Self Driving Car thread

[MilleniX]

I understand the desire for some kind of generalized benchmark but much like in the HPC world, the real most useful benchmark is your real workload; in my case, I can get into my car in my driveway and tap "start self-driving" and it will drive all the way to my office parking spot in another city. Perhaps already better than a median Uber driver.

That's great, but does it do it with reliability sufficient for society to permit it on public roads?

Incidentally, my work in the HPC space has moved into a domain that has demands of reliability similar to licensing autonomous driving systems. I'm working on code that will be used as part of the US's operational weather forecast system. The operations folks definitely do not accept "it ran fast enough without crashing in the cases we happened to try" as sufficient assurance to deploy some new code to production. They want worst-case timings, they want fail-safes, they want debuggability/diagnosability in the event of errors, etc. Tesla continues to make no claims about any of those things, except in the narrow Robotaxi deployments, since FSD(S) is otherwise strictly an SAE Level 2 ADAS.

Incidentally, Waymo is going beyond that with external auditing of their processes and systems: https://waymo.com/blog/2025/11/independent-audits

 

[Shavano]

So I don't follow the news about this topic as much as some but I would assume that closed highway (ie all self driving, no pedestrians ) is a mostly solved issue at this point?

Besides the large issues of cost and inertia (probably need at least 70+% self driving cars), is there any reason to not start creating self driving highways where it is illegal to manually drive on them?

Several ginormous ones. A tiny fraction of cars on the road have self-driving features, and of those (mostly Teslas), a minority have self-driving service enabled. The drivers of manually driven cars are not going to pay for it. The owners of self-driving cars can't pay for them. That major obstacle aside, now you'll be reserving ginormous physical areas for the use of a
tiny minority of people. That's insanely inefficient.

The DARPA Grand Challenge and the fact that if everything is computer controlled it would behave in a predictable manner which can be accounted for.

You need to learn more about complex, dynamic computer control systems if you think that. It's simply not true in general. Not even in the special case where all the cars are running the same software, which they wouldn't be unless the government forced them.

By the way, if you're interested in these benchmarks I'm talking about: I'm interested too, because they don't exist. Or, well, there doesn't exist a well-defined self-driving relevant computer vision benchmark right now. This is one of the big regulatory issues - nobody has so far defined what kind of tests or capabilities are required for a self-driving system to be considered compliant in some way. UL has come out with benchmarks for computer vision over the years, but none are designed for video and none do realtime environmental scanning. It's kind of obvious that before we put any of these systems on cars, we should probably have tests to verify they do what they claim to do and are fit for purpose, like we do for literally every other safety system on cars.

While I agree that we're woefully behind the curve in defining requirements and have defaulted to let the billionaires do what they want with not even a system of deciding who is at fault in an accident when the self-driving system fucks up. That at the very least should have been debated by public safety experts, or by Congress, and a system should have been put in place before any automation at all was allowed to touch the controls of cars.

The particular questions about what the technical standards should be for an automation system are not easy to define. Humans don't read every traffic sign. Does a self driving system have to? What percentage of traffic signals can the system misread or react inappropriately to and still be considered "safe enough?" It's not zero for people, and people are allowed to drive. People vary considerably in their reaction time to expected and unexpected happenings in traffic, and probably in what they can readily identify and what they can't.

 

[chalex]

Maybe we can make a 15-min "driving test" and have the self-driving car drive the test circuit to prove that it knows how to do it all, just like we do for humans.

 

[w00key]

Maybe we can make a 15-min "driving test" and have the self-driving car drive the test circuit to prove that it knows how to do it all, just like we do for humans.

Let's not. Tesla can start with not hitting kids and obeying the rules around school busses and not crashing into first responders with on. The edge cases are what makes humans better, a big yellow thing is a context cue "F"SD misses easily.

 

[Xenocrates]

Maybe we can make a 15-min "driving test" and have the self-driving car drive the test circuit to prove that it knows how to do it all, just like we do for humans.

We have had those, in the form of test environments for a while. It turns out, it's possible to ace those environments, and still suck as a self driving algorithm. Humans we test the mechanical skills, because we assume they have judgement. We need to test the judgement of self driving, because the mechanical skills are relatively easy to program.

Trouble is, a human can generally recognize an edge case (flat tire, recently repaved road with no markings, first responders blocking lanes) and take appropriate action. They also communicate with people in the environment around them to get necessary information.
With self driving cars, they're limited to what sensors are designed in.
So, do you have the car refuse to operate if the TPMS system is inop or low? How do you deal with people who just put the TPMS sensors in a pipe and pressurize it, like mechanics did to my MiL's car to get the idiot light to stay off?
Do you infer shoulder sizes and where the markings should be, or do you cut out when you hit a newly paved section?
If there's a bridge plate in the road to cover large potholes or other roadwork, how much do you slow down, and can you tell it apart from general debris?
Can they tell if a cable in the road is high voltage, or just coax? Will they respond to a road closed with tape appropriately?

Self driving cars don't work at scale, because they don't have an interface to take direction from construction workers, first responders, or anyone around them. And trying to give them an interface programmatically will either be useless (IE, it's sufficiently privileged that a bunch of construction/utility people won't have access), or a massive security vulnerability (Car jacked by anyone in a hi-vis).

Self driving works just enough to be a serious hazard to not just the operator, but to everyone around it, and especially without proper V2X communications, should not be allowed to be operated on any public roadways without explicitly licensed and monitored test drivers and strict reporting requirements for incidents. Because those incidents can be used to help other self driving programs improve without reproducing deadly crashes like Tesla has had dozens of times, or Uber, or others, by designing to address those edge cases once they're raised.
I'd like to note that there are no fatalities associated with MB or Toyota's self driving development, and I've seen those systems first hand, in operation. Meanwhile, Tesla et al just keep YOLOing and deploying death traps to production.

 

[sryan2k1]

I've been on probably a dozen autonomous trains at various US airports in recent years.

A point to point glorified tram doesn't count.

 

[redleader]

A point to point glorified tram doesn't count.

Isn't that what he is proposing?

 

[demultiplexer]

The particular questions about what the technical standards should be for an automation system are not easy to define. Humans don't read every traffic sign. Does a self driving system have to? What percentage of traffic signals can the system misread or react inappropriately to and still be considered "safe enough?" It's not zero for people, and people are allowed to drive. People vary considerably in their reaction time to expected and unexpected happenings in traffic, and probably in what they can readily identify and what they can't.

There's many ways to approach this problem, but what's important here is to not approach self-driving cars as human-replacement robots driving cars. They're fundamentally different machines with fundamentally different failure modes. So whatever benchmarks or requirements we pose on SDCs, they should be aimed towards the expected failure modes of these cars, not failure modes of humans.

I don't have any answers on what that would be, but at least it's not the same as trying to emulate an acceptable level of failure as it relates to human drivers. And the acceptable level of safety is likely going to change over time, as this is one of those policy knobs politicians can manipulate to improve safety over time, a bit like CAFE is designed to improve gas mileage over time.

 

[w00key]

Humans don't read every traffic sign. Does a self driving system have to?

Eh, automated systems make the huge mistake reading wrong signs all the time. Off the side meant for an exit lane, nah, better bloop bliep you drive too fast, 50 kph limit now! On a 130 kph Autobahn.

Some cars have the option to adjust adaptive cruise control limit to read signs. It is the first thing you disable so you don't randomly brake check on the highway.

These systems also fail to detect end of restrictions when it is not explicitly marked like 8 out of 10 times, and stick to a 30 kph crawl limit long after construction zone / dangerous exit.

 

[sryan2k1]

I got on a residential road the other day where the limit was 40 MPH and for a while the car showed a limit of 70. The human in my knew that the 70 was wrong. What would the SDC do?

 

[rain shadow]

The Tesla folks say that the "end to end" model for FSD basically takes raw CCD data as input and produces steering/acceleration control as output, so that's 70ms between something showing up in the input and the car slamming on the breaks. I think you are right there are a lot of complications but just like with all the autonomous drones and humanoid robots, they can process/react/do much faster than humans.

I have no specific knowledge of their sensor fusion stack, but I would think that it's only 70ms for some tasks, like lane keeping. IFF to emergency braking could require several confirmation steps. IIRC some early versions had to be dialed back due to unnecessary (and scary) braking. The interval could be tested, something along the lines of a human figure popping up out of a manhole, or spotting (part of) a stopped vehicle when going around a blind curve.

I assume the reaction time will still be faster than a human, but I don't believe it will be 70ms. My gut instinct is the speed advantage is mostly from the car being able to activate the brakes electronically much faster than a human can move their foot off the accelerator to the brake pedal, and then push the brake pedal far enough to activate full braking to the point of ABS kicking in, and then a few more msec until ABS starts to manage wheel slip.

Mercedes has a system on some cars where it tries to mitigate the human foot delay by tracking how long it takes for the foot to move off the accelerator to the brake pedal. If that interval seems to be panic level, the car will go to full braking before the human has pushed anywhere near the full-braking level of force on the pedal.

 

[sryan2k1]

Mercedes has a system on some cars where it tries to mitigate the human foot delay by tracking how long it takes for the foot to move off the accelerator to the brake pedal. If that interval seems to be panic level, the car will go to full braking before the human has pushed anywhere near the full-braking level of force on the pedal.

Ford uses the front camera and radar to apply more brake than the user is commanding in emergencies.

 

[rain shadow]

Ford uses the front camera and radar to apply more brake than the user is commanding in emergencies.

Full braking is needed so seldom that few people even realize how quickly a modern car with good tires can stop. I strongly suspect human drivers could benefit from some training.

 

[w00key]

Full braking is needed so seldom that few people even realize how quickly a modern car with good tires can stop. I strongly suspect human drivers could benefit from some training.

Once in a while, when safe and no one behind and near, I command a full power stop. Did it once on snow (rare in NL) to see what level of traction I can expect - it's much faster than you expect.

But watch out for ice. I have slipped through a red light before at like 10 kph because the whole intersection was a skating rink, got flashed but no one got a ticket that day, "force majeure" I guess. In that case, better try to keep directional control over stopping, you don't want to spin, crash or fall into a ditch.

 

[chalex]

I assume the reaction time will still be faster than a human, but I don't believe it will be 70ms.

You don't have to assume or believe anything, you can just go try it empirically (I posted the video earlier in the thread).

 

[Chanur64]

Ford uses the front camera and radar to apply more brake than the user is commanding in emergencies.

Wouldn't that be a serious drawback in poor road conditions, such as slush or simply driving on snow? Or is the system smart enough to factor in road conditions before it takes action? Sharp breaking is the fastest way to break traction on snow or ice, or risk hydroplaning in heavy rain. Supposed to use gradual braking adjustments...

 

[rain shadow]

You don't have to assume or believe anything, you can just go try it empirically (I posted the video earlier in the thread).

The video is taken from behind and to the right of the driver. A camera positioned where the driver's eyes are would get back let's say (2ft/32mph) 40 msec of reaction time (relative to how the video portrays the situation) just from being ahead of the video cam, and probably another 10 or 20 msec from being positioned to left and able to see around the suburban sooner (not going to do the geometry computation for that, sorry)

The car is going 32mph in what appears to be 25 zone. The human driver may set the speed higher knowing the autopilot has a better reaction time, or more likely because they wanted to make the video more impressive. However, in day to day driving of a Tesla, the Peltzman Effect is real and probably mitigates some or even most of whatever reaction time advantage autopilot may have.

My personal experience is that I can see other things like a child chasing a ball, someone visible through car's windows, or their shadow, I have some knowledge of school schedules, and while my reaction time is worse than autopilot, I'm reacting to a set of information that includes a more diverse data set with earlier information. And I won't be going 30% over the speed limit in a residential area.

Whatever advantage Tesla's have in reaction time is either lost in the noise or is working against other safety problems:
https://www.roadandtrack.com/news/a...fatal-accident-rate-of-all-auto-brands-study/

Anyway, the video is kind of impressive but it doesn't really change my mind about safety, which is that Teslas using autopilot are probably still pretty safe, probably do have better reaction time than most drivers under most circumstances, but 70msec is not the whole story.

 

[MilleniX]

You don't have to assume or believe anything, you can just go try it empirically (I posted the video earlier in the thread).

Is that any sort of design assurance, based on WCET analysis? (trick question - Tesla won't tell us shit about their safety standards) Or, is it literally a single sample of an arbitrary version in arbitrary circumstances?

 

[CUclimber]

Rivian is making a huge vertically-integrated push into self-driving: https://rivian.com/autonomy

I'm not bitter as a Gen1 owner who will never see this on my car, and I trust RJ more than Elon when it comes to anticipated timelines and features.

 

[sryan2k1]

They have a lot of features RJ promised on the Gen1 that still don't exist or they've flipped on and now say won't exist. I don't believe anything until it exists anymore.

The launch edition R2's not having Lidar is kind of a bummer though.

 

[CUclimber]

They have a lot of features RJ promised on the Gen1 that still don't exist or they've flipped on and now say won't exist. I don't believe anything until it exists anymore.

Tank Turn is the only one I can think of. Unless there was a V2H charging feature as well?

 

[sryan2k1]

Launch mode is the other thing I can think of off the top of my head but I recall there being some other stuff.

 

[ramases]

Full braking is needed so seldom that few people even realize how quickly a modern car with good tires can stop. I strongly suspect human drivers could benefit from some training.

US driver ed problem. Other parts of the world sorted that one out decades ago.

In Austria you must take a road safety and basic emergency recovery maneuver course 3-9 month after you get your license. This requirement was introduced in the early 2000s.

During the course the instructor will make you slam your brakes as hard as you can on various surfaces, starting from dry racing-grade tarmac to dry regular street tarmac to wet tarmac to simulated snow (the courses come equipped with floodable sections whose surfaces have similar traction to snow when flooded).

The 'basic' required version of the training does this starting from 50 kph to at least 80 kph for dry tarmac, but you can also book additional courses (often heavily subsidized by your insurance provider) where you get to do that at whatever speed you want to. Okay, not quite whatever, somewhere above 130 kph the instructor will tell you to cut it out.

 

[MilleniX]

US driver ed problem. Other parts of the world sorted that one out decades ago.

In Austria you must take a road safety and basic emergency recovery maneuver course 3-9 month after you get your license. This requirement was introduced in the early 2000s.

During the course the instructor will make you slam your brakes as hard as you can on various surfaces, starting from dry racing-grade tarmac to dry regular street tarmac to wet tarmac to simulated snow (the courses come equipped with floodable sections whose surfaces have similar traction to snow when flooded).

The 'basic' required version of the training does this starting from 50 kph to at least 80 kph for dry tarmac, but you can also book additional courses (often heavily subsidized by your insurance provider) where you get to do that at whatever speed you want to. Okay, not quite whatever, somewhere above 130 kph the instructor will tell you to cut it out.

I can only wish such things were required in the US.

Out of curiosity, do these courses require drivers to demonstrate any particular proficiency while executing those maneuvers, or just to experience them? Does a new driver's license get revoked if they fail the course?

Or is the initially-issued license more akin to the US "learner's permit" system, in which new drivers initially have to have supervision while driving, and then maybe have limited independent privileges, but take a (minimal) competency test to graduate to a less restricted license? (In many US states, licensed teenage drivers still have limits on passengers and time-of-day to contain the scope of peer-influenced bad decision making)

 

[rain shadow]

Or is the initially-issued license more akin to the US "learner's permit" system, in which new drivers initially have to have supervision while driving, and then maybe have limited independent privileges, but take a (minimal) competency test to graduate to a less restricted license? (In many US states, licensed teenage drivers still have limits on passengers and time-of-day to contain the scope of peer-influenced bad decision making)

Also far as the US goes, it's become more common (although still state-by-state) for teenaged driver blood alcohol limits to be much lower.

 

[sryan2k1]

Also far as the US goes, it's become more common (although still state-by-state) for teenaged driver blood alcohol limits to be much lower.

It's 0 in every US state, not sure how it can be less than that.

 

[ramases]

Out of curiosity, do these courses require drivers to demonstrate any particular proficiency while executing those maneuvers, or just to experience them? Does a new driver's license get revoked if they fail the course?

You can't fail, it is a teaching and not an exam environment. Some of the stuff they make you do can be a pretty terrifying experience for new drivers.

On the higher end of the stress spectrum (and IRC this is something that's optional in the sense that you can opt out) you have exercises like this:
1) Switch off electronic stability control program
2) Accelerate to about 50 kph on a wet surface that simulates a snowcovered road
3) Drive over a plate that will jank the rear wheels to left or right (random choice), causing the car to enter a skid
4) Accept the spin and just get the car to stop ("only" requires nerves; do not apply steering input, apply full brake and keep it that way until you stop. As long as you keep doing that the tires will not regain directional traction and you'll continue in a straight line until whatever friction is to be had has stopped the car) or try to recover from the skid using steering inputs (not every new driver will be able to do that; this one is definitely optional because if you fuck it up in certain ways you can flip the car)

As you might imagine this can be an extremely high-stress exercise for people who've not driven that much before and who never rode out a spin or recovered from a skid. Adding the pressure of an exam to that would simply be counterproductive. Hence the focus on exposure and experience instead of proficiency.

Or is the initially-issued license more akin to the US "learner's permit" system, in which new drivers initially have to have supervision while driving, and then maybe have limited independent privileges, but take a (minimal) competency test to graduate to a less restricted license? (In many US states, licensed teenage drivers still have limits on passengers and time-of-day to contain the scope of peer-influenced bad decision making)

No, once you pass the theoretical and road driving test (which will include at least 40 minutes in life traffic) you just have to take the course. You do get a provisional license that is easier to yank if you speed or do other stupid stuff for the first three years. Booze + driving (even below 0.05 BAC) = immediate yank. The provisional status automatically expires if you manage to drive for three years without getting busted for certain things.

The closest we have to the US learner's permit, where starting with 16 you have to drive a total of 3000km under supervision of your parents (or any other named and approved guardian; you must have supervision and cannot drive on your own). Then you have to pass the regular road driving test, and can get your license with 17 instead of the usual 18.

Statistically those that get the license via this scheme instead of the traditional drivers ed in a driving school have, despite obtaining their license at 17 instead of at 18 (or older), a 15% lower chance of an accident.

 

[rain shadow]

It's 0 in every US state, not sure how it can be less than that.

Lower than limits on adults is what I meant

 

[sryan2k1]

I still don't understand. Drinking age is 21 in all US states. The legal BAC for anyone under 21 is 0%

 

[rain shadow]

I still don't understand. Drinking age is 21 in all US states. The legal BAC for anyone under 21 is 0%

MilleniX was listing some teenage driving restrictions in the US, but did not list the lower BAC levels compared to 21 and over. I was only attempting to complete the list. The limit is not exactly 0%, it varies state by state, but is usually very low, between 0.00% to 0.02%

 

[Quarthinos]

MilleniX was listing some teenage driving restrictions in the US, but did not list the lower BAC levels compared to 21 and over. I was only attempting to complete the list. The limit is not exactly 0%, it varies state by state, but is usually very low, between 0.00% to 0.02%

But why is it not 0%? If you can't legally drink alcohol, why can you legally drive with a measurable BAC?

 

[Chuckles]

But why is it not 0%? If you can't legally drink alcohol, why can you legally drive with a measurable BAC?

Measurement error bars. There are things that the body produces that are or look like alcohols, and the air itself can have things that are or look like alcohols. If your machine is set to zero, it wills give false failures.

 

[sryan2k1]

Maybe that 0.02 is okay for a FST but I assume a blood draw limit is 0%

 

[curih]

It would make sense for some allowance of a very minimal level. I can't say how the rules are actually written. But alcohol is not exactly uncommon as a cooking ingredient for example. It's not going to get you drunk used that way, but might just tick into detectable levels. Also sometimes in medicines.

 

[sryan2k1]

Says someone who hasn't chugged bitters (Which are normally around 45% ABV and not age restricted)

 

[smartalco]

To nitpick here, it’s not illegal for under-21 to drink in a lot of states (over half), it’s just illegal to give/sell/whatever alcohol to someone under 21, except under specific circumstances (generally, in your home under parental supervision). It’s perfectly legal for a 20 year old to be drunk, they just obviously shouldn’t be driving.

 

[AbidingArs]

The two would be different crimes (underage drinking vs driving while intoxicated). As this is state law, it will vary by state; Pennsylvania has the following law against underage drinking:

§ 6308. Purchase, consumption, possession or transportation of liquor or malt or brewed beverages.
(a) Offense defined.--A person commits a summary offense if he, being less than 21 years of age, attempts to purchase, purchases, consumes, possesses or knowingly and intentionally transports any liquor or malt or brewed beverages, as defined in section 6310.6 (relating to definitions). For the purposes of this section, it shall not be a defense that the liquor or malt or brewed beverage was consumed in a jurisdiction other than the jurisdiction where the citation for underage drinking was issued.

But has a different standard for driving under the influence.

a) General impairment.--
(1) An individual may not drive, operate or be in actual physical control of the movement of a vehicle after imbibing a sufficient amount of alcohol such that the individual is rendered incapable of safely driving, operating or being in actual physical control of the movement of the vehicle.
(2) An individual may not drive, operate or be in actual physical control of the movement of a vehicle after imbibing a sufficient amount of alcohol such that the alcohol concentration in the individual's blood or breath is at least 0.08% but less than 0.10% within two hours after the individual has driven, operated or been in actual physical control of the movement of the vehicle.
....
(e) Minors.--A minor may not drive, operate or be in actual physical control of the movement of a vehicle after imbibing a sufficient amount of alcohol such that the alcohol concentration in the minor's blood or breath is 0.02% or higher within two hours after the minor has driven, operated or been in actual physical control of the movement of the vehicle.

I'm not sure if this is due to limitations in detection, calibration of the test, or some other factor. Presumably there are also differing penalties for the different crimes.

Edit: posted too soon

 

[sryan2k1]

To nitpick here, it’s not illegal for under-21 to drink in a lot of states (over half), it’s just illegal to give/sell/whatever alcohol to someone under 21, except under specific circumstances (generally, in your home under parental supervision). It’s perfectly legal for a 20 year old to be drunk, they just obviously shouldn’t be driving.

This gives Ron White "Drunk in public" vibes.

 

[chalex]

Rivian is making a huge vertically-integrated push into self-driving: https://rivian.com/autonomy

I'm not bitter as a Gen1 owner who will never see this on my car, and I trust RJ more than Elon when it comes to anticipated timelines and features.

Gotta read the fine print:

2. Universal Hands-Free will not stop or slow down for traffic lights or stop signs. Review the Rivian Owner's Guide for additional details about this feature.

 

[smartalco]

This gives Ron White "Drunk in public" vibes.

Right, I’m not recommending it as a life choice, just as a counter to:

I still don't understand. Drinking age is 21 in all US states. The legal BAC for anyone under 21 is 0%

Which is just false in over half the country.

Edit: this is way too off topic at this point, so I should have let it go.