Why didn’t auto-braking stop these crashes? | Hacker News Reader

Why didn’t auto-braking stop these crashes?

latimes.com

96 pointsheavyset_go138 comments

Because Tesla effectively doesn't have AEB while AutoPilot is enabled.

In a Tesla AEB should be an entirely self-contained ("dumber") system. Its only job should be to look for any ground-level object ahead and brake. It should do this using simpler but more reliable ("dumber") techniques like forward bumper-level radar or camera parallax, not NN/deep learning/object recognition/AI.

To use an analogy, AEB should be like two-factor authentication. If your "second" factor is the same as your first, then it isn't two-factor. In this case AutoPilot is the "first" and "second" factor i.e. it is AEB and in control of the vehicle.

So Tesla's vehicles absolutely have AEB, when AutoPilot is disabled, because it is offering a "second opinion" relative to the driver. So IIHS's ratings are correct in that instance. But as soon as AutoPilot is enabled it is no longer a "second" opinion and therefore may as well not exist.

Unfortunately Tesla aren't alone in making this mistake. We're seeing other vehicle manufacturers stepping into the vehicle automation space double-dipping their auto-drive systems into AEB. It is a bad practice regardless of who, AEB should be its own system with its own simpler/dumber logic that protects people from auto-drive mistakes, not compounds them.

Does Tesla even have a radar?

I dont think so. I thought their whole gimmick or goal was to drive entirely by cameras and nothing else. Because humans only have eyes.

Though just speculation and no source.

Due to the current chip shortage they've moved to a vision-only system, as they had difficulty getting radar units in the quantity required. Until recently however, all Teslas had radar since ~2014. The high profile crashes are mostly/all radar-enabled vehicles.

The thing about automotive radar, and this isn't just Tesla, is that stationary objects are normally excluded. Otherwise the system brakes for things like overhead signs that have high radar reflectivity. Stopped cars, concrete barriers, etc are all basically invisible to most automotive radar.

Tesla also isn't the only manufacturer taking the vision approach. Subaru, for example, was doing vision-only emergency braking in 2016.

> The thing about automotive radar, and this isn't just Tesla, is that stationary objects are normally excluded. Otherwise the system brakes for things like overhead signs that have high radar reflectivity. Stopped cars, concrete barriers, etc are all basically invisible to most automotive radar.

I believe most automotive radars in the last few years are FMCW exactly to accommodate adaptive cruise control.

When we bought our used Suzuki, one of the differences to another car we looked at was that ours had stereo vision AEB and the other had radar - one of the selling points by the dealer was that the camera version could detect stopped objects, which I always figured was half the point of AEB.

I'm quite sure I've heard something pretty close to those words from Musk himself in some video.

Humans only have eyes... and the most powerful supercomputer in the world between their ears.

The goal is to cut costs and keep shipping during the chip shortage. Ignore the excuses.

Humans also have ears. And a sense of touch. And a sense of motion. And that supercomputer between the ears has decades of experience doing sensor fusion amongst all the various inputs down to the level of milliseconds of variation, and decades of experience of rising in cars and knowing what they feel is safe versus what is not, and presumably at least months of experience of control behind the wheel.

Tesla doesn’t have a lot of any of those things.

To my knowledge, they don't even have binocular vision. For several angles there is only one camera. Even where there are multiple cameras, they are clustered together and so of limited use for binocular vision. It seems there are multiple cameras to improve range and FoV rather than depth perception.

So it's like a small animal with one eye closed.

> Because humans only have eyes.

This is the most ridiculous line of reasoning I've heard in a long time. Wow.

Humans only have eyes? So Teslas can smell the milk and throw it out if its gone bad? Or listen to a faraway sound and know that its getting closer at an alarming or not-alarming rate? Or feel the road getting harsher and understand that the GPS is incorrect?

> So Teslas can smell the milk and throw it out if its gone bad?

Cars do not need that skill.

> Or listen to a faraway sound and know that its getting closer at an alarming or not-alarming rate?

I'm sure the car has multiple microphones, it just needs the brains to process the input.

> Or feel the road getting harsher and understand that the GPS is incorrect?

Again, I am sure the car has multiple G sensors, accelerometers and gyroscopes. It just needs the smarts to process the input.

> Cars do not need that skill.

As a driver, I can certainly sense my car smelling funny if something goes wrong :)

Yes, at least some do, and they recently disabled it in a recent software update

Any info as to why? Simply because it is redundant or is there a technical problem with it?

IIRC there's concern of crosstalk as it becomes more common. So relying on it in a radar-noisy environment would be risky.

Or they just don't want to maintain two different versions of their software for cars with and without it.

They removed it on new models early this year.

Many manufacturers are removing radar in their latest model spins fwiw.

Suddenly my 2016 car just got much more valuable to me.

Except you are at the whim and mercy of Tesla to keep providing updated support using lidar. They will probably just not use the lidar any more and merely use the cameras(which you also have) like they are on the 2021 Model 3/Y

Notably, Tesla cars have never used Lidar, which is a lot more precise than radar, but it costs more.

Lidar costs have come down. Now it is around $1,000 soon it will be under $500

https://www.barrons.com/articles/lidar-is-the-future-of-auto...

Tesla uses common technology available. Mass-produced consumer cars haven't yet adopted lidar, because the technology has been too expensive and bulky, until very recent development. The next-gen premium cars will likely offer lidar as an option in near future.

One little known fact is that lidar has been in cars for a long time. Volvo for example has had lidar in some cars going back almost 10 years IRRC (I saw a presentation by a Volvo engineer talking about this, can't find any link right now because it's all swamped by the Volvo luminar collaboration). It was very different type of lidar though mainly used for keeping distance to the car ahead (and maybe AEB?). The thing about "modern" Lidar is that there are still so many open questions (weather, how to keep the apertures clean...) , apart from the fact that they still are way too expensive.

I like the 2-factor analogy for the driver-assist technologies, which are ordinarily framed as more primitive automation, rather than the low-cost safety boons that they are.

The Level 3 automation systems that Tesla and Uber built rely on two factors too, but they invert them, expecting disengaged humans to snap to alertness at super-human speed, which is just not how our brains work. So of course the second factor only exists as a legal cover to reduce their liability, which is all these players care about as far as safety is concerned.

It always bugs me they trot out the Mountain View crash as an example where a Model X struck an unmarked and unprotected jersey barrier end in what visually looks like a traffic lane. Further that multiple human drivers had hit the divider in the same spot and a near fatal accident occurred literally the next day into the same barrier.

If I put a brick wall on the freeway and paint it to look like a tunnel ala wil-e-coyote am I responsible for building the wall or is the driver responsible for hitting it?

Is this the one you are talking about?

https://www.mercurynews.com/2018/03/30/tesla-autopilot-was-o...

I'd blame a human driver for hitting it in this case.

Were any of the person-driven accidents fatal, do you know?

If it fools .1% of human drivers and 50% of automated Tesla's I'd say is Tesla's fault.

It’s Mountain View California - fully 50% of the cars on the roads are Teslas

The Mountain View crash where the driver reported that specific location to Tesla on multiple occasions and still let the robocar drive full speed into the jersey barrier on autopilot?

It sounds a bit like the 3 foot wide concrete pillar that came out of nowhere and scratched half way down the passenger side of the car recently.

You can blame the driver but if tesla calls their systems "auto-pilot" or "full self-driving" mode then they're creating expectations in people's minds that they aren't going to be kamikaze'd into a concrete wall... When they create these expectations and then people die, something needs to change.

Even if they didn't call it that, anything that allows drivers to redirect their attention seems dangerous to me. Because it is always going to be hard for a disengaged driver to regain control and evade an accident in the last second.

Automatic braking doesn't get much attention, but it seems like something that could be incredibly valuable if it got to the "clearly better than human" level.

This is purely anecdotal, but I find the automatic braking system on my Mazda CX-9 to be fairly annoying. Once every couple months or so, it has a false alarm and brakes while blaring out alerts, in an urban environment where I don't observe anything out of the ordinary at all. And it's never activated in a situation where I found it to be helpful.

That said, it's possible that an annoyance every two months is just the cost of a system that really would help me if I got into bigger trouble. It's hard to say from purely my own anecdotes.

> Automatic braking doesn't get much attention, but it seems like something that could be incredibly valuable if it got to the "clearly better than human" level.

It's important to point out here that a handful of back of the envelope analyses are out there that show that Tesla's autopilot is already "clearly better than human" level at avoiding emergency vehicles.

These are very common accidents. Emergency vehicles in travel lanes get hit all the time, crews are carefully trained to avoid that situation if at all practical, and taught how to reduce personal risk when that's not possible.

In fact (and I don't have a link handy) if you extrapolate out from a few different data sources it looks like Tesla's on AP get in these crashes about half as much as vehicles in the general fleet.

Now, that doesn't mean that these 13 events don't constitute a cluster worth investigating. And it doesn't mean that there can't be a bug worth fixing. It doesn't even mean that the media shouldn't write about it.

It does mean that we should be very careful with pronouncements about safety from a sample set of this size, though.

I think the problem here is calling something "better than human" because it manages to avoid some accidents. If a car avoids one accident for every 100,000 times it intervenes, that isn't really a good thing. It just means that drivers start to ignore flashing lights and warnings because they are always meaningless. Maybe it even causes some accidents by distracting the driver, and they don't get categorized as "caused by the automatic braking system" because nothing really tracks that.

Every time I start my car, about 15 seconds into the drive a popup appears on the screen, saying something like "Warning: automated driving assistance system can cause distraction. Select OK to dismiss". Thanks, you have reduced your own liability in exchange for further distracting the driver.

>In fact (and I don't have a link handy) if you extrapolate out from a few different data sources it looks like Tesla's on AP get in these crashes about half as much as vehicles in the general fleet.

Why use that as a comparison? The obvious confounder would be that people would disable autopilot in poor conditions, making such crashes a priori more likely. The relevant comparison would be Teslas (with and without AP) against a similar class of cars, maybe BMWs.

That's nitpicking. There's an initial contention: "Teslas hit emergency vehicles frequently", and a simple refutation: "No, they don't, here are some numbers[1]". Your response only makes it "not impossible" that the initial contention is correct, but you have to show numbers to prove that.

All we have is a cluster of 13 accidents over 3-4 years, and there are something like thousands of such accidents in the US every year. That's not proof that this is noise and not signal, but the obvious hypothesis is that it's probably noise[2]

[1] Which I don't have, though I suppose I could track it down.

[2] For the record: I believe it's probably a real bug, but in a product that on the whole is clearly better, so it's well hidden in the noise.

The initial contention was that tesla's autopilot was superior to humans, at a specific task. If this were true, it would be very interesting, and represent real progress in the self driving space. Thus I'm interested in whether it's true, and proposed a method by which that might actually be determined.

What's the source for this claim? Are you accounting for the fact different classes of vehicle have differing accident statistics (e.g. comparing teslas to similarly old and expensive alternatives), and that autopilot's miles are not a random sample of all driven miles?

No, because it's enough that it exists as an existence proof refuting the hypothesis. The burden is on people contending that e.g. AP can't stop for emergency vehicles (or the related point in the linked article that emergency braking doesn't work) to show that it is unsafe. That analysis doesn't hold for obvious interpretations of the data we have. That doesn't mean it's wrong, it means you need to come to the table with more work. It's not enough to argue with me on HN.

If you're merely questioning whether Tesla's really do crash into emergency vehicles more often that would be expected, I get it.

But "Tesla's on AP get in these crashes about half as much as vehicles in the general fleet" is unsupported; even in Tesla's own figures from https://www.tesla.com/VehicleSafetyReport the difference between AP and non-AP Tesla's isn't large enough (for me anyhow) to actually draw any conclusions from, since AP usage is obviously correlated with road conditions. To be clear: I'm not saying it's at all obvious that AP does worse than a human driver, but it's neither is that possibility ruled out by anything they've posted there - Tesla simply choose to present statistics without the context required to interpret them, likely because it suits their marketing aims. News at 11.

The human instinct to retain control even when that's not to their advantage gives me the hunch that it's more plausible that the apparent numerical benefits of AP are simply due to selection bias; people turn it on in situations that are less messy, and therefore it's involved in fewer accidents. Perhaps these investigations will reveal more.

Put it this way: if autopilot actually were better than a human driver, and humans choose to use it in less dangerous situations, and those drivers at least sometimes pay enough attention to catch the occasional AP oversight (better isn't perfect!) - well, then I'd expect the AP crash numbers to be dramatically lower than the human crash numbers. They're not.

I do think the Tesla AP numbers look less bad than they did a few years ago anyhow, to be fair. I just don't think it paints a clear picture yet.

> the difference between AP and non-AP Tesla's isn't large enough (for me anyhow) to actually draw any conclusions from

Fine. But then why do you feel so certain about a data set of 13 points over three years? The point under discussion is a contention that AP is less safe because the autobraking doesn't work with emergency vehicles.

I'm not sure what you're talking about? Which 13 points? Maybe you're confusing me with a different poster? Anyhow, my objection is to the positive assertion that the AP is safer based on published data. And even there - I can understand the positive interpretation at first glance; I just don't share it. As to the formal investigation - there aren't any conclusions yet, so no news there.

The one on my 2016 C-class Mercedes was so bad that it almost caused two accidents. If not for years of pingpong... I got rid of it after having it checked out and declared healthy, if that system wasn't broken it might as well be. Worse than useless. Don't get a MB with automatic emergency braking unless you want to covertly commit suicide.

"The one on my 2016 C-class Mercedes was so bad that it almost caused two accidents. If not for years of pingpong..."

Care to elaborate? Both how it nearly caused an accident, and how you avoided one seem interesting!

There were two instances that were particularly memorable, the first was when passing through a steel bridge with fairly narrow spacing, the car - without any warning - slammed on the brakes just as I was passing the entrance to the bridge. The road surface was wet from recent rain and the car immediately started to spin due to uneven grip. I managed to avoid the posts :) This happened once more in the same spot but then I was already more alert to this kind of nonsense so that time it was pretty uneventful even if the car - again - slammed on the brakes.

The other situation was when in a tight right hand turn a large advertising sign caused the car to believe I was about to have a head on collision (when in fact the object was merely passing right-to-left through my field of view). This caused me to end up on the wrong side of the road, if there had been opposing traffic it would have been a head on collision, exactly the thing it was trying to avoid.

I brought the car to the local dealership, they checked it over and pronounced it sound, I sold it a week later. Murderous beta grade - or not even beta grade, proof of concept - software in that car. I was pretty loyal to the brand but I don't think I'll ever have another one, besides the plastic bling interior the fact that they shipped that junk has made the brand worthless to me.

Now I drive a 24 year old car that was re-built end-to-end. It has an ECU, ABS, an aftermarket mp3 player / navigation combo and that's it.

How would you define "clearly better than human" for a driver assist technology? Like if 'driver + automatic breaking' gets into fewer crashes than 'driver', is that a success? What if the assist is constantly doing phantom breaking, but never hits anything?

If it's a lot fewer crashes, and phantom braking happens rarely, that would be pretty strong evidence. After some brief looking around the internet, I didn't really find any good evidence, or any weak evidence, it seems like we don't really know. I wouldn't be surprised if some of these systems perform far better than others and the general lack of data and transparency makes it hard to figure out whether they are better than human, or not, or it's a mix.

Toyota's safety sense system has saved me from minor accidents a few times. It has never auto applied the brakes, but the alarms and brake assist have been really helpful on freeways during abrupt stops. My girlfriend, on the other hand, hates her Honda's auto braking system, as it triggers too easily, even on the lowest setting. I've seen it regularly trigger when pulling up to sensor controlled gates.

I have a Toyota and my auto-braking alarm thing has only gone off a few times... all very clearly my fault for losing attention for a brief moment. To be honest, I have no idea if it's gone into applying-brake-mode. I'm guessing yes, but... it hasn't been often nor can I even remember really what happened. I do know that it prevented crashes, whether it applied the brakes or alerted/scared me into doing so.

There has never, ever been a false alarm though. Not even once.

Just as another point of reference, my (2021) Mazda 3 AEB has only ever activated when another driver stopped short in front of me. I believe it's configurable in settings, though your experience may be worse due to environmental concerns (bugs, rain, dust), different model year, or simply being a bigger vehicle (with a different bumper height or more mass requiring more time to stop).

What happens when the auto braking engages? Does it just slam the brakes on and stop dead? Wouldn't that be a big risk of you getting rear ended by someone behind you (especially if it's a false alarm)?

Yes, it goes into full anti-lock braking, but it only does this at the very last possible second it can be done without a collision - the driver usually gets enough time hearing an audible alarm to stop, provided there is enough time.

I haven’t heard of any rear end accidents due to false alarms with AEB, though that doesn’t mean they don’t exist. I’d guess the system is tuned to need enough confidence that a forward collision is likely to engage braking, knowing of the potential consequences of suddenly and dramatically stopping.

In my Passat it just barely taps the brakes rapidly to get your attention, I've never had it go further then that but in theory it will lock them up if you don't respond. If it detects someone in the lane you are about to merge into, it cuts off the power steering for a millisecond, just long enough for you to notice the car in your blind spot but not long enough to prevent you from going into the lane if you really want to. It's pretty close to perfect in my opinion.

Depends on the vehicle. My 2014 Chevy pretensions the brakes, but doesn't apply meaningful force.

My 2021 Hyundai alerts and pretensions as well. The two situations it's come on, I've been aware of the situation and actively moving to brake. Had I not actually applied braking, I believe the vehicle would have in my place.

I somewhat suspect AEB on the Hyundai will lead to a collision, but at a significantly reduced speed. In other words, a fender bender instead of mangled mess.

Automatic Emergency Braking is amazing and it never gets talked about. I’ve logged 40k miles in my 2019 Subaru with Eyesight and it very reliably notices dangerous situations. Other vehicles hard stopping in front of me is the most common occurrence, and it always hits that sweet spot of me being glad it alerted me so I could brake before it had to.

I’ve had one instance where it went straight to braking - I took a right turn at low speed onto a neighborhood street and encountered a cyclist in my lane coming quickly and directly towards me, about 20 feet away. The car stopped itself. It was jarring but I wouldn’t say it was an overreaction.

I can only think of 1-2 cases where it’s truly given a false alarm - both at low speed on a city street with thick steam coming from a manhole in front of me, which Subaru says in the manual is a difficult situation. Both times I only got a momentary audible alarm; tap the brakes and it stops.

The reliability of AEB definitely varies by manufacturer, but I’m impressed by it.

Second this. Really impressed with Subaru's eyesight system. No false alarms till date. Twice in 3 years it alarmed when cars were turning in front of me and it was too close for the car to stay comfortable but i would have been okay with it. 3 times it has saved me from frontal crash at highway speeds by applying AEB.

I have a Hyundai with a similar system. In urban highway traffic, I've found it exceptionally good at reacting to drivers going from "barely braking" to "stopping hard" before I can physically react. Pretty incredible system.

> Other vehicles hard stopping in front of me is the most common occurrence

Wow, is mine defective? I have the same vehicle. My most common occurrence is backing out of my steep driveway at faster than 1 mph. We made a habit of disabling it as soon as we start the engine. The second most common occurrence is driving past the narrow neighborhood roundabouts everywhere in Seattle.

It’s worth mentioning that I’m only talking about Eyesight while driving forward. The rear ARB is radar controlled and very sensitive. I’d never say it’s false alarmed, but it’s very, very sensitive. It may be a matter of preference to some extent. At low speed in reverse I appreciate the added layer of caution.

It’s really the type of safety system consumers cannot reliably judge from first hand experiences since preventing a near accident will never be as vivid or memorable of an experience as it malfunctioning and nearly causing an accident.

My 2018 SEAT (VW group tech) has twice saved me from an expensive mistake by slamming on the brakes rather than reverse into an obstacle.

The first time I'd only had the car a couple of days and had accidentally engaged reverse when I thought I was in first (the previous car had a different gear layout).

Another couple of times it's squeaked at me at low speed for no obvious reason. On balance, I'm extremely pleased with it.

My hypothesis on this is that all automaker's automatic emergency braking systems likely function at approximately the same level, but Tesla drivers are disproportionately likely to rely on it due to the use of Autopilot being so incredibly common. Statistically, drivers of other brands that include similar lanekeeping abilities are far less likely to use them compared to Tesla owners' use of Autopilot.

Should it be happening? Well the NHTSA doesn't seem to mind it. Maybe we should be asking them why their automatic emergency braking tests are so stupidly out of agreement with what drivers expect and automakers promise that the systems are able to do. The freight industry has a high speed test and their systems are top notch.

Sorry, your hypothesis makes no sense. There are WAY more cars sold by other brands, so more Tesla incidents is even more galling. Using autopilot should only HELP the auto braking not make it worse.

Other systems use radar and camera. They are thoroughly tested by the OEMs and suppliers to strict tolerances.

Meanwhile Tesla is shipping beta software and having people test it. Including removing radar because of “phantom” braking.

OPs point was tesla drivers are more likely to use the features. More cars being sold by other brands, could refute that but it doesnt necessarily. Its about the rate of usage of the software X total sales X avg. miles driven

This article addresses automatic emergency breaking, which is ON by default in the vehicles discussed.

Right but his point is that Teslas are more likely to crash in the first place because people are using Autopilot, which isn't as good as human drivers. So automatic braking gets used more.

I have a Hyundai with radar based adaptive cruise and AEB. Notably, with cruise control enabled, it gives visual proximity to a leading vehicle.

The only case I've ever had an issue is at the crest or base of a steep hill on a secondary road. The angle does not allow the radar to work at the right angle.

It's kind of interesting that this appears to happen with police cars and emergency vehicles mostly. Is there a particular way of parking such vehicles that makes it hard(er) for a Tesla to detect it?

It might also indeed be about phantom braking. My relatively old (probably designed around 10 to 20 years ago) system gets confused sometimes when I pass a truck or a bus and once even when I passed a bicycle. It happened at least 5 times in less than 10,000km. So I suspect there is some merit to the Autopilot disabling the feature.

I wonder if the cameras or ML are thrown off by flashing emergency lights or emergency vehicles' livery (most cars on the road use just 1 paint color)

Yeah. I'm as big a Tesla booster as you'll find but I'm in the "this is probably a bug in recognition" camp. They can presumably fix this via properly parametrized modelling. It's not a rare edge case at all.

I have a hunch that at some point, there'll be an investigation and a discovery process, and we'll learn that the training dataset deliberately _excluded_ emergency vehicles because they didn't represent "normal driving" or something.

That's purely uninformed speculation, but I could see it happening like this: Early on, someone says "The totality of situations encountered on the road is too complex. We're not building a system to do everything, we're building it to take over during the boring parts of a drive. Only teach it the boring stuff." Then a few years later, someone else, forgetting those initial assumptions, says "Hey this thing has done like a billion miles of boring stuff, why wouldn't you trust it for more?"

No, the car sees emergency vehicles. It sees stopped vehicles. It sees vehicles stopped across lanes. This situation is far, far more common than the count of actual accidents we've seen. There probably is a bug, but it's not remotely as simple as "can't see fire engines".

Until recently Tesla relied on forward microwave radar to detect collisions which is very bad at detecting stationary objects.

Great majority of stationary cars on a highway are emergency vehicles, hence the 14 or so collisions.

I think the flashing lights and appearance might degrade autopilot. Other cars also have issues because they use radar (except for some new models with camera-based systems like the 2022 Honda Civic) and ignore stationary targets when they see a moving vehicle as a more likely "target". Some systems use a combination of radar and camera (with camera providing low-speed, parked car, bike, and pedestrian detection). Then it's up to the quality of the algorithms and tests have shown that some cars are pretty poor at detection.

The human visual system is amazing at normalizing across many orders of magnitudes (cf HDR imaging), but yet with the evolution of LED lights and a lack of accountability, the lights on police cars have gotten way too bright for even human drivers at night. I wouldn't be surprised if the ultimate cause is a complete swamping of the cameras for everything that isn't immediately in front of the car, and even the lane lines disappearing when close enough to the blinding lights. What is a computer supposed to do when confronted with complete loss of visual input? A human will squint, slow down by 10-20mph, extrapolate the lane lines, and play the odds. But there is no surefire way to proceed safely, and thus no straightforward action to program a computer with.

Well, if the computer can recognize the situation it can always slow down, alert the driver, and eventually stop. That's the nice thing about cars: There's always a safe default mode. I suspect that the computer cannot recognize this problem reliably, though.

Slowing down to under 40mph is not a safe thing to do on a highway and is illegal in many places. And suddenly slowing down is extra unsafe when there's a car right behind you that's also being blinded.

Slowing down on a highway is required by law in many areas, in the presence of emergency vehicles.

Yes, by 10-20mph. Not down to <40mph. I thought it was a smart idea when I first saw signs about it on a road trip many years ago, and incorporated it into my own driving before the law had been widely adopted. I aim to do most of my driving at less busy times, so I often end up moving over a lane even if I'm not in the rightmost lane. And actually it's a good idea to do the same for all stopped vehicles, regardless of lights - amateurs are even more likely to make mistakes.

I see flashy lights on the highway and I can’t tell for sure what’s going on I for sure slow down <40 mph because for all you know they could be up to anything that requires you to come to a complete stop. Especially if the blinding lights are especially blinding.

So just keep on trucking along at full speed with your eyes closed, you reckon.

Spoken like a true Tesla engineer.

I specifically laid out what I do, and what I believe most other people do, as humans having to cope with increased danger. I'm in no way trying to defend Tesla, but rather pointing out a general road hazard that needlessly exists and that perhaps isn't being handled gracefully by automation. This idea that a vehicle can always fall back to stopping comes from the same mistaken place as asserting that a human driver can take over at a moment's notice.

Well there's really only two options for a self driving car that has lost situational awareness - (1) keep going at the same speed, or (2) slow down and alert the driver (and possibly stop if the driver doesn't respond).

Choeger was suggesting that (2) was the prudent option. You replied and appeared to be suggesting the first option because slowing down was dangerous. But I contend that (1) is the more dangerous option if situational awareness has been lost.

Like handing control back to the human, it depends on the timescale. Over longer periods, yes, obviously the only sane option is to stop. But if confidence in sensors or analysis suddenly drops, especially when the system is still under development and unknowns are expected, the response needs to be graceful rather than sudden.

This is reflected in their comment generally, but talks past the context of my comment. In the scenario I'm envisioning, most of the view is fine right up until say the last ten seconds, where the dynamic range of the camera is suddenly overloaded. That's not enough time to talk about stopping or handing off control, and slowing down isn't going to help with the automation having lost the lane. A human gets through the situation by making a plan ahead of time, including taking note of how a road curves and guessing that the lane will remain clear. Whereas I can imagine automation responding in a horribly wrong way.

I wonder if it would make sense to hand over control to the human, or at least slow down a lot, if the car detects any emergency vehicles with lights flashing.

For emergency vehicles parked on the side of the road, there could be all sorts of reasons why you should slow down and move to the other lane. Blasting through 3ft from the vehicle at 80mph isn't the smartest move.

I've been in situations where I've had to stop suddenly because a emergency vehicle was coming across my path at a cross road, where the traffic lights were green for me. But as I heard sirens and saw the reflection of blue lights, I knew to slow down well before I saw the vehicle.

> I wonder if it would make sense to hand over control to the human

Handing over control to the driver having a nap seems like a bad idea. Slowing down certainly makes sense, I for one have an inbuilt instinct, when I see flashing lights and a parked police car, I tend to ease off a bit and prepare to stop.

There are also issues with other makers more simple automatic emergency braking systems. But overall it seems like the systems are a net improvement in avoiding at least some accidents.

https://www.usatoday.com/story/money/cars/2019/10/29/iihs-au...

Ive a Jeep Grand Cherokee. It has the front sensing collision detector and auto braking. It may have hit the brakes 1 time that I wasnt paying attention and 9,999,999 times when there was nothing in front of me. I disable it and thus far have avoided crashing. The point is that driver assist systems are only assist and you have to pay attention when you drive. Ive said this here before, Autopilot is a terrible misnomer. It seems to sit squarely in the valley of being good enough to 'trust' but not good enough for real world use. If I were in charge of the world, it would be rebranded immediately.

Dunno, autopilot (at least originally) was just a stupid switch on a plane that would try to hold altitude and speed. It wouldn't try to avoid objects, couldn't take off, couldn't land.

Very similar to Tesla's auto pilot, that can't manage your driveway, random surface streets, but can mostly (but not completely) handle highways.

Good perspective.

I think, as you noted, that 'autopilot' is probably fine, but 'full self driving' is actively misleading.

Note that I use FSD for 90+% of all of my driving, and I think I'm safer because of it, but it has to be used with understanding and care. The very name 'full self driving' is something it might be some day, but it's nowhere close to that now.

I read in another HN thread, that some Tesla owners have experienced "ghost braking," and have adopted the technique of holding their foot just above the accelerator, in order to counteract it.

When I read that, I made a mental note to slow down a bit if a Tesla gets behind me, so they will pass me. Usually that's not a problem, as I don't drive particularly fast.

I don't know if this is an issue with other brands of cars. My fancy new Subaru has never ghost-braked.

I’ve experienced this in a Tesla. It slammed the brakes on without warning in the middle of a freeway and dropped our speed from 70 to 30 in seconds. Luckily there wasn’t a truck behind us otherwise we would have all died. Tesla fans tend to waive stories like that away with “yeah it happens sometimes” which I think is a pretty astonishing response to what really felt like attempted murder by a car.

I'd much prefer a system that stopped my vehicle as smoothly as possible to one that I would rely on to autonomously brake for me.

I suspect that if all cars were fully autonomous, that we would have fewer collisions and fewer high-speed collisions simply because we wouldn't need to use traffic signals to push traffic along (and we'd have fewer traffic-stopping accidents).

A few months back, an older gentleman collided with my car on US 75 in Dallas while I was at a full stop. Would automatic braking have slowed him? Probably. Would he have collided with me at all if his car were driving autonomously? Probably not. He was distracted and took his eyes off the road, never seeing that the traffic ahead of him was at a complete stop. It took nearly 10 minutes for us to get off the road for a fender bender because nobody would allow us to get over. Traffic simply flowed around us despite the issue.

I suspect that within the next 50 years, driving at all in most major cities will require a special license, and it will become a civil rights issue for both those who don't know how to behave like adults but want the freedom to move about anyway, and those who simply do not want to be tracked by any entity.

This is the fundamental problem with artificial "intelligence" driving: It works fantastic 99.9999% of the time. But the 0.0001% of the time when doesn't work will kill you.

There's no way it's anywhere close to 99.9999% right now.

You're right

Trying not to be a Tesl fanboy, but they have a nice graph showing the number of auto pilot complaints but fail to mention if they normalized for the number of cars on the road (with autopilot) of that brand. Maybe there are a ton of Tesla complaints because they have way more cars on the road? What's the percentage? I guess we'll never know. Not great reporting.

Not sure about the US but I can’t believe there’s more Tesla’s than Audi’s on the road in the U.S?

Edit: oh there is, that’s quite cool honestly. https://www.goodcarbadcar.net/2021-us-vehicle-sales-figures-...

As far as I know all new model luxury cars like Audi and Mercedes come with automatic braking for frontal impacts. So considering the number of Mercedes sold in the last few years, I think it's safe to say there are more of them than Teslas.

Automatic braking has been a common feature even in non-luxury vehicles since at least 2018.

Yeah, once the Toyota Corolla has something, you can pretty much assume it's just standard now.

FTA: “The rate of complaints about Tesla, relative to the number of its cars sold in the U.S. in 2020, was more than three times that of the other automakers.”

Still a bad comparison, but I would think that the number relative to the number of cars in use of each brand wouldn’t be better for Tesla.

The real question would be how this would be relative to the number of miles driven (which, I guess, is about as fair as we could get)

The Tesla NTHSA complaints are kind of a mess and why we can't have nice things (data).

Many are valid and from justifiably disgruntled owners, and I could see how Tesla would have the most driver assistance related complaints, but there are subset that are suspect and/or cranky.

https://www.nhtsa.gov/vehicle/2020/TESLA/MODEL%2525203/4%252...

Several begin with "The contact ", which I guess is someone filing a complaint based on second hand info?

There's also Keef, who is still filing complaints based off of insurance auction listings and has now started filing complaints based on FSD Beta videos.

THIS IS A COMPLAINT ABOUT THE BETA RELEASE AUTOPILOT. A RANDOM MODEL 3 VIN HAS BEEN USED TO BE ABLE TO FILE A COMPLAINT. THIS COMPLAINT APPLIES TO ALL MODELS OF TESLA. HERE IS A VIDEO OF A PROUD BETA TESTER. HTTPS://WWW.YOUTUBE.COM/WATCH?V=PQ-RGY1V8UG THIS OWNER HAS A MISPLACED TRUST IN THE SAFETY OF THE AUTOPILOT. HE IS USING IT IN HIGH RAFFIC AREAS WITH PEDESTRIANS AND CYCLISTS HE IS ALSO USING IT AT EXCESSIVE SPEED ON NARROW ROADS ONLY INCHES AWAY FROM ONCOMING TRAFFIC. RATHER THAN HAVING HIS HANDS HOLDING THE WHEEL IN THE CORRECT WAY HE ONLY HAS HIS FINGERS TOUCHING THE WHEEL WITH THE PALMS FACING THE WRONG WAY. IF ANYTHING GOES WRONG HE WILL NOT BE ABLE GRAB THE WHEEL PROPERLY IN TIME TO PREVENT A COLLISION. FINGERTIP DRIVING IS UNSAFE AT THE BEST OF TIMES. IF AUTOPILOT HAS A GLITCH OR A SPASM THE DRIVER WOULD BE UNABLE TO QUICKLY REGAIN CONTROL. DESPITE THE CLAIMS OF ELON MUSK SUCH DRIVER BEHAVIOR DOES NOT LEAD TO THE BEAT RELEASE AUTOPILOT BEING REVOKED. THIS EXPERIMENT IS DEADLY. PLEASE GET IT STOPPED IMMEDIATELY. THANK YOU KEEF.

After several officers where nearly injured/injured/killed because drivers ran into them as they were stopped on the side of the road, some states instituted laws that say you either slow down to 45 or you move over a lane.

Is that built into AutoPilot and whatever it's called on other brands? Humans have to do it, so cars should have to as well.

Every lane change is a possible collision. I wonder if that was taken into account.

Considering how far away flashing lights can be sensed (in both humans and cars), that shouldn't be a problem.

Can autopilot even change lanes? It seems like that would be impossible to code for freeways in cities in California. You have to muscle your way into the lane and force people to back down or get hit or else they aren't ever letting you in.

Except it's not. Every lane change in semi-dense and dense traffic is a possible collision; I'd say if you have a car moving in the same direction as you are, and it's closer than ~50 feet, then it's a non-negligible possibility, other than that you just have nobody to collide with during lane change . We're talking highway speeds, of course, not racetrack

Yes, that's why slowing down is the other action you can take.

> I wonder if that was taken into account.

It was, but as the moral accounting stands, it is better for ten 'civilians' to be hurt, than for one police officer.

If all lanes are full, in practice you don't need to, and shouldn't, evacuate the right lane. If there's space for you to change lanes safely, then do so.

A couple weeks ago Tesla pushed an update that causes autopilot to brake and alert the driver when emergency vehicle lights are detected.[1]

1. https://twitter.com/Analytic_ETH/status/1440537502272536581

If a human driver hits a pulled-over police car, or any other stationary object that can be seen well in advance, the problem isn't simply lack of braking.

Why would that be the talking point if an autopilot does that? Because we expect it not to be aware of the obstruction until within emergency braking distance?

Wouldn't be surprised if this is their radar whitelist blinding it to sensor input.

Quite possible. They've physically removed the radar hardware on the newer models.

Musk is very adamant that cameras are enough even going so far as to stop including radar in Model 3 and Y units shipped to the US. So even if it were the radar being confused according to the Tesla line it shouldn't matter the cameras should be enough.

https://techcrunch.com/2021/05/25/tesla-is-no-longer-using-r...

We have two Teslas and they are great but I wish they used LIDAR to support the CV powered AutoPilot.

AEB above 55mph is only designed to slow down, to lessen the impact, not to stop.

giving new meaning to 'move fast and break things'

> Tesla does not have a media relations department and Chief Executive Elon Musk did not respond to attempts to seek comment.

Tesla has ongoing response to this.

https://www.tesla.com/VehicleSafetyReport

Somebody said it is a hard problem to solve.

Has journalism been this lousy forever or is it just recently? Quoting numbers of complaints per manufacturer is meaningless.

How many cars with the given feature turned on exist from each manufacturer?

How many miles driven with the given feature on?

What speed were these cars at?

That’s just from top of my head.

Oh wait? Complaints? How is that a measure of safety? Let me add some more:

How many people were injured/killed?

What was the cost of repair from the accidents?

Damn it — fell for this again.

All that matters is whether you are more likely to get into an accident/get injured in a Tesla vs other cars in the same price bracket. That's it! All these slice and dice reporting is of little value.

OT: I wonder why sites like the LA Times don’t offer a redirect to Apple News when the content is hosted there. I do subscribe to Apple News+ along with additional iCloud storage, Music, etc. I assume Apple gives LA Times something when I read the article vs. when I click away because I’m not willing to subscribe to more news sites than I already am.

Apple News link for those in a similar boat: https://apple.news/A9PXJ2JwHRouQSYO9lQh9tA

It is not as convenient as a redirect, but on iPhone and iPad you can use the "Share" function when on the page in Safari and select "News" as what to share with and that usually works.

I don't know how to do that on a Mac.

Thank you for the tip!

Maybe I’m making too many assumptions about how much Apple pays news sources when someone reads a news article. I’d have thought there’d be an alignment of financial interest in sources like LA Times to facilitating some sort of “read this article on X”, and for Apple in boosting value of News+. I wish there were more of a happy path to supporting journalism without needing to have 50 individual subscriptions.

I guess the moral hazard in a “pay per article” model would be a system that rewards clickbait (though I guess that’s the same as status quo with advertising)

I’ve been dinged several times for sharing Apple News links, because they can’t be resolved unless you’re using a recent Apple OS.

I’ve stopped using Apple News.

yet another tesla hit piece from russ mitchell and the la times.

is it just me or does the la times drag tesla at every opportunity? could it be because the owner of the la times, patrick soon shiong, has a company called nantenergy that seeks to compete with tesla? at a minimum, there’s a conflict of interest.

and, kind of a tangent, but where are the articles about the thousands of lives lost due to the incompatibility of trucks/suvs and sedans in crashes actually causing thousands of lives lost per year?

I'm pretty neutral on Tesla/Musk, but I looked at this guy's (Russ Mitchell) writing history and it appears to be an unhealthy fixation on Tesla. His Twitter, and almost every article he writes, is about bashing Tesla. It's like he has real personal beef with Musk, a person he's probably never met.

He's a journalist, they typically get assigned to cover certain topics. Not surprising that he focuses on the EV space, that's probably in his job description for LA times.

focusing on the ev space != relentlessly bashing one ev company

They are kind of the big cheese of the EV space and you could argue people relentlessly bash Tesla here too. Maybe there is something to these critiques of the company, rather than everyone and anyone just having an axe to grind.

Once again an article with NO context.

TACC and ebreaking are in MANY auto's. We get no data on the number of accidents and fatalities for cars

1) Driven by humans

2) Other systems of ebreaking

For all we know, despite these 17 accidents, teslas may be far safer.

For example, a quick google shows that there are 1.7 MILLION rear end collisions EACH YEAR in the US.

https://www.washingtonpost.com/news/dr-gridlock/wp/2015/06/0...

Note that the 17 items discussed here with Tesla are over multiple years.

So in the time teslas had these 17 crashes there were perhaps 3M+ rear end collisions alone?

The sad reality is that the clickbait headlines sell, and any effort to actually look at data does not (on the web and more recently on HN). Folks just do not want to engage with an idea or with numbers, but the outrage (or downvote) button is usually close by :)

Reply on news.ycombinator.com