There is a good reason why Elon is starting the robots in factories where the robots have little or no contact with humans, safety! Every FSD accident is greatly amplified by scare mongers. Starting Optimus in factory settings reduces this problem. The other issue is production ramp-up. The first jobs need to be replacing high value added humans. Universal adoption is not going to be overnight.
The Captain
The factory business is relatively easy. The tasks are repetitive, or semi-repetitive, or serially repetitive and generally well defined. A bot for the home doing laundry, caring for kids, washing the dog, or in ânursing homes, hospitals, restaurants and other locations where the capacity to do multiple tasks in a human environmentâ is a looooong way away.
What does this have to do with what I wrote above? Iâll repeat it here -
⌠that have âstandardâ cars do some of the driving - like stopping at stop signs and stoplights, like turning left and right, adjusting to speed limits, and pulling into a parking spot upon arrival.
I read the article and watched the video, and that Mercedes feature does none of these things. What is does do, and does well, is handle stop and go traffic, in one lane, on a highway, up to 40 mi/hr. If the traffic flow speeds up to 55, then down to 20 ⌠15 ⌠20 ⌠50 ⌠20 ⌠etc, it canât adjust to them and will either remain at 40 or will ask the driver to take over. It wonât change lanes when necessary, it wonât exit the highway when you reach your exit, etc. I rented a bunch of different cars a few years ago (more than 5 years ago) and tried out their cruise control with lane and distance keeping feature, and they all worked reasonably well. As I recall, Volvo (an S90 I think) performed the best at the time. I wouldnât be surprised to see this Mercedes system do an even better job, though I donât understand why it wonât do it at 65 or 70 mi/hr as well. That would make it far more useful.
I donât know how long âlooooongâ is going to be but most likely shorter than pessimists think it will be. But that is not all that relevant. We need to solve two adoption issues which I believe are safety and affordability. Jobs in location with little risk to humans and jobs costly enough me make the robots economically feasible.
I repeat:
One thought to keep in mind, Tony Seba has documented that technology adoption âSâ curves are getting steeper all the time. Futurologist Ray Kurzweil wrote in The Age of Spiritual Machines that time runs faster as systems get more organized (or words to that effect). The future is now closer than ever before.
The Captain
I donât understand this statement. Can you explain what it means? Iâm assuming that hardware costs per mile means the cost of the vehicle (the âhardwareâ) divided by the useful miles the vehicle can provide. So 30 cents for Waymo could be letâs say $90,000 for the vehicle divided by 300,000 miles = 30 cents. But how did someone calculate the uberlyft number? Letâs say the car costs $40,000 and can go only 150,000 miles (really a typical Toyota, Honda, or Tesla could easily go 200,000 miles), that comes to 27 cents per mile. And even so, you would have to explain why the Waymo car (a Jaguar) would get twice as many miles as a Toyota, Honda, or Tesla would get.
Minor detail: There is no Robotaxi. There are only plans for a Robotaxi.
Iâm not sure geofencing is a huge liability. You donât have to replace every human driver. You can carve out the easiest, most lucrative areas first, and leave the rest to Uber until later. In the meantime, there are self-driving taxis accepting passengers from the general public right now. That seems to be a bit of a head start over the Robotaxi, which again, does not exist.
And Iâm not sure Tesla will be the first to crack the nut on autopilot. Lately, other manufacturers are consistently getting higher ratings and Teslaâs FSD has fallen to sort of mid-range. There is even an open-source autopilot that gets good reviews.
Article from yesterday on this topic. Waymo is testing a purpose-built robotaxi right now:
Not just plans, most of the pieces of the Robotaxi also exist. itâs a question of putting them together in the right order.
The Captain
I canât speak to the video but it does do everything you mentioned in your quote.
From the article:
The Mercedes-Benz EQS sedan expertly steered, braked and accelerated on Southern Californiaâs Interstate 10 as I flipped through the pages of the Los Angeles Times, oblivious to the traffic encircling us.
I had deliberately relinquished all driving duties to the carâs computer. Instead of monitoring the road, I was free to turn my attention to âother activities inside the vehicle,â my passenger, Mercedes-Benz engineer Lucas Bolster, reminded me.
The key here is that unlike L2, a driver is not required to pay attention. Again, there is no other personal vehicle certified as L3 currently on the road.
Perhaps this fills in the blanks:
Oops, missed this part:
It probably CAN, but at a much higher level of liability. Or, perhaps it canât because the reaction times have to be twice as fast at 70 than they are at 35 (currently the driver gets 10 seconds to take over - that might be cut to five at higher speeds).
It wouldnât be for robotaxi, or indeed most of the TaaS market. If you geofence to a single metro, your vehicle will still be able to service the overwhelming majority of trips taken in that metro.
Plus, in these early years of robotaxis (if weâre there yet), itâs pretty unlikely that these services will be completely autonomous. Theyâll still occasionally find themselves in circumstances where human intervention is required - either a remote takeover, or a human physically going to wherever they are to help them out (a door thatâs stuck ajar, a traffic situation that they canât figure a way out of, etc.). And theyâll need regular cleaning. To service those cars, youâll need personnel in the area - which is itself a de facto geofence.
Unlimited Level 5 is the sell for people who own their own vehicles. Robotaxis and TaaS donât need more than Level 4.
I mentioned 4 things (but there are scores more):
Which of these items do you think it does? Before you answer, seek more information about what exactly the âDrive Pilotâ feature in the 2024 US Mercedes EQS comprises. I suspect some misconception is at play here.
With Uber or Lyft you have the hardware cost, but you also have to pay a driver. That makes your cost per mile significantly higher.
Okay. So back to the original discussion. If you were to compare a Tesla driverless solution (letâs say a model Y costing $45,000+$8000 for the self-driving software) to a Waymo driverless solution (letâs say a Jaguar i-Pace costing $80,000+$60,000 extra equipment+$? for the software), which one do you think will have the lower cost per mile?
Today, it would be the Waymo solution. Because Tesla doesnât have a driverless solution. FSD is only Level 2, while Waymo is Level 4. So the Model Y would need a driver. Which would make it much more expensive.
Baked into your hypothetical discussion is the assumption that Tesla solves Level 5 autonomy, and that it can do that without having to add any additional equipment. They have not yet done so.
I THINK it does all of those. L3 has limitations but it seems their L2 setting does all of the above. Admittedly, the switch from L3 to L2 is not automatic and it requires the driver to make that switch but the best I can tell, it does all of the above under one setting or another.
Edit: Here is vid on Park Assist.
Not according to MB. Their tutorial video claims that they only work in traffic jams, under 40 mph on certain pre-mapped freeways. So Iâm not sure how it would handle stop signs and turns. The FAQ says it does not change lanes.
This page has tutorial videos a map of allowed roads and FAQ, etc.
(Maybe it is different outside the US)?
Already said upthread: âliabilityâ. But I have another, although admittedly itâs a total guess. Based on mine which has stay-in-lane and cruise control with 3 levels of choice on how close to follow, I find that even with the most distant, anytime someone moves into the open space the car âbrakesâ (not really, itâs regen; but same loss of momentum) which makes for an uncomfortable push-pull driving experience. At even higher speeds the software will want even more space for reaction time and braking, and as a practical matter itâs near impossible to keep that much space between you and the car ahead.
I donât know how we ever overcome that until everything is autonomous, but thatâs for a far distant generation in a far distant galaxy.
The discussion was about the future ⌠as we were discussing the statement by the Waymo CEO about reaching the 30 cent per mile cost sometime in the future.
It canât!!! None of the popular Teslaâs on the road today will ever be fully autonomous (except perhaps the Cybertruck). Thatâs because they canât see what is directly in front of them ⌠so the car will drive right over the bike carelessly left in the driveway on the ground in front of the car.
Only after a bumper camera is installed on some future model will full autonomy even be feasible in the first place. But there are still plenty of other things that will stymie full autonomy for quite some time. The cars will need to be able to read signs and follow directions first. If you want another example, hereâs one. Many housing developments have two lanes for entrance, one for cars with a sticker permitting entrance and a second where a gate guard checks ID and calls the home you are visiting before allowing the vehicle in. Any fully autonomous vehicle will need to know if it has a sticker for that particular place, and it will need to know which lane to choose, and it will need to know how to tell the guard which family or address it is headed to. Tesla, with the most advanced general purpose driving system, still canât do any of this.
Yes! A year ago, Tesla had the same problem. Theyâve greatly improved it in recent versions, and it is much less jarring. It almost drives like a human now, but not in ever scenario. There are still plenty of scenarios where it doesnât drive particularly well (Iâve mentioned them in the many threads weâve had about the topic) and I report every single one of them to Tesla. This evening, for example, I was using FSD on the highway, and I was 1 mile from the exit, driving in the second to left lane. I was wondering why it hadnât started moving over to the right in preparation to exit, so I blinked ârightâ to tell it to do so (even though it should have done it alone). There were cars in the lane to my right (just a few, and it would clear in seconds), so it decided to ignore my blink to the right AND IT SIGNALED LEFT instead and prepared to move into the leftmost lane! That is ridiculous when less than 1 mile to your exit. I left a quick bug report to Tesla, and I think I even shouted at it a bit. At that point I took over, exited properly, and turned FSD back on. I have not received the latest version (12.5) because I drive a model 3 and so far they are only releasing the latest version to model Y vehicles. The newer version is reportedly quite a bit better at many driving tasks. We shall see.
My Model Y gives me all sorts of warnings if I get within a foot or so of a six inch curb from the front or back.
Gotcha. If weâre talking about the future, though, then you wouldnât use those present-day costs for the car or the equipment.
Waymo is counting on being able to drive down the cost of the sensor array. Teslaâs approach uses low-cost sensors that need a super-âsmartâ AI driver; Waymo uses higher-cost sensors (LIDAR) so that they use a âdumberâ AI driver. Itâs been a while, but my dim recollection is that their expectation was that they could get their sensor array below $8-10K once they optimized it and entered larger-scale production.
Theyâre probably not going to use Jaguar iPace as their vehicle in the longer term, either - at some point theyâll use something a little more purpose-fit for a robotaxi, not a midrange luxury car.
Can they do it? We donât know. It might be that neither approach will work any time soon. We might simply be in a world where near-term tech just canât make an AI smart enough for vision-based autonomy, and where LIDAR-based sensors remain too expensive to get the cost of robotaxis significantly below that of human drivers.
Waymo is testing a purpose built vehicle built by Zeekr in China that lacks a steering wheel and other driver controls.
