you want more videos on this theme - it takes time to prepare but I also find the topic really interesting.
Great video! I have subscribed hoping for more content like this in the future. It would be awesome to have progress updates 2-4 times/year. Thank you for this video!
More, please :) Great work!!
Would love more, this was great
My friend, in your opinion, which system is the most future-proof? Thank you for your assessment. Many greetings from Germany.
Awesome video. Thanks a lot for posting this and all the comments below are helpful too.
This is a great video. Good production values, interesting points of learning, and interesting subjects. ML is such a breakthrough in computing but also something that is almost foreign for older IT professionals. When I took a degree in computer science in the mid-to-late 1990is this didn't exist. I was thought religiously that the "out data" can never be better than the "in data". In layman terms: you can't take a digital photo and upscale it with a better resolution because you are limited to the indata. But with ML you can... With self-driving cars. How will the government exploit the camera data to map our lives or is a strong government a key to get safe self-driving? In China, they are making everything networked and data-linked to self-driving cars so the AI knows every school crossing, signal, and so on by data signals and not just a camera with AI. Well.. sorry for the old man rant at the end.
Great video! Just a quick note: LIDAR is NOT redundant for vision, as LIDAR cannot see. LIDAR is a "hardware diverse" method to simply detect the distance to a point in space. Hardware diversity and redundancy are completely different things. Only cameras will well-trained neural nets can give you "vision". I think of LIDAR like a blind person with a walking cane. It is a hardware diverse way of detecting the distance to an object if vision is unable to perform that function.
Thanks! I agree that it is not redundant for vision. However you do get a 3D map which should be good for driving with if you also have some other things like signs in HD maps. Traffic lights will be hard though... However, LIDAR actually does see a bit. Because in addition to measuring the "time of flight" which gets converted to a distance, you also get a value for the strength of the returning beam. This strength indicates how reflective the object is. So this can look a bit like a gray scale image for some objects.
Great content. I rarely subscribe to RUclips channels and I subscribed to yours. I’m really interested in learning about how viable Tesla’s approach is and if you think there are any fundamental flaws. I really appreciate also understanding more about MobilEye and Waymo.
Excellent video, thank you. You should have more subs. Thanks for the no nonesense informative approach.
Great video. Tesla may not be just uploading edge cases these days. We are in the fsd beta program and when we return home, it regularly uploads > 20 GB of data even when the drive has no disengagements (which are happening more frequently). It's quite possible entire videos of drives are being uploaded for automatic annotation and even possible mapping purposes (there are technical papers that show cameras can create similar datasets to lidar to give 3d maps of its environment). I suspect Tesla maybe quietly moving to replicate what Mobileye is proposing by using data collected from each car to develop high resolution maps. If each Tesla out of the 100K enrolled in the FSD beta program is uploading 20 GB each drive, it may provide a more rich dataset than what Mobileye is proposing. The dataset would be even more enhanced if the remaining Tesla fleet is also involved in data collection beyond just edge cases. It would be interesting to hear from other Tesla owners that are not presently involved with the FSD beta program are also experiencing increased data uploads after drives.
I agree that once the differences in driving experience declines between the mass and rich data sets, it will be hard to justify continuing to rich data sets.
Thanks, very interesting comment! I also have this suspicion, somehow without "real intelligence", which humans have and computers do not, I think HD maps with at least lane lines and road edges might be what is needed to get enough robustness. Because the roads of the world are so wildly different. Humans use reasoning and logic to solve problems as they come, but cars might need HD maps...
I think there might be diminishing returns of the HD map beyond lanes, road edges and voxels (mine-craft ich blocks of the surroundings). However as more compute is available offline maybe it is still better that the cars upload the raw image data and not the output.
If Tesla gets HD maps, they will have "redundancy" for the world structure. Then maybe the question is if they can make vision only robust enough for cars/pedestrians/animals etc.
Eventually maybe the companies will meet in the middle - Tesla adding a few cameras and increasing the camera resolution, Waymo perhaps removing some LIDARs. Question is who gets there faster. :) Will be interesting to follow.
Which tech has the highest level of ASIL's for Safety?
Driving assistance systems and autonomous system are typically classified using SAE J3016. A few cars has autonomy in a limited operational domain. The Mercedes EQS and the S-class has a level 3 system a so called "eyes off" system (you can read a book or play a game on your phone) in stop and go traffic in speeds up to 60km/h. Honda has a similar system in Japan. All other systems are not autonomous, levels 0-2. In a Level 4 car you do not need to take over control ever, within the operational domain. It's like riding in a taxi. You are not driving. Such system are not legal anywhere in the world (for private operation), and you will not be able to buy such a car before 2030 likely,.
The real promise of self driving goes much further than safety. Subscribing to a robotaxi service promises to be way cheaper than owning your own vehicle. With such a fleet vehicles are utilized far more than private vehicles. They can also be sized to match the average occupancy rather than buying a large car for the rare occasion you might need it.
Result is far fewer vehicles produced most of which are far smaller and more efficient.
The real winner from this will be the environment, through quieter safer streets and vastly reduced emissions.
Good point, I agree - it will really reduce the cost for mobility. However, when it is very cheap to take a robotaxi, is there maybe a risk that the streets will not become quieter but instead more busy? I am thinking that people rather drive in their own private robotaxi than taking the subway etc.
@@lightinspace2963 I think it's definitely possible people will stop taking buses and trains, after all the will get door to door service. This is not as bad as it seems though because the load factor is so poor on buses especially that it's quite easy for a lightweight 2 person vehicle to emit less CO2 per passenger than a bus.
The efficiency of the subway is great though, and yes moving people off this and onto the streets would also be bad from a traffic POV.
Can ML save the Wasa so we can take Moscow soon?
What few discuss with regard to the autonomous ride hailing approaches taken by Waymo, Mobileye, and Tesla is the economics of each. Nothing succeeds if it is not economically viable. For that reason, I think Waymo is driving down a blind alley. It's an over-engineered, expensive system that will never be profitable. Tesla's system is the most likely to succeed from an economics standpoint if the engineering issues can be solved, and it appears they will be. Mobileye will find a place, as well, but on a more limited scale.
Mass data/rich data is not an accurate characterization of the two approaches. The data contained in Waymo's high definition maps is certainly voluminous, but unnecessarily so because vehicle localization can be achieved much more efficiently using GPS and vision. Tesla's occupancy networks serve the purpose of 3-D lidar maps for most purposes with more than adequate resolution.
I like Mobileye's use of low resolution dynamic maps as a form of shared memory. I think Tesla would do well to implement this in their FSD system to deal with confusing lane structures, parking lots, driveways, speed bumps and potholes, road closures, etc.
Thanks for the video.
Seems to me like the rich data approach is winning. The reliability has proven to be very difficult to improve. Every subsequent 9 after 99% seems to be exponentially more difficult to get. And there are a lot of 9 required to build public confidence. Unsurprisingly, it seems to be much easier to train networks on higher quality data. Still, it seems extremely hard. So I don't see mass data approach reaching the minimum threshold anytime soon. In fact the rich data companies will be in a better position to scale down their sensors once it starts working. They will have the ultimate baseline to compare these scaled down approaches against.
Anyone can put lidar on their cars, not everyone can do Tesla approach
Elon said Level 5 autonomy was a "solved problem" in 2015, so his projections are totally meaningless. He's also been on record saying "next year" since 2016. As a note, Tesla is using HD maps. They may not have all of the detail that Waymo collects, but certainly they're much higher detail than their non-FSD beta maps. They've also said they're running their own maps collection program which seems to be similar to MobilEye's program as run on VW vehicles globally.
Considering the industry history, given projects like NavLab, I see little evidence that there's going to be a large enough breakthrough to offer Level 4 on a majority of roads in a majority of the world. I'm highly skeptical we'll even have Level 4 on major highways in a majority of the world. Being a user of the FSD Beta, there's absolutely no merit to Elon's claims whatsoever.
Yeah, Elon has been wrong many times now so it gets harder and harder to believe his projections. However, he is a master at inspiring his own employees, investors and at attracting talent so now Tesla has an incredibly good team working very hard on this problem.
Interesting with the HD maps, where did you hear about that? Would like to research that a bit more especially if it is also covering things such as lanes and road boundaries.
@@lightinspace2963 There's a limit to what inspiration can do, in the face of real technical challenges. "Vision" based self driving has been around since the 1908s with CMU's NavLab, so clearly the challenge isn't getting solved simply by a determined team of engineers. There are technical challenges to solve.
As for HD maps, Tesla announced they were collecting them starting in parking lots, and continuing from there. People like Green The Only have pulled apart some of the map data, and we know that FSD specific maps are required for city streets autopilot.
@@TomCameron Tesla's "fsd" maps doesn't map the traffic lights to lanes and other stuff like that. And the drivable path is not on the centimeter level as a typical hd-map would be.
@@StefanNorberg These are all distinctions that don't make much difference, honestly. The fact of the matter is, Tesla is using an HD maps product from TomTom in their data tiles.
@@TomCameron Yeah, I agree. The computer vision space is still very much a research area, and it's probably a decade away at the very least.
I don't see the point in comparing Teslas system to Waymo/Cruise/Mobileye.
Waymo works today. You can get into a driverless Waymo taxi.
Tesla FSD doesn't work without a lot of driver interventions.
We can't compare them until Teslas system manages to drive around without a human driver.
Sure, but you can compare the technologies I think. I think no one has won this race until they deploy a system that works profitable in 100+ cities including connecting highways and larger roads.
Waymo is operating in the small area of Chandler. If that was all Tesla had tasked their system to do, then I'm sure they would be able to operate there too. They are trying to solve the much bigger problem of general ai. Which if they solve is a far more scalable solution.
Yeah because who cares about the future?
@@mfpears Are you trying to make a point. If so what is it?
Tesla dropping Lidar for cameras is not a good thing.
I actually agree with Elon. Lidar is a crutch, and is actually holding back self driving development.
Read "The Bitter Lesson"
@@daydreamer8373 Correct. LIDAR is NOT redundant for vision, as LIDAR cannot see. LIDAR is a "hardware diverse" method to simply detect the distance to a point in space. Hardware diversity and redundancy are completely different things. Only cameras will well-trained neural nets can give you "vision". I think of LIDAR like a blind person with a walking cane. It is a hardware diverse way of detecting the distance to an object if vision is unable to perform that function.