I think you are letting Elon’s incompetence in how our government works color your thinking. Just because he is so clueless on that front doesn’t mean his companies are also clueless. He also seems to have very top-notch and highly-motivated people working for him.
With regard to promising FSD within two years for the last ten years and w…
I think you are letting Elon’s incompetence in how our government works color your thinking. Just because he is so clueless on that front doesn’t mean his companies are also clueless. He also seems to have very top-notch and highly-motivated people working for him.
With regard to promising FSD within two years for the last ten years and within a year for the last six, the thing to do is listen to the other people at Tesla working on FSD. Elon is always grossly optimistic on really hard things that others doesn’t think are possible (FSD, landing and reusing rocket boosters, brain interface chips, tunnel-boring equipment in miles per day instead of days per mile, to name a few). Part of that is setting very-stretch goals, but he actually believes there’s a chance he could meet those. He acknowledged that trait is saying that his companies turn “impossible” into “late.” The problem with saying “he’s been wrong about that so many times” is that they do seem to eventually get those to work.
The cars at launch will be “Robotaxis”: Model Y’s with no driver. What they will not be is “Cybercabs”: the new two-seater cars with no steering wheel still in development that had prototypes shown and driven at the Robotaxi event 10/10/24. Eventually, Cybercabs will dominate any successful Robotaxi network because their amortized cost per mile will be less than the about $0.15 of the Model Y. Since about 90% of rideshares are just one or two people, Cybercabs can handle those. Even once Cybercab are in full production, there will still be the need for larger cars for when there are more than two people.
Waymo and Tesla took very different approaches to the problem. Waymo went for LIDAR to get precision mapping of the environment around it combined with very detailed maps of where it would be operating. Tesla went for radar and cameras initially, but the radar angular resolution was too poor to be helpful. Radar cannot tell the difference between a car stopped in the road and a car stopped beside the road. It can tell that there is something large, that it is stopped, and exactly how far away it is, but cannot tell if it matters or not. The radar was dropped not to save cost, but because it gave so many false alarms that it had to practically be ignored.
The advantage for Tesla is the cost of a couple of thousand dollars for the inference computer along with all the cameras. That is much cheaper than the additional $100,000 that Waymo spends. The result is Tesla decided to put it into every Tesla built. Not having it as an option cuts manufacturing changes from car-to-car and brings the unit cost of FSD hardware down due to the much higher volume. It also means that Tesla has over two million cars on the road that customers could eventually “hire out” to the Robotaxi fleet to earn money when they are not using it. Initially, Tesla will only be using company cars. They don’t need the large number until it proves successful and needs to scale up. Their plan is to start with 10-20 and geofenced to a area of Austin with only easy turns. They expect to be up to several hundred cars within a few months. Waymo has 700 cars nationally and has an order for 300 more next year. Tesla currently produces about 400,000 cars a quarter (all of which are Robotaxi ready), but that will go up if they start producing Cybercabs in volume
Waymo is already operating driverless taxi fleets and have been for a number of years. They started with by-invitation and for their own employees for a while before offering paid rides to the general public. Tesla has been doing the same thing in California for a while for their own employees, but with a driver there that could take over if necessary. All driverless fleets need to have the ability of a remote teleoperator that can direct the car if necessary. The teleoperator is not monitoring the car, but rather is there to do something when the car reports it doesn’t know what to do. In some cases, this could just be giving a direction to the car (turn around or back up, go around some obstruction, take a different route) or in others, it might be to remotely drive the vehicle. The question is how often that is necessary. With Cruise, after they went under, it came out that their teleoperators were doing something an average of every two miles or so. That doesn’t scale. If you have one teleoperator for every 100 cars, that could work with a small fleet, but to scale to a large fleet, interventions would need to be low enough that there are a lot fewer than one remote driver per hundred cars.
Regarding LIDAR vs cameras, generally, the precision of LIDAR (knowing where things are to a tenth of an inch) isn’t critical to being able to drivie. 4K video is sufficient to know where to go and what to avoid. The really big benefit is having eight cameras instead of just seeing one direction at a time. I’m a little surprised that there aren’t cameras looking sideways at the four corners of the car. That would increase visibility even more when needing to stick ones nose out to see if it is clear to go. The cameras also go farther into the infrared (I can see a lot more at night in my backup camera than my mirror). Tesla does have some test cars fitted with LIDAR for precision truth data for validating the 3D model the software generates from the cameras.
I think you are letting Elon’s incompetence in how our government works color your thinking. Just because he is so clueless on that front doesn’t mean his companies are also clueless. He also seems to have very top-notch and highly-motivated people working for him.
With regard to promising FSD within two years for the last ten years and within a year for the last six, the thing to do is listen to the other people at Tesla working on FSD. Elon is always grossly optimistic on really hard things that others doesn’t think are possible (FSD, landing and reusing rocket boosters, brain interface chips, tunnel-boring equipment in miles per day instead of days per mile, to name a few). Part of that is setting very-stretch goals, but he actually believes there’s a chance he could meet those. He acknowledged that trait is saying that his companies turn “impossible” into “late.” The problem with saying “he’s been wrong about that so many times” is that they do seem to eventually get those to work.
The cars at launch will be “Robotaxis”: Model Y’s with no driver. What they will not be is “Cybercabs”: the new two-seater cars with no steering wheel still in development that had prototypes shown and driven at the Robotaxi event 10/10/24. Eventually, Cybercabs will dominate any successful Robotaxi network because their amortized cost per mile will be less than the about $0.15 of the Model Y. Since about 90% of rideshares are just one or two people, Cybercabs can handle those. Even once Cybercab are in full production, there will still be the need for larger cars for when there are more than two people.
Waymo and Tesla took very different approaches to the problem. Waymo went for LIDAR to get precision mapping of the environment around it combined with very detailed maps of where it would be operating. Tesla went for radar and cameras initially, but the radar angular resolution was too poor to be helpful. Radar cannot tell the difference between a car stopped in the road and a car stopped beside the road. It can tell that there is something large, that it is stopped, and exactly how far away it is, but cannot tell if it matters or not. The radar was dropped not to save cost, but because it gave so many false alarms that it had to practically be ignored.
The advantage for Tesla is the cost of a couple of thousand dollars for the inference computer along with all the cameras. That is much cheaper than the additional $100,000 that Waymo spends. The result is Tesla decided to put it into every Tesla built. Not having it as an option cuts manufacturing changes from car-to-car and brings the unit cost of FSD hardware down due to the much higher volume. It also means that Tesla has over two million cars on the road that customers could eventually “hire out” to the Robotaxi fleet to earn money when they are not using it. Initially, Tesla will only be using company cars. They don’t need the large number until it proves successful and needs to scale up. Their plan is to start with 10-20 and geofenced to a area of Austin with only easy turns. They expect to be up to several hundred cars within a few months. Waymo has 700 cars nationally and has an order for 300 more next year. Tesla currently produces about 400,000 cars a quarter (all of which are Robotaxi ready), but that will go up if they start producing Cybercabs in volume
Waymo is already operating driverless taxi fleets and have been for a number of years. They started with by-invitation and for their own employees for a while before offering paid rides to the general public. Tesla has been doing the same thing in California for a while for their own employees, but with a driver there that could take over if necessary. All driverless fleets need to have the ability of a remote teleoperator that can direct the car if necessary. The teleoperator is not monitoring the car, but rather is there to do something when the car reports it doesn’t know what to do. In some cases, this could just be giving a direction to the car (turn around or back up, go around some obstruction, take a different route) or in others, it might be to remotely drive the vehicle. The question is how often that is necessary. With Cruise, after they went under, it came out that their teleoperators were doing something an average of every two miles or so. That doesn’t scale. If you have one teleoperator for every 100 cars, that could work with a small fleet, but to scale to a large fleet, interventions would need to be low enough that there are a lot fewer than one remote driver per hundred cars.
Regarding LIDAR vs cameras, generally, the precision of LIDAR (knowing where things are to a tenth of an inch) isn’t critical to being able to drivie. 4K video is sufficient to know where to go and what to avoid. The really big benefit is having eight cameras instead of just seeing one direction at a time. I’m a little surprised that there aren’t cameras looking sideways at the four corners of the car. That would increase visibility even more when needing to stick ones nose out to see if it is clear to go. The cameras also go farther into the infrared (I can see a lot more at night in my backup camera than my mirror). Tesla does have some test cars fitted with LIDAR for precision truth data for validating the 3D model the software generates from the cameras.