Self Driving Car thread
- Thread starter ChaoticUnreal
- Start date
the discussion online about these numbers is still ongoing, here are some more numbers: "by end 2022, US MY miles driven >7B, M3 ~19B"
So at ~5 deaths per 1B mile driven, that would be like 35 deaths total in a model Y before end of 2022? Seems low. Though from this tracker, only ~300 deaths in or by a Tesla tracked by end of 2022: https://www.tesladeaths.com/, and the Model Y is only a fraction of that.
So at ~5 deaths per 1B mile driven, that would be like 35 deaths total in a model Y before end of 2022? Seems low. Though from this tracker, only ~300 deaths in or by a Tesla tracked by end of 2022: https://www.tesladeaths.com/, and the Model Y is only a fraction of that.
Tesla has released FSD v13.2 to the usual early cheerleaders.
One of the headlines is the ability to move directly from driving mode into parking mode and complete the action. This seems to be somewhat unreliable at the moment with numerous early videos showing the car often not recognising spaces, or parking incorrectly (e.g. one on a hashed area). I suspect this will be improved as manually moving the car to parking mode seems to find the spaces fairly well.
One of the headlines is the ability to move directly from driving mode into parking mode and complete the action. This seems to be somewhat unreliable at the moment with numerous early videos showing the car often not recognising spaces, or parking incorrectly (e.g. one on a hashed area). I suspect this will be improved as manually moving the car to parking mode seems to find the spaces fairly well.
I suspect there will be a full Ars story about this soon
https://www.bbc.co.uk/news/articles/cj902y4ez71o
https://www.bbc.co.uk/news/articles/cj902y4ez71o
Not unexpected, but I'm super interested to hear from them or people in the team what caused the collapse of their self-driving branch. From a performance point of view they were certainly not trailing, they were leading along with Argo.ai and Waymo. It's clear that none are ready for prime-time, but there is still a surprising amount of optimism in the industry.
I'm not seeing much optimism, esp. w.r.t. the cost of needed HW sensors to make sure there are no single points of failure, and how long it'll take to get to what usedto be called SAE L5 (no driver in vehicle, no user controls necessary).
(personally, I think SDCs without central traffic control much of the time don't make sense)
Sounds like the criminal case effectively ended the project last year, they laid off a lot of the team and now they're figuring out what to do with what's left of the company.
I see it as the inevitable result of the economic situation being less optimistic. Capital gets more focused on things that will likely be able to sell even if the economy doesn't expand a lot. There's also competition in the robotaxi space and I don't think anybody's making money at it.
Automation that can help you drive but doesn't have to be perfect and function without a backup is a lot easier to pull off, costs less, and doesn't expose you to as many risks. How mad is a Tesla self-driving owner when their car refuses to do something because of a situation it doesn't have the programming for or a broken sensor and they have to operate it directly? Not that mad is Tesla's experience, right? But a taxi that won't go and can't easily be operated from manual controls is going to have to be towed.
Automation that can help you drive but doesn't have to be perfect and function without a backup is a lot easier to pull off, costs less, and doesn't expose you to as many risks. How mad is a Tesla self-driving owner when their car refuses to do something because of a situation it doesn't have the programming for or a broken sensor and they have to operate it directly? Not that mad is Tesla's experience, right? But a taxi that won't go and can't easily be operated from manual controls is going to have to be towed.
Forgive me for being cynical, but if they really think they'll pull 50 billion in revenue out of their robotaxis, they're probably going to find a way to overlook that as a rectifiable error rather than a showstopper. Especially because all their competitors also have blood on their hands, some quite a bit more.
IMO, but I'm an outlier pessimist, it has to do with some form of realism seeping into these companies, finally not just figuring out but also convincing management that self-driving cars aren't close and won't happen anytime soon, so it's better to cut their losses. But since I've been wrong about this attitude before, I'd like to hear from them rather than speculate based on my own beliefs.
I don't know about realism or tempered expectations.
You don't see Musk backing off his ridiculous promises, because TSLA valuation depends a lot on SDC and the general AI hype.
I read that Uber, which seemed to give up on SDC after the accident in Phoenix, is now testing in the UAE.
GM has usually been penny-wise, pound foolish, not really looking long term.
It just may be that self-driving progresses more in places like China, because the government there will push beyond safety concerns -- that's probably why Uber is in UAE.
With Trump in office, Musk may bypass states like CA and get more lasses-faire treatment from the feds.
You don't see Musk backing off his ridiculous promises, because TSLA valuation depends a lot on SDC and the general AI hype.
I read that Uber, which seemed to give up on SDC after the accident in Phoenix, is now testing in the UAE.
GM has usually been penny-wise, pound foolish, not really looking long term.
It just may be that self-driving progresses more in places like China, because the government there will push beyond safety concerns -- that's probably why Uber is in UAE.
With Trump in office, Musk may bypass states like CA and get more lasses-faire treatment from the feds.
Waymo has allowed Swiss Re to examine their data.
Waymo come out way safer than humans.
” The study uses auto liability claims aggregate statistics as a proxy for at-fault collisions and expands on our previous research. It demonstrates that as we've scaled operations across Phoenix, San Francisco, Los Angeles, and Austin, the Waymo Driver significantly outperforms both the overall driving population and the latest generation of human-driven vehicles equipped with advanced driver assistance systems (ADAS)……..Waymo Driver demonstrated better safety performance when compared to human-driven vehicles, with an 88% reduction in property damage claims and 92% reduction in bodily injury claims. In real numbers, across 25.3 million miles, the Waymo Driver was involved in just nine property damage claims and two bodily injury claims”
overview
https://waymo.com/blog/2024/12/new-swiss-re-study-waymo
lots of data in the preprint
https://waymo.com/research/do-auton...ation-human-driven-vehicles-25-million-miles/
Waymo come out way safer than humans.
” The study uses auto liability claims aggregate statistics as a proxy for at-fault collisions and expands on our previous research. It demonstrates that as we've scaled operations across Phoenix, San Francisco, Los Angeles, and Austin, the Waymo Driver significantly outperforms both the overall driving population and the latest generation of human-driven vehicles equipped with advanced driver assistance systems (ADAS)……..Waymo Driver demonstrated better safety performance when compared to human-driven vehicles, with an 88% reduction in property damage claims and 92% reduction in bodily injury claims. In real numbers, across 25.3 million miles, the Waymo Driver was involved in just nine property damage claims and two bodily injury claims”
overview
https://waymo.com/blog/2024/12/new-swiss-re-study-waymo
lots of data in the preprint
https://waymo.com/research/do-auton...ation-human-driven-vehicles-25-million-miles/
Full paper: https://storage.googleapis.com/waym...mo and Human-Driven Vehicles at 25M miles.pdf
What I can't find:
Waymo is currently limited to low speed and non highway roads. How did they segment the insurance data to get only the city driving miles out of it?
I guess it's still better than Tesla's comparison, highway miles vs mixed, as cities cause more accidents, but either way it's not a apples to apples comparison.
Do you lot have telematics/black box insurance over there? If so, then that could be a starting point...
https://blog.aurora.tech/safety/preparing-for-driverless-operations
Aurora just published a white paper on their safety process, in advance of plans to launch operator-free operations.
Aurora just published a white paper on their safety process, in advance of plans to launch operator-free operations.
Video of a ride on a Chinese self-driving system.
View: https://youtu.be/VuDSz06BT2g?si=0vriL2PMC1AgzsAp
This is a Huawei EV with 3 lidars. The car is around $30k. The lidars are about $200 each now.
The software isn't done but the YouTuber seems more impressed with it than anything he's used on a Tesla FSD.
View: https://youtu.be/VuDSz06BT2g?si=0vriL2PMC1AgzsAp
This is a Huawei EV with 3 lidars. The car is around $30k. The lidars are about $200 each now.
The software isn't done but the YouTuber seems more impressed with it than anything he's used on a Tesla FSD.
What? There's a safety process? I thought everybody was just YOLOing it.
Can't tell if serious.
Waymo, Cruise, and Aurora, at least, have all been quite methodical about their development and its implications. Uber, Tesla, et al have obviously not been. No idea about how the Chinese or other ones are approaching things.
Nevertheless, Waymo ran over and dragged a pedestrian. So I guess?
Yes, you are correct.
Waymo cars have had different problems.
https://www.theverge.com/2024/5/14/24156238/waymo-nhtsa-investigation-crash-wrong-side-road
Not a car exactly, but my wife's family owns a 250 acre farm around here and we're deep in the middle of adding an open source autosteer system to a tractor as a prototype and may do multiple later. A bunch of farmers in Europe got stick of paying Deere for the hardware and subscriptions and made their own, with blackjack and hookers. The project is AgOpenGPS, and it's the most poorly run hardware/software project I've seen in a long time and it's all glorious.
We got PCBs made via JLCPCB, added a RTK base station to one of the barns driven by a Pi4 and RTKBase. The GPS receivers are all uBlox F9P's and we expect about 1.5" of repeatable accuracy in the field. It's driven by a Teensy (Arduino knockoff) and a Cytron MD13S motor driver. We're bolting a steering motor to the wheel in the cab with some 3d printed gears but the PCB supports hydraulic valves to interface with power steering directly or CAN native (for tractors that are autosteer ready). It uses a land rover ride height sensor for wheel angle, they're resilient and about $50 brand new or about $20 for knockoffs. We estimate this is going to cost ~$1000 per tractor with no 3rd party dependancies or recurring costs (other than GPS being available)
We're pretty excited, should be a few weeks until we have it ready to drive something.
We got PCBs made via JLCPCB, added a RTK base station to one of the barns driven by a Pi4 and RTKBase. The GPS receivers are all uBlox F9P's and we expect about 1.5" of repeatable accuracy in the field. It's driven by a Teensy (Arduino knockoff) and a Cytron MD13S motor driver. We're bolting a steering motor to the wheel in the cab with some 3d printed gears but the PCB supports hydraulic valves to interface with power steering directly or CAN native (for tractors that are autosteer ready). It uses a land rover ride height sensor for wheel angle, they're resilient and about $50 brand new or about $20 for knockoffs. We estimate this is going to cost ~$1000 per tractor with no 3rd party dependancies or recurring costs (other than GPS being available)
We're pretty excited, should be a few weeks until we have it ready to drive something.
Are you doing dGPS with markers on the field? Because 1.5" repeatability is pretty fancy even for that. Didn't know you could just buy modules that did that!
Nothing in the fields. I'm no GPS expert but I do have an embedded systems background. It's my understanding that this generation of GPS receivers, specifically the uBlox stuff is like 90% as good as "Survey grade" gear from Trimble/etc at like 1/10th to 1/20th the cost. For a farm this is more than enough. Below 10km between the base and the rovers you're looking at about a 2-5cm minimum absolute error or basically 0.75" to 2"
RTK is siginifcantly more accurate than dGPS
For the RTK base we set up the same model receiver plugged into that Pi running RTKBase, you let it collect 1hz data for ~24 hours and then you send that data to one of several worldwide government PPP processing services, we used NRCan's PPP (Canada). They take all your data and basically give you back the exact location of the base (I think the accuracy on the position we got back was quoted at 0.02 meters for X/Y and 0.05M for Z and you program this into the base station.
Then you can either send this data to the rover directly with Tailscale or some overlay, or use a free service like RTK2Go (which we did), that allows anyone else in the area to benefit from it should they want.
Inside the tractor there's a windows tablet running the AgOpen software that talks to the PCB and part of that is a NTRIP client that gets the correction data from the base station, that sends the correction data down into the receivers, which do their own magic and pop out with a standard GPS message the software can consume. Oh, right, these receivers are also 10hz capable at full performance and 50hz capable at lower accuracy. They're pretty nuts.
My understanding is that because the base knows exactly where it is, any atmospheric interference will affect receivers the same in a given region. So the base says "Hey you're getting this data from these GPS sats (it can only correct for data from shared visible birds) but you actually need to adjust what you think about the data from those by XXX because I know where I am and I can see the interference"
The rover (technical name for the moving thing) has two receivers that it uses for extremely accurate roll/heading calculations. Basically the right GPS receiver does position and the left receiver is used to calculate direction/roll. You can use 1 receiver and an IMU instead but it's not as good to save $200 bux on the F9P.
tl;dr we should get roughly 1-2" of repeatable year over year row over row accuracy.
RTK is siginifcantly more accurate than dGPS
For the RTK base we set up the same model receiver plugged into that Pi running RTKBase, you let it collect 1hz data for ~24 hours and then you send that data to one of several worldwide government PPP processing services, we used NRCan's PPP (Canada). They take all your data and basically give you back the exact location of the base (I think the accuracy on the position we got back was quoted at 0.02 meters for X/Y and 0.05M for Z and you program this into the base station.
Then you can either send this data to the rover directly with Tailscale or some overlay, or use a free service like RTK2Go (which we did), that allows anyone else in the area to benefit from it should they want.
Inside the tractor there's a windows tablet running the AgOpen software that talks to the PCB and part of that is a NTRIP client that gets the correction data from the base station, that sends the correction data down into the receivers, which do their own magic and pop out with a standard GPS message the software can consume. Oh, right, these receivers are also 10hz capable at full performance and 50hz capable at lower accuracy. They're pretty nuts.
My understanding is that because the base knows exactly where it is, any atmospheric interference will affect receivers the same in a given region. So the base says "Hey you're getting this data from these GPS sats (it can only correct for data from shared visible birds) but you actually need to adjust what you think about the data from those by XXX because I know where I am and I can see the interference"
The rover (technical name for the moving thing) has two receivers that it uses for extremely accurate roll/heading calculations. Basically the right GPS receiver does position and the left receiver is used to calculate direction/roll. You can use 1 receiver and an IMU instead but it's not as good to save $200 bux on the F9P.
tl;dr we should get roughly 1-2" of repeatable year over year row over row accuracy.
Last edited:
Ah so it's like DIY WAAS but local and much more accurate. Nice.
You're literally doing dGPS, but with DIY stuff. That's amazing. dGPS receivers end up being about 1cm accurate if you integrate for long enough, but they cost a lot more. RTK is another word for dGPS using not just GPS satellites but also the glonass etc. satellites.
It's kind of bonkers that this can be done with COTS receivers. Back in the day, and by that I mean 2007, you needed a temperature stabilized, calibrated clock source that by itself was $1000. And that was just one part of the system!
It's a square rectangle thing. dGPS is a blanket term that is often used to talk about any corrective GPS, but DGNSS (dGPS) and RTK are two different ways of doing it, out of the main 4 correction types.
The F9P's are ~$220 in single quantities.
The F9P's are ~$220 in single quantities.
Last edited:
I've spent enough time with Tesla's self driving system to be really unimpressed with it. I suspect that the underlying tech is fine, the problem is the regulatory/liability controls. For instance: when it rains, the Tesla system tends to turn itself off. Super-irritating. Also, the driver attention monitoring is really aggressive and it has a high rate of false positives. Furthermore, when you get a strikeout, you don't even have cruise control anymore.
I really wish that there were a way to jailbreak the autopilot system in Tesla to turn off all of the annoying aspects.
I really wish that there were a way to jailbreak the autopilot system in Tesla to turn off all of the annoying aspects.
I've had two ford vehicles with radar based ACC and in both you can pick in a menu if you want ACC or standard cruise. If the radar is blocked or disabled in some way you can switch it and use it like traditional cruise. Tesla doesn't allow that, which is kind of insane.
Tesla released the FSD version in China recently and a lot of Chinese videos are up comparing the various Chinese ADAS vs FSD. The ones I've seen are mostly favorable to FSD but who knows for real.
Either way, Tesla is probably the only US company that can compete in the ADAS/FSD world. Could be BYD or Huawei or similar wins out in the end.
Either way, Tesla is probably the only US company that can compete in the ADAS/FSD world. Could be BYD or Huawei or similar wins out in the end.
In China there's very agressive censoring of anything bad related to Tesla. One of the major chiefs is a Musk fan and they are grateful for the tech transfer making the country the leader / top x in battery and EV tech.
I wouldn't trust the reporting. Then again, some tests of Chinese ADAS outside of China like in Thailand are also rather unfavorable. Both are maybe a mess?
It would be difficult for FSD to outperform the multi-lidar systems which are starting to become common in China now.
In the Out of Spec video I posted above about the Huawei EV with self-driving, the thing was looking around for parking spaces and letting you tap the spot you wanted and it would park for you as well as pull out of a tight space.
I think also the driver gets out of the car near a garage entrance and it drives itself into the garage and finds a parking spot and parks itself.
Mercedes is rumored to be making a deal with China's Hesai to purchase lidar units.
It would be the first time a foreign auto brand would be using Chinese lidar components for cars destined for sale outside of China.
Apparently cars with lidar in China are fairly common.
https://www.reuters.com/business/au...l-markets-with-chinas-hesai-lidar-2025-03-11/
So just over 2 years ago, Mercedes announced a similar deal with Luminar, with the goal of scaling up production and driving down costs:
https://www.theverge.com/2023/2/22/23608857/mercedes-benz-luminar-lidar-expand-adas-drive-pilot
Either Luminar failed to drive down costs to some threshold which it promised or Mercedes may have concluded that Hesai is more likely to produce in volume and at costs already lower than Luminar.
It makes you wonder if the main obstacle to adoption of lidar was their cost or the cost of the computing power and sensor fusion algorithms to integrate lidar data. Seems like almost every other SDC platform than Tesla are using lidar.
It would be the first time a foreign auto brand would be using Chinese lidar components for cars destined for sale outside of China.
Apparently cars with lidar in China are fairly common.
https://www.reuters.com/business/au...l-markets-with-chinas-hesai-lidar-2025-03-11/
So just over 2 years ago, Mercedes announced a similar deal with Luminar, with the goal of scaling up production and driving down costs:
https://www.theverge.com/2023/2/22/23608857/mercedes-benz-luminar-lidar-expand-adas-drive-pilot
Either Luminar failed to drive down costs to some threshold which it promised or Mercedes may have concluded that Hesai is more likely to produce in volume and at costs already lower than Luminar.
It makes you wonder if the main obstacle to adoption of lidar was their cost or the cost of the computing power and sensor fusion algorithms to integrate lidar data. Seems like almost every other SDC platform than Tesla are using lidar.
random data point "As with the P7+ sedan launched last year, Xpeng has removed LiDAR from the updated G6 and G9, with the smart driving system shifting to a pure vision solution."
It feels really silly but…I just got yet another week long strikeout for Tesla self driving. I really do not like tesla auto steer. It is way too cautious and it seems geared more towards minimizing possible liability for Tesla rather than being useful.
kind of thinking of installing a comma3x in the Tesla. Ridiculous, but Tesla‘s driver monitoring is way too aggressive and it won’t turn on half the time.
kind of thinking of installing a comma3x in the Tesla. Ridiculous, but Tesla‘s driver monitoring is way too aggressive and it won’t turn on half the time.
The other day, my mom and I were talking about how Waymo is just starting up in Miami, and touting that they finally feel like they can handle rain well enough to roll out there.
Then today I notice that this thread was spun out of the Tesla thread about eleven-and-a-half years ago. 80/20 rule at work, I guess. That last 20% really isn’t easy!
Then today I notice that this thread was spun out of the Tesla thread about eleven-and-a-half years ago. 80/20 rule at work, I guess. That last 20% really isn’t easy!
Unfortunate. Plain old camera vision systems can't compete with lidar for object recognition or seeing though rain, fog, and snow.
It needs to be really aggressive unless it's truly ready for SAE Level 5, and it's not. That annoying driver monitoring and reluctance to engage if everything is not right is saving your life.
Actually, no, I disagree. I have a comma.ai which has significantly less compute and fewer cameras. The comma.ai handles those situations just fine. The Tesla could too— i think hw4 is like the equivalent of a ps5 or something? The comma3x in comparison uses a snapdragon 845 which might be as powerful as a ps2 or something. It handles all of those situations just fine.