Proof that this system lacks night vision. No different to the video of the deer getting run over by another Tesla on autopilot [0] and the near miss that happened when the system malfunctioned and tried to hit another car [1] - both cases at night.
Sounds like the risks in using both either autopilot or FSD systems at night is even more dangerous but very surprising that these systems still have no night-vision system.
Active sensing like Lidar is even better than night vision. The side view of a damaged truck is really hard to train as a relevant class reliably, even for Karpathy.
Even if... Tesla (like many others) suppressed Radar standstill targets to prioritize driving over stopping. Without another modality, they are indistinguishable from ground clutter in some scenes and who would want to full stop for a manhole cover every now and then... Elon maybe thought, getting rid of radar might solve that problem. It doesn't.
It seems like it does use the radar. The generation after 2.5 (3.0) did appear to improve the "frames per second" processing from ~110-200/sec to ~2300/sec.
I'm guessing the comment means something like "if this was related to sensor processing throughput or latency, the 2.5 version might be meaningful".
Not only are they unpredictable humans in general, but truckers are often the most sleep-deprived inattentive drivers on the road.
Having been personally involved in a collision with one where my vehicle ended up under their trailer at highway speeds, when the trucker abruptly swerved into my lane to avoid rear-ending another truck they were tail-gating without paying enough attention to brake in time, I find this thread of discussion patently absurd.
People don't realise how much "swerving to get out of the way of your own mistakes" costs the person you're swerving into.
It's so sad to see it, but then they claim "oh but it was empty". People use it as their resort when following too closely, I wish they'd just accept their mistake and crash into the back of the car in front of them. But that would be too clean and logical.
Watching the video I'm not sure I would have been able to stop in time. There's like a quarter second between first seeing the truck and slamming into it.
Human eyes have much higher dynamic range than typical video cameras. What the video shows is not likely representative of what your eyes would have seen.
Watching the video a few times and knowing after the fact the first frame in which you can see something there, there's 4+ seconds to react. From the first indication of something being wrong (oncoming traffic headlamp occlusion) there's even more time (at least 7s). A sober, awake driver would have had "much higher dynamic range" as you put it, not only in their eyesight, but also in their acuity, and would have been able to come to a full stop without hitting the truck at all, much less hitting it at high speed.
With respect, while I broadly agree that a good driver could significantly mitigate the impact, I think you're looking at this with too much hindsight.
1. Even a sober, awake driver glances down to confirm their speedometer and scan their fuel gauge, check engine light, tire pressure, etc.
2. A sober, awake driver should be scanning the road along and beside the projected vehicle path. But they doesn't necessarily mean the would be looking in the direction of the occluded headlights at the very first instant they tell of a crashed car.
3. A sober, awake driver has to process what they see.
4. A sober, awake driver still has a good 0.2 to 0.25 second reaction time.
5. I actually do look for occluded lights when driving at night, but my experience talking about this with others in person is that this is a totally foreign concept.
In other words, you know where and how and when to look for the obstruction because you've watched the video a few times.
This is sort of like the Air France 447 flight or the Boeing 737 Max trim problem. If someone tells you exactly what kind of scenario you're going to face, how to handle it and approximately when you'll face it: you can win every time. 100% of pilots who were briefed on the specifics of AF447 avoided crashing the sim. Same for the 737 Max.
Yet, hundreds of people are dead because pilots in actual cockpits, who were not given prior notification failed to act correctly.
I think it's pretty likely that a human driver would have collided with the truck.
A major difference is that (from watching the uncompressed footage), a human driver would likely have stomped on the breaks before impact. That last minute breaking could make a world of difference.
A collision at 75 mph has a lot more energy (4.5x ish) than a collision at 35 mph. So every 1/10s of breaking will do a lot to reduce the damage of the impact.
That's one of my concerns about driverless vehicles that I want to see the statistics on. I have a hunch that automation will reduce the absolute number of impacts, but may lead the impacts that remain could be higher energy impacts, because they result from a classification error.
dodging sudden obstacle is not something one usually does with higher level reasoning. if you fullscreen the better quality version from streamable and immerse in the situation, while the truck is hard to spot, brain can definitely see something is occluding road painting and other cars light.
in person one might not realize a truck is there, but brain would probably think "something not right" and alert senses maybe even reflexively release pressure on the accellerator, so that when higher level consciousness makes out the truck, body is primed to act.
I’m reasonably sure I would have avoided that based on the uncompressed, from pretty early on you can see “somethings not right”.
Wife works in automotive engineering and I asked her why FCW and AEB didn’t detect it and her only thought is that because the truck is on the same angle as the central reservation and already in a bit of state it might not have detected it, although she thinks that more likely something just went wrong
If I'm being honest with myself, I would have hit the truck going just as fast and not noticed until it was far too late. The first sign of a problem was the skid marks, but but then the collision was unavoidable. I tried going back and looking to see if the oncoming headlights were obscured by the body of the truck, but I wasn't able to discern that.
I also thought I might have been able to take the 18 wheeler's reaction as a clue, but it passed the wreck in the other lane and only starting braking right as it passed it.
I think the issue with watching it like this is that you’re not currently driving so not attuned to what you’re looking for. I guess the issue with autopilot is that the driver likewise isn’t attunded to driving as they are on autopilot.
With that in mind, in this exact scenario, I would have hit it also. Only with full manual control of the car do I think I could have avoided it.
> The first sign of a problem was the skid marks, but but then the collision was unavoidable.
I'm pretty sure I would have hit the truck too[1], but there actually was an earlier clue: Headlights from oncoming traffic are occluded by the two crashed vehicles ahead.
When rewatching the video (so I know both that I should be looking for something and approximately where that something is), the occluded headlights are pretty obvious. An attentive driver may have noticed the vehicles, but I think most people saying the truck was obvious are deluding themselves.
[1] The first time I watched the video, my reaction was "Where is this tru-OHSHIT".
> I’m reasonably sure I would have avoided that based on the uncompressed
I'm not so sure I would have.
If I'm driving at 120km/h , at night, chances are I'm also tiered and reacting that quickly at that speed with limited vision would have been a coin toss for me.
I'm all for making these cars safer but we are watching that video, knowing the truck would be there but fail to consider all the extra context around it. I think accidents can and will happen with self driving cars, no matter what we do.
Thanks for posting! Having watched this version there's no way any human driver paying attention would hit that truck. It would have been more visible in actuality as the dynamic range of video doesn't compare to eyesight.
Not saying I wouldn't have crashed, but it's way more visible in this video. It's not even close. Definitely would have at least tried to brake, whereas in the compressed video the truck seems to just pop into existence out of the darkness.
big big difference, the first video i thought to myself "yea i probably would have hit it too", this video shows me that i could have definitely slammed the brakes to have minimal impact or maybe even stop entirely.
Wow that is a huge difference, thank you for sharing. I am fucking against this auto pilot stuff a good deal and even I was looking at the video in the original link thinking "Fuck I dunno if a human would do any better".
In the uncompressed video it seems very obvious that there is an obstacle in the road much earlier
The other question to ask is what was the Tesla doing bombing along in the outer lane when there was (seemingly) negligible traffic anywhere except the truck in the inner lane.
Whilst its unlikely the Tesla autopilot would have fared much better in the centre lane, the impact could have been substantially less.
Did Tesla respond to this? I am curious if this an algorithm issue or an issue with real time scheduling where the necessary signal did not arrive in time and so the whole system failed to react.
The truck still blocks the view of headlights of the traffic in the other lane. It also blocked view of the lights of the right-lane truck near the end. I don't know what the state of the art is like, but humans can tell something is "off" up ahead when they see this kind of "disappearing headlights" pattern.
In any case, the tweet claims there was no braking of any kind, and I'd expect braking at least in the split second before the crash, where the truck is visible to the camera.
I'm reminded of the time the Uber SDC hit the pedestrian. Uber initally released one camera's footage, where the ped wasn't visible until it was too late.[1] But then others drove the same path in the same lighting conditions and, even with their (unspecialized) dashcams it was clear you had plenty of time to see someone in the street -- and even that much is ignoring how humans see better at night than the dashcam would!
(Oddly, I can't seem to find any such videos at the moment or know what terms I should be using to search, but I definitely remember people uploading such videos to YouTube.)
That's part of the issue - Elon says that LIDAR is "a deadend" and removed radar from Teslas as being "unnecessary", because their cameras "could do everything needed".
Either one of those would have detected this.
My Audi's radar works great in pea soup fog. Has prevented me from nearly rear-ending a gray vehicle with no lights on in heavy fog.
Removing LIDAR could still be the right call even if it means accidents like this still happen or even happen more often.
If removing LIDAR lowers the cost of the vehicles which means more of them are sold which results in a fewer accidents and fatalities overall, then I think it's the right call.
> more of them are sold which results in a fewer accidents and fatalities overall
This argument needs to be fleshed out more after a video like this surfaces. I can't accept at face value that this style of autonomous car, which has demonstrated issues running into parked and stationary objects, is the platform which when mass distributed will reduce overall accidents and fatalities.
It's clear from this video that this vehicle doesn't know when it's blind, and will merrily carry on at 75 mph even though it has no idea what's going on around it. If you are driving in the dark or the fog and you can't see 12 seconds in front of you, you slow down until you can. Not Tesla apparently -- it will continue on at 75mph even though it can't break in time if something emerges in front of it. How is this the technological basis of a something that promises to lower overall fatalities? How does this get better with scale? Seems to me, based on this repeatedly demonstrated mode of failure, scale will only reveal how broken their self driving system is.
Why is it accepted here that filling the road with autonomous cars that don't know when they're blind and run into parked objects will somehow result in a net positive outcome?
Moreover, how is Tesla in particular the company which is going to deliver these safer vehicles that make our roads safer? Tesla has a history and culture of deception in their marketing. They haven't taken these accidents seriously in their engineering and product design. Teslas have been running into stationary objects for 5+ years now and they show no sign of fixing that. Call me crazy but if I were shipping a driverless car technology, I'd want to make sure it didn't run into things. Maybe that's just me?
Tesla recklessly released this flawed technology and product to actual consumers for use on actual roads. How are these the actions of a safety-focused company? From my perspective, if you flood the roads with Teslas, you're not going to get safer roads, you're going to get more Twitter threads with videos like this. The claim that this will necessarily lead to safer roads is yet to be proven.
Look at all the fatalities and less serious crashes and you should be able to calculate if Teslas are over or under represented in that data. I honestly have no idea if they are or are not.
I think you underestimate how bad human drivers are. In the US, something like 100 people per day are killed in car crashes. Depending on how many Teslas are on the road, seeing a video like this every day may not be that surprising. If bugs in self driving tech kills 10,000 people per year (30 a day) but the overall fatality rate drops from 35k to 25k per year, it could be seen as a win.
I get it. I understand that argument and it is a good utilitarian argument. What I'm saying is Tesla has proven they aren't the company to bring about that future.
Tesla as an organization has demonstrated a level of corporate sociopathy that tells me they don't take safety seriously as a culture. Surely, they tout safety as a selling point but their actions tell me they are far more interested in getting driverless cars on the street quickly than doing so safely.
You can't achieve sweeping victories in safety without intentionality. Tesla may claim that safety is a priority for them, but if that were true they wouldn't lie to their customers about the capabilities of their own products. They wouldn't eschew industry best practices as far has hardware is concerned. They wouldn't beta test unproven software on their customers and charge them for it, leading to bodily harm of their customers and others. They would fully investigate problems as they occur and definitively fix them instead of allowing concerning issues to persist for half a decade now (running into large stationary objects). There are many organizations working on driverless cars, and yet Tesla is the only one that behaves this way.
I really hope we as a community have grown past the idea that simply throwing technology into the world will have positive net benefits as long as there are good or good-adjacent intentions behind them. In fact, it takes planning and execution to ensure good outcomes. Simply doing something that can be cast as a boon for safety does not guarantee a boon for safety if safety isn't a top goal. And by top I mean it has to be #1 or #2, because any lower and other priorities eat into it.
> more of them are sold which results in a fewer accidents and fatalities overall, then I think it's the right call.
Except that every bit of available evidence shows that not only do self driving cars not reduce fatalities, there is no reason to expect that they will.
People are surprisingly good drivers - the reason there are so many accidents is simply that people drive so much. I ran some numbers many years ago showing this: https://news.ycombinator.com/item?id=12012166 check them if you like.
I think folks here are misunderstanding what I am trying to say above.
The system seems it, at least the sensors will and then they send the image to the module that processes this and runs the algorithm to detect that its a stationary object and that it needs to hit the brakes. Then the signal needs to be sent to the module that brakes. Being a real time system, all of these need to happen within a time window for the brakes to be applied at the right time. Its possible that because of some resource constraints or poor programming/design one of the steps in the chain is not triggered and so the system did not brake.
As a human, I would probably crash into it, depending on what I seen from non-compressed and compressed version. However, camera works differently from human eyes.
Idk. The non-compressed version makes it seem like the truck was very visible. I think any driver paying attention would have at least applied some braking.
Two issues that come to mind are (1) keeping the lights on while otherwise shutting off the vehicle and (b) ensuring that the lights don't all get smashed to bits in the accident.
Yes, and your human eyes have far better dynamic range than these cameras. Plus, I could tell there was something in the road early on despite the truck not being illuminated, because I saw the oncoming headlines disappearing behind an obstruction.
Really shows how Tesla has probably made a big mistake going all-in on camera only vision. And their vision system couldn't make the inference of an obstruction based on the oncoming headlights.
How did they even justify the use of a camera-only sensor system for low-light conditions? At least the radar sensors had some complementarity with visible light sensors but they decided to get rid of those too.
To me the biggest issue is that there doesn't seem to have been any braking attempt, even at very close distance. Even if it was imperfect or hard to see at a distance, I would expect to see the 75mph reduced even slightly before impact. Every speed reduction is a statistical increase in surviving a crash.
You have to remember that you have a much greater "dynamic range" of vision as a human. That means that you can see much darker and brighter things at the same time.
The thing that would(or should) have alerted you as a driver is the reflecting of the truck's lights on the paint of the broken down pickup as it drove past. Although why the truck driver didn't slowdown as it passed a crash is a bit of a mystery.
However, Teslas have radar for this sort of things. On the other hand this does kinda underscore the problem of using cameras with global shutter control for self driving. The auto contrast couldnt cope with the oncoming lights, so stopped down the sensitivity of the camera, leaving it blind to that broken truck.
What is the practical solution for cameras only besides sensors with more dynamic range? Secondary cameras with faster apertures and higher ISOs(or different exposure settings)? Would that also help by allowing triangulation for distances and minimizing false positives for stationary objects? It is more of a technology or cost problem?
Tesla already have limited number of sensors, and they do most of the driving using monocular(single camera, or single viewpoint with three different zoom levels) inference. This is absolute madness, given the state of current tech.
For that night situation, I would be pushing for a different camera type. One that doesn't have a global exposure setting. This means that groups of pixels can be lumped together to have local ISO settings(its more complex than that, and involves integrating processing silicon directly under the sensor, on the same die.). This is a way to get vastly greater dynamic range in a situation where you have lots of point lights (ie headlights)
This would require developing and qualifying a new camera sensor from scratch (I don't think generic versions of this type of camera exist yet) Musk is unlikely to do this as he has signaled that everything should be done via ML.
Second, using a different wavelength, IR, or near IR.
Thirdly, higher wavelength radar. It gives you this information for free, they are cheap and already qualified for vehicles.
I was driving from Bay Area to a nearby city at night. Lighting was similar. Was driving at about 70-75mph. Saw a flash in the far distance. Had no idea a crash had happened. Out of extreme caution slowed down to 60mph. In about a minute had a truck right in my lane just like this video. Swerved to another lane to avoid crash. The only reason we survived was that we had slowed down based on a random flash.
Another thing I don’t understand is why wont Tesla invest in a front only radar sensor to avoid such collisions.
op is slightly confused on terminology. It does not have lasers\lidar but some models do have sensors such as the ars4-b. Other models use cameras only
Elon Musk is adamant about but needing such systems because "after all, humans are fine without them as well". I don't know what systems the current hardware has, but am sure there's no LIDAR.
Devil's Advocate: You already knew what to look for* when you watched the video. I would say that a typical driver in the same scenario certainly wouldn't have had nearly as much of three seconds between noticing and collision.
Edit: *Specifically, to be on the lookout for a black sideways truck on an unlit road directly in ahead of where the car is traveling.
If a kid running out to get their ball is a the standard example of a hazard that drivers should always be expecting on a residential road, I think a broken down vehicle is the standard example for a highway.
Maybe a human could have braked even a little bit, but the autopilot didn’t brake at all. That’s a problem for advocates for it
Because the car had radar. There are plenty of other car brands that ship similar emergency breaking features that wouldn't have missed this. Also, I, as a human viewing a video with crappy color balance (much inferior to human night vision), noticed the car long before it was struck. At least I would have begun breaking before hitting the car.
My emergency responder friend added this material to the bottom of his family vehicle as well as to the side.
He has responded to so many night crashes where they have had trouble locating the vehicle, that he wanted to make the next emergency responder’s job easier by helping them see the flipped vehicle if that were ever to (unfortunately) occur.
Keep in mind the human eye is very good at detecting low level lighting, better than a camera. Some studies say it can even detect a single photon. It was likely easier to see in person.
Single photon detecting eyes or not, most of us drive and know how hard that scenario would be to prevent.
I doubt I would have seen it or reacted in time.
There was a human in the Tesla that apparently didn't see it either.
- edit -
Sure... maybe a human could have "lessened the blow", maybe some humans, but now we're hindsight backseat driving and pulling things out of our ass.
As a human there's a lot that goes into it, it depends on how long you were driving, your vision at the time, whether you were talking, your driving skills, your reaction time, your handling of accidents, knowing where you can swerve to.
Some humans with full attention may have prevented it or weakened the crash, but that's a maybe, and assuming full attention.
Many people in this thread who think they would have noticed the subtle lighting effects on a pitch black pickup in the middle of the road. I've nearly hit blacked out cars on the road, you don't see them. You could go back and probably see it in a dashcam, but you know it's there then.
The uncompressed version you're still expecting a truck, it seems like it's much slower than what rounding a corner going 75mph on a freeway is, I don't think any of us could judge without being in the seat at the time.
All in all, this is the most speculative thread I've ever seen on HN. Very reaching, but anti-Tesla is hot news here so...
The big takeaway for me is how that Tesla just took that head-on collision at 75mph without compromising the cabin.
Don’t be so sure. The lines on the road have a obvious break where the truck is. You might have noticed that and slowed down as a result, even slowed down enough to be able to notice the truck in time.
you would have seen the light reflecting off the wreck when the truck drove past it. Because the light would be moving and in a place you wouldn't expect it, you would have been alerted.
You might not have been able recognise it was a pickup, but you would have at least braked earlier.
You may not have reacted in time to prevent a collision, but you very likely certainly would have stomped on the breaks before the impact.
Changing a 75mph collision into a 35mph collision is a world of difference in terms of the total energy.
I think it's important not to view "collision" as a binary event. With human drivers, it's pretty rare not to heal the squeal of breaks just before an impact. I suspect with automated driving (at least for a while) these kinds of "full speed" impacts will be more common than with human drivers.
It would have been difficult, but contrast is a lot different in real life. This video is pretty LDR, and the truck would have been visible at approximately 0:05, giving about 3 seconds warning, which is a lot when you are driving on a highway.
You're misunderstanding what people are saying. You've watched a highly compressed video from a camera sensor that is known to have relatively poor dynamic range (ie: it doesn't see well in the dark when lights are around).
The point OP was making (that you completely dismissed) was that our eyes are actually quite good at sensing this scenario. Here's a better version: https://twitter.com/haenschengross/status/147332613675511808...
When you watch the uncompressed version, there are tons of visual cues to a human that there is an object in the way. This is no where near as subtle as the janky video makes it out. You can say you might have missed this at night but either you don't pay enough attention when you're driving, or you're giving tesla a pass here they don't deserve.
I think it is kind of unfair to apply your own intuition without any tests. Perhaps the average human will see there is something weird in front without being conscious of it (e.g. the lack of the white lines etc.) that is you would see it without seeing it. Or at least enough to slow down a little and be more cautious going forward. It is pretty hard to know this without doing tests.
But from that video this looks like the kind of thing an attentive driver could easily miss until it was too late. Maybe they’d start breaking but would they have stopped anywhere near in time? I doubt it.
We don’t know unless we test it. But an attentive drive might at least increase their alertness when the stop seeing the headlights of the oncoming traffic, step off their gas pedal when the lane lines starts to disappear, and apply a significant amount of breaking when they finally notice the car, making the accident at least not as bad.
> Wow. I don’t know how anything without night vision wouldn’t have hit that. It’s practically invisible.
Radar and Lidar work in the dark. But yes, this is a problem with optical-only, especially since the camera-systems Tesla uses are pretty badly spec'd (1.2MP, not even 1080p)
Tesla _used_ to have a RADAR system installed. But stopped installing them and rewrote the software to get rid of it earlier this year.
Your iPhone 13 has a 1/1.9" wide sensor, significantly larger... and multiples of them working as a team.
Even if we go back to 2018: the Google Pixel 2 has a 1/2.6" sensor, still much larger (and with better resolution) than what this Tesla is using to see the road.
----------
If you want better night vision, then use a bigger sensor. Its just basic physics. A sensor with 2x the area will capture 2x the photons.
Complaining that this car can't see the road at night... well.... yeah. Its got an tiny camera sensor. What did you expect? Its got bad night vision.
ADAS must use qualified chips and that one was a pretty standard one some years ago. I like others better, but you must see them in system context and be able to continuously process that stream. Given the power budget and other constraints this was what a number of manufacturers ended up with. After all these guys are not dumb, they just need to desing in many more aspects than mobile phones.
I hope so, I'm in that field since 2003, started with the first night vision system in series production and built an number of those systems since.
One of the problems is tenperature range. Driving continuiusly in Death Valley can heat up your camera to the point it runs away into thermal self destruction, if you don't take counter measures. This is not a problem for mobile phones.
Another issue is continuously processing uncompressed data. Mobile phones compress with hardware support, especially in video mode. I had one of my mobiles shut down during a long video conference. This must not happen to an assistance system.
Couldn’t an infrared camera have helped? black paint is not anyways black in infrared, and the recently-crashed car would likely have a heat signature. I wonder if they would have to start from scratch with their training data, or if black and white could be used for everything.
The RADAR that Tesla used to ship with (and older Teslas used to use before FSD12 disabled it) would have almost certainly seen the truck in this specific case.
The issue with RADAR was allegedly phantom braking, if Tesla marketing was to be believed. But other companies seem to have RADAR systems that work.
------
Its fine to use 1.2 MP 1/3" CMOS sensors as part of a general-purpose sensor network (which may include RADAR, ultrasonic, and cheap Cameras all working together). I think LIDAR is still needed to really make things work, but RADAR + Cameras seem to cover a wide variety of cases.
What's a problem is when Tesla starts removing RADAR and pretends its an advantage. (Look, they couldn't figure out the RADAR false-positive problem, and there's a chip shortage right now anyway. So they just... stopped using RADAR)
> The issue with RADAR was allegedly phantom braking, if Tesla marketing was to be believed. But other companies seem to have RADAR systems that work.
And the phantom braking hasn't improved without radar. As with so many things at Tesla, this was clearly a cost-cutting or supply-constraint inspired move masquerading as some technological advance.
Have you ever tried to take a picture on your phone at night? Or even dusk? Things that are clearly visible render as 'practically invisible.' It's a safe bet that this video isn't a good representation of how visible the thing is, and it's guaranteed that the truck was more visible than it appears. The question of how much more visible probably isn't easy to answer other than "much more."
A car’s autopilot braking in the nick of time to avoid an “invisible” obstacle is basically a perfect viral video, surely that would be a top post here and on Facebook and Reddit if it existed
That would be real magic. Magic is numerous cases where Autopilot does end up avoiding accidents in otherwise mundane situations that most people don't even think about, but over time does make a big dent in the number of fatalities.
At this point it is very clear, to me, that anyone using FSD/Autopilot is risking their lives by trusting their lives on a beta, or perhaps even alpha product.
It has issues during the day, at night it is downright dangerous.
I do agree that while watching the video the stopped truck is not visible till very late and based just on the video it is not clear if a human would have done any better. However, I'm not sure we can make that call since it's not clear what the quality/aperture of the video is and what the human eye was able to see.
Still, even then, a human would have slammed the breaks or swerved to the right to a free lane and depending on how good the Tesla breaks are, I don't own one, it may have stopped/slowed down substantially.
EDIT:
I think some people are underestimating their reaction time. I have had a case, once in my life, where I was driving on a Freeway at ~65 mph, 87 N for those from the Bay Area, and due to the curving of the Freeway I did not see the traffic stopped in front of me in the left most lane. It was very similar to this, but while the freeway was curving. I did not slam the breaks, but swerved to the right lane, which very luckily for me was empty till I could slam the breaks. It was purely reflex action driven, no thinking was involved.
In this case there are two things:
- human eye can see farther at night than a camera as others have also noted.
- swerving would/could have prevented a head on collision since there was no traffic in the middle lane (though the stopped truck is occupying a bit of it) and would have hit the Tesla on the left side. Much better than a full head on collision.
> even normal adaptive cruise control makes me very nervous
Thank you, it seems I'm not the only one. My foot hurts much more when I'm using adaptive cruise control since it's always hovering over the brake but not touching it - unless there is absolutely not traffic whatsoever. Due to this, I almost never use it.
I used adaptive cruise in a Nissan Murano and it maintained assured safe distance perfectly. First time using adaptive cruise and I found it very confidence inspiring.
> At this point it is very clear that anyone using FSD/Autopilot is risking their lives by trusting their lives on a beta, or perhaps even alpha product.
Somewhat. The product costs $10,000 and has been enabled since 2015 (originally $5000). Its reasonable for people to think that this thing works. Sure, perhaps those individuals should do better research, but there's a huge degree of misinformation coming in from the top here.
The root-cause of the problem is Tesla. Its unreasonable for Tesla to pretend it works, charge significant sums of money for the "feature", and ignore the difficult questions when this thing crashes into stopped Firetrucks over-and-over again (https://www.consumerreports.org/autonomous-driving/nhtsa-saf...)
At some point, we as a society need to accept and admit that Tesla is __lying__ about this feature. That it is dangerous, and always has been dangerous.
There's a degree of blame we can assign to the Tesla-fanboys who accept this marketing without any level of critical thinking. But the root cause of the problem is Tesla's marketing.
It isn't just Tesla that is leading to confusion here. For example, both you and the comment you are responding to treat Autopilot and Full Self Driving as interchangeable when they are distinct features.
Autopilot has been on cars for years, has driven multiple billion miles, and now comes free with every Tesla. In reality it is just lane and traffic aware cruise control which numerous cars come with today.
Full Self Driving doesn't have near the track record, is still in beta (and maybe should still be in alpha), and has that 5 figure price tag you mentioned.
When you blur the lines between the two, it also results in people thinking Autopilot is more capable than it is or that Full Self Driving is safer than it is.
Not crashing into Firetrucks is an emergency-brake assist problem. This isn't a "beta" part of Tesla's suite.
I'd rather not use the crap marketing terms of Tesla. They're clearly misleading and just propaganda. You and I know that its emergency braking.
It doesn't matter if its on an intelligent-cruise control / lane assist mode, or if its in the hands of a driver. Emergency braking is supposed to kick in and work in these cases
Your original comment was speaking very broadly and not about this specific accident. I responded broadly and not about the specific accident. Your response was only about this accident and ignored my broader point.
Yes, this accident is a problem that Tesla needs to work to fix. But no system is going to ever be flawless and I would bet many humans and other manufacturer emergency braking systems would fail to avoid this accident.
In the meantime everyone, from Tesla fanboys to Tesla critics, should be mindful about how they talk about these things because confusion is dangerous regardless of its source.
Which is Tesla crashing into stopped trucks in the middle of the road. The fact is, we have a pattern here, and it is now documented.
As sad as it is, this Twitter-thread isn't "news". The NHTSA opened an investigation last year, but you know how slowly the government works. Everyone has known this.
There is a pattern here about Tesla's emergency braking, be it on the default package, the 2015-era Autopilot for $5000, or the modern 2021-era Autopilot for $10,000.
Are you just trolling with that comment. You are making the same mistake again that I corrected you on in your first comment. Autopilot is free with the car.
Emergency braking systems are not foolproof. Tesla is not a market leader in that category, but these systems from other manufactures also fail in similar situations[1].
I don't see the distinction. Both FSD and Autopilot would have crashed into this truck at full speed.
We both know they're on the same sensor suite on largely the same technology. "FSD" may have more software and "AI" behind it, but the fundamental "is a stopped car in front of me" logic is the same neural-network training through the Dojo-supercomputer or whatever else Tesla wants to talk about. Its just a software switch that's locked behind a $10,000 paywall, nothing more.
But the customer, and the people involved in crashes, don't care about the names of supercomputers, the different features of the car, or whether or not its in Beta or not.
--------------
Can Tesla's sensor suite + computer see a stopped car in front of it? Well... no. It can't. Not only does this case prove it, so have many other cases.
Once again, I am speaking generally and not about this specific accident. We have no idea how any other person or car would perform in this exact situation and I don't see much value in speculating.
I care much less about a singular accident than I do about trying to help prevent future accidents by clearing up the confusion. You are continuing to add to that confusion, seemingly intentionally by continuing to blur the lines that I have already clarified in previous comments.
You already admitted that a product being on the market for years would lead to a consumer being more confident in it. By telling people that FSD has been tested for years on public roads rather than months your criticism is counterintuitively also contributing to a false confidence in the technology.
> Once again, I am speaking generally and not about this specific accident. We have no idea how any other person or car would perform in this exact situation and I don't see much value in speculating.
But we have an NHTSA complaint documenting 11 other accidents of this form. That is, a stopped car in front of a Tesla that was provably on Autopilot. And then the Tesla crashes into the stopped car (most commonly: emergency vehicles like Firetrucks or Police cars). EDIT: These other examples from NHTSA were in broad-daylight, with flashing lights, at night, and a variety of different situations. So something is definitely wrong here.
And who knows about all the other situations where a crash was going to happen but the driver took emergency action and saved the day.
-----
This accident, the one in the twitter post, as well as the 11 other documented cases in the NHTSA investigation, prove that there's a systemic problem with the technology.
There are probably other accidents too where the emergency braking should have worked but aren't documented yet.
------
This is no longer a Rusell's Teapot scenario. We've got 12-examples (11 from the NHTSA complaint, and +1 from this Twitter thread / Twitter video). We've got the examples, live on camera. We can absolutely say without a doubt that Teslas have difficulty in this case.
Autopilot has not been on cars for years, Adaptive cruise control has been.
Moreover, while I’ll accept that my comment above confuses between them, I’ll argue its naming it as “ Autipilot” is what is at fault here since people overestimate its capabilities.
I am neither going to defend the feature's name or get into a debate about it. I am simply pointing out that the confusion does not have one single source like you originally implied.
Autopilot may be a marketing name and people can voice their opinion about it as not being appropriate or whatever but where did the notion come from that "Autopilot" means it works on its own without human?
Autopilot is a term most commonly used in aircraft and that's where almost everyone has head from it before. At the same time almost everyone should be fully aware that aircraft on autopilot do in fact need at least one human pilot. Almost all aircraft inevitably crash if left on autopilot with no further human interaction.
So where does this idea come from that the word "autopilot" somehow misleads people into believing that it does not require a human.
Most people aren’t pilots, most people have never flown an aircraft, most people have no idea how actual autopilot works. I bet most people think that autopilot is something you turn on while in air and makes the airplane fly on course without further inputs from the pilot. I bet for most people autopilot means computer does driving.
You dont have to be pilot or even remotely interested in aircraft to know the term. Its like "Blackbox" everyone with basic education has heard about it.
>I bet for most people autopilot means computer does driving.
That is the claim but there is no evidence at all. Everyone knows there is a pilot in almost every aircraft and that autopilot is just a computer that takes work away. If it would fly alone there would be no pilot.
There should be no people with a driver license who think autopilot means no human needed. Its completely absurd. That word never meant that and its not people misinterpreting the word that sleep in their cars and crash it. These people are not mislead by a word they are just irresponsible and fully aware of it or drunk.
I guess we won’t know until we ask a random sample. Until then—and unless Tesla has conducted a survey and is not sharing—we can assume at best negligence and at worst intentional market manipulation.
I guess we assume word have meaning and are mostly understood by everyone according to the meanings or people dont understand the word at all if they never heard it.
Assuming wide spread misinterpretation is a bit far fetched especially for a word that the average teenager knows what it means.
I think people do think that for the most part, the airplane goes on autopilot, requiring minimal input from the pilot, and when everything goes smoothly, no input is required from the pilot. Input is only required when things go wrong. I think this is what people think, but I am just speculating.
> Autopilot may be a marketing name and people can voice their opinion about it as not being appropriate or whatever but where did the notion come from that "Autopilot" means it works on its own without human?
When Tesla Marketing videos said "The person in the driver's seat is only there for legal reasons. He is not doing anything. The car is driving itself".
You can find this quote all over the internet. Since this is Hacker News: https://news.ycombinator.com/item?id=26858525, you can see people discussing this and proving that the video said that.
Quote end-quote. Tesla literally said that and posted it all over the internet. Elon Musk then came on after that video was aired and said they'll have "Coast to coast Full-self driving" in 2017, and repeatedly double-downed on that claim.
I didn't say anything about Tesla, I dont care about Tesla. The word was "Autopilot" and the claim was that people interpret the word wrong. There is zero evidence that this is the case. I've never talked to anyone who doesn't know what the word means and no one ever expressed that it means "drives/flies without a human". Not even Tesla said that. The "Full Self-Driving Capability" that they promised and never delivered was not named "Autopilot" yet people keep bring up that "misleading" word as the reason why people crash cars. It makes no sense.
Because the claim that people dont understand the word comes up in such threads.
Other than that my post had nothing to do with the thread/article it was a reply to the parent post stating the claim.
That depend on the plane and all kinds of external circumstances as well as how it is programmed etc. etc.
But I can assure you even modern passenger plane on autopilot do not avoid crashes with other planes, buildings or ground (mountains etc.)
It also wont start or land a plane.
In a worst case scenario the autopilot would crash the plane immediately because it was give the task to fly into something and in best case scenario the plane flies according to the programmed route and corrects for wind and all that and if nothing is in the way it eventually runs out of fuel and crashes.
AFAIK it would not even slow down and certainly it would not do an emergency landing.
“Auto” here being short for automatic, so “automatic pilot.”
I understand your explanation of the term with regards to airplanes, but not everybody understands that. Even if the term ‘autopilot’ is technically correct, it might be better to just call it driver assist for now.
I’m not laying the blame on Tesla for these sorts of accidents though. Regardless of the name, I think Tesla makes it explicitly clear how the function should be used.
I do think there is some friction here between Tesla’s branding and marketing goals, and the long and complex road to safe autonomous driving.
That is correct and therefore the autopilot does automatic piloting which is obviously spectrum that can include more or less capabilities but certainly its not to be confuse with "autonomous piloting" or "self-driving/operating".
This is a marketing page, directly from Tesla in 2016. It is clearly misleading and/or lying in several ways, and Tesla has yet to update their direct marketing on their direct website.
Are also marketing. And this was a lie of course, "coast to coast FSD" never happened, not in 2017, and probably never. It doesn't seem like the technology is capable of seeing stopped cars in front of it (let alone of a coast-to-coast drive without disengagements).
I think you're being pedantic about inane details.
The CEO participates in marketing, on Twitter, on Facebook. Tesla markets on their website, and through their employees and service centers. There are clearly a ton of Tesla influencers who are also being paid to shill the car and their features online (much like any other car company, social media shilling strategies are very effective). There are many, many web-based marketing strategies Tesla has, from referral links, to search-engine-optimization, to blogging, to Youtube videos, etc. etc.
That Tesla doesn't participate in the traditional TV-ad marketing is a bit of a non-sequitur. They clearly have a huge online presence and leverage that to influence their buyers. If you've ever done SEO before, you'd clearly see how Tesla (and Elon Musk) optimize for SEO.
Yeah, Musk says that they don't do marketing. But since when was Musk a trustworthy source of information? Think for a bit.
at 75mph, even with a very quick reaction time of 0.7 seconds to shift your foot from gas to brake, your already approching 80', and haven't even started slowing down. Yanking the wheel sideways on most vehicles that speed means you either roll-over, or side-swipe a vehicle in the next lane, since you didn't have time to check that lane before swerving. I think you are drastically overrating your driving abilities.
Really if its that dark out and you can't see that far, on autopilot or otherwise, you shouldn't be going that fast. It's the same thing with people speeding in dense fog.
In most states a driver who runs into a stationary object is always considered at fault and can be cited for a moving violation. That rule applies even if the object is difficult to see. The driver is legally required to reduce speed based on the conditions.
Not to be pedantic but do you have a source? I actually want to know. It seems like such laws would have a reasonable exemption for unexpected objects in roadways. For example, I have a friend who successfully sued the LA City government after running into a downed light pole on a city street that was not visible due to low lighting.
Personally I consider it my responsibility, as a driver, to avoid any collisions with stationary objects. That means managing my speed and visibility accordingly.
IMO your foot should be on break if you're in FSD mode. That's where my foot is using just cruise control even. The slightest tap on the break turns off cruise control.
Also, if it's so dark that you can't stop before hitting a car in front of you, then you're going too fast. The speed limit is a max, not a min, and it changes depending on prevailing road conditions no matter what the posted signs say. For example, if it's raining and you get into an accident, you can be fined for speeding even if you were going the speed limit. Going 75mph in pea-soup fog is a recipe for disaster. The dead of night has the same calculus. You should only go as fast as you can see. A good rule of thumb is you should always be able to see 12 seconds clearly ahead of you. That's how some professional human drivers manage to drive literal millions of miles over a career without a single accident.
Frankly, all this talk about how Teslas will improve the safety of traffic because they perform at a superhuman level is belied by this video evidence. Apparently these machines only drive as well as the drivers who programmed them. I wonder if there are even any professional drivers on the engineering team consulting on this project.
If the car can't even operate within the envelope of its own sensing and computing limitations, then how can we make any claims about its safety? It's driving at top speed in situations where it's essentially blind. When I'm blinded, I pull over. When the Tesla is blinded, it doesn't even know. See also, the guy who ran full speed into a tractor trailer and lost his head. It's been years since that incident and Teslas are still driving full speed into things they can't see. How is that a basis for platform that will save us all due to how safe it is?
They're not going to roll unless it's something incredibly top heavy like a Jeep. It's a car, turning the wheel 90 degrees only turns the wheels like 10 degrees.
Also, a sideswipe is almost certainly preferable for everyone than plowing head on at full speed.
Take a guess where that truck is going to go after it gets hit at 75 mph. Probably into another lane, potentially into oncoming traffic.
Human reactions are poor. Houseplant reactions are worse. Tesla acted more like a houseplant than a human
> a sideswipe is almost certainly preferable for everyone than plowing head on at full speed.
I’m fairly sure this is not true. Modern cars have lots of energy-absorbing crumple zones in the front, plus airbags. Head-on is the most highly-optimized crash orientation. Side-swipes are I believe the worst case, because there is so little protection for the driver.
Assuming the driver was paying attention to the road, even taking ~2 seconds (About how long it took me to clearly recognize the truck in the video) and slowed down from 70MPH from 40MPH, it would reduce amount of kinetic energy in the collision by ~300%. It could be the difference between a survivable accident and a lethal one.
7^2 is 49, 4^2 is 16, all the other factors cancel. Yes, of course the right way to say it is "reduces the energy by 67%," but it's not that hard to see what was meant.
Based on my experience with drivers and driving at night, it is very unlikely that humans would do better in most circumstances.
> till, even then, a human would have slammed the breaks or swerved to the right to a free lane
Most humans unfortunately do not have the experience of knowing what to do at the right time; they may swerve to the left here in this video and crash into a more dangerous object (another high-speed car or truck) and cause a more dangerous pileup. There is no right or wrong answer in this situation, we don't do drills (at least in US) that at least tell people what to do in bad situations like this. This may be a practical question but in reality, no one has the training or experience to deal with this _instantly_ and worse, in dark environments.
Also, most people is likely to be tired if they're driving at night such as coming home from work.
What this proves is that we need emergency lighting systems in the event of a crash. Like if an airbag has been deployed, lights around need to be visible to avoid this kind of scenario.
It would be awesome if we could provide VR training and use videos like this as examples to train folks.
Mercedes does this, along with a lot of other cool things they call “pre safe”, like playing a loud noise on the speakers to prepare your ear drum for the sudden change in air pressure.
I’m sure other cars are doing the flashing lights now but few companies have invested as much in safety as Mercedes.
I set off the hazard lights in the '94 BMW I fixed up in college when I turned up my subwoofer too loud once, so I think the accelerometer-triggered lights are, or should be, widespread already.
First, robust response to a crash. Maybe this is lighting (flood LEDs on the back of first responder vehicles), maybe this is emergency vehicles with crash shock absorbers (Scorpion Impact Attenuators is a well known brand) that sit behind the accident in lanes of travel to protect responders at the accident scene. Cars should attempt to detect this, but it should be expected this may fail. The crash absorber will take the brunt of the crash, but everyone is going home to their families with some tears shed by the insurers. This applies to all vehicles with traffic aware cruise control and lane keeping who can’t reliably detect stoped vehicles (which is all automakers).
Second, first responders (through dispatch) should be able to issue an alert to Google Maps and any other traffic aggregator to alert about an accident in a lat long “bubble” (think a TFR in air navigation, pushed out over ADS-B or a data link) with an expiration time. Vehicles would receive this event and temporarily apply a reduced speed limit in the area (one that the driver could override, but that’ll be logged in the black box when they push the accelerator past the temporary limit). This also provides alerting to the driver to be attentive about an upcoming event, just like Waze alerts you to an upcoming accident or speed trap. The vehicle integration is key, as the robotics is what needs the safety context and awareness.
Now, if you’re not paying attention as a driver, no amount of tech and effort helps, and these folks end up killing themselves and others every day with inattentive and distracted driving.
I believe there was an attempt to make car to car communication system mandatory in US so that every single car on the road can talk to each other but I admit, I do not know how all of this worked and what happened in the end or if this is still ongoing.
I can’t speak to consortium work such that you mention (V2V comms), but Google and Tesla could, today, use their existing traffic data feeds and mapping tools to both allow first responders to create events and for the vehicles to interpret and action that data.
This is also vastly cheaper than LiDAR on every vehicle in the short term.
> What's worrying isn't that the autopilot didn't "know what to do"; but that it didn't detect the oncoming parked car at all.
This is not simply a parked car, this is a dark non-reflective crashed pick-up truck sideway on a highway lane that's CURVED at night. There are multiple variables all of which are bad; non-reflective, driving high speed (75mph) in a curved lane, and night-time.
Worse, this car's detection system is based on video camera only, it does not have LIDAR. That's where Tesla has a big problem with and it clearly shows here. Video camera-only system has blind spots and using auto-pilot or self-driving system with this only is not a good idea, period.
In addition, we're not supposed to rely on autopilot only regardless of how good anything is; this is all is supposed to be a complimentary tool just like every other safety feature within the car; the cruise control feature is not supposed to be used blindly on its own and excellent human drivers with cruise control can still hit that car.
There's no 100% foolproof algorithm or method to predict every single scenario.
If the driver was relying on the auto-pilot and some kind of detection system and didn't pay attention, then it is not the system that failed, it is the driver's fault entirely.
> "Most" (meaning > 50 percent) is just off the page, and doesn't jibe with reality. Sounds like you're a bit too eager to give Tesla a pass, here.
I'm not giving anyone a pass, I'm pointing out that humans have their own weaknesses and blind spots that machines can help with but humans on their own, will not be doing better in this specific scenario with all of these variables. I can actually understand why the system failed to do anything given the technology with limited sensor data but I don't understand why they didn't use more sensors like lidar, infrared or emergency breaking system or so on. I may be missing something here.
Regardless, we're not looking at a scenario where it happens every day, it only takes one bad accident and one tired human to turn an uncommon scenario into a disaster. So, it doesn't matter what percentage we deal with.
> If the driver was relying on the auto-pilot and some kind of detection system and didn't pay attention, then it is not the system that failed, it is the driver's fault entirely.
IMO it’s the fault of the engineers who did not take human factors into account. We all know that humans are terrible at following instructions or even using common sense — hence warnings about not using your toaster in the bathtub — so when engineers pretend that users will do the right thing, the engineers are at fault for not addressing the behaviours of the multitudes who will not do the right thing.
TLDR: the driver is at fault for being a dumbass using autopilot at night, and Tesla is at fault for not considering how real humans will use the product.
Tesla is at fault for not considering how real humans will use the product.
I think they know perfectly well, on some level.
As they are of the fact that large portion of their drivers will be perfect dumbasses.
But they're invested too deeply in the promise of an "Autopilot" product (which of course consumers will take to mean exactly what it says) by this point that they're not going to backpedal or change course -- unless forced to do so.
I also feel like the way Tesla drives is really aggressive, I basically can't use is adaptive cruise control in slow/stop-and-go traffic because it ZOOMs right up the the car in front and then slams on the brakes, rather than just going along at a reasonable pace. Similarly, in this scenario with this visibility I might choose not to drive 10mph over the speed limit.
There are cases where a human driver is better, but there are significantly more cases where Autopilot is better. For example when you drive for 4 hours straight or driving through stop and go traffic. Think about about the lives being saved and be sure any conclusion we reach is based on rigorous analytics and not just gut feeling after having watched a horrific video, which also happens numerous times on daily basis to all types of cars.
> For example when you drive for 4 hours straight or driving through stop and go traffic.
If I had a uninterrupted four hours of drive as part of my regular commute I wouldn't buy a Tesla, I would buy a gun and a single bullet with my name on it. Two hours one way every day were already enough to drive me almost insane.
> watching the video [...] it is not clear if a human would have done any better.
Anyone who's ever taken a photograph in low light conditions knows that in the dark, cameras fail to capture loads of detail that humans can make out just fine.
I'd wager a human driving in such circumstances would have seen the truck much more easily than we can see it in some dark, blurry dashcam footage.
And as bad as it looks like on that camera, its a way better camera than the 1.2MP 1/3" CMOS sensor that Tesla uses for its "self-driving" cameras.
In any case, our human eyes are actually really, really good. Far better than typical CMOS technology at capturing dynamic range and whatnot. Our Iris opens/closes depending on a variety of conditions, and our retina itself is 22 mm, far larger than typical camera CMOS sensors (but smaller than the top-of-the-line 35mm cameras).
Not only is our surface-area larger... our rods / cones seem to be superior in dynamic range than CMOS technology.
----------
That being said: I think a big problem for Tesla's system is that they use tiny and cheap 1/3" CMOS sensors at low resolutions, and then expect AI to magically fix all their problems.
Other companies are willing to not only buy more expensive cameras, but also more-expensive sensors, such as RADAR or LIDAR, to assist the camera.
I think some people are underestimating their reaction time.
Agreed; the human brain has special chemicals which significantly amplify our processing capabilities during moments of acute danger. I was once going up freeway, in the dead of night, and spotted a (darkly painted) car fully flipped over, about 2 seconds ahead of me. Managed to both decelerate (just enough) and bear gently to the right (not swerve), with room to spare.
Granted I was going 60, not 75. But I'm not the one telling people to drive on AP at 75, in the dead of night, no less
> At this point it is very clear, to me, that anyone using FSD/Autopilot is risking their lives by trusting their lives on a beta, or perhaps even alpha product
Are you using population-level accident statistics to make this assessment?
I’ve seen analysis that suggests the overall safety numbers for self-driving mode are better than human drivers, though I’ve not dug deeply enough to be confident in them.
I’d caution against extrapolating from individual incidents though. I think we should expect that at the point that self driving is safer than a human driver, it will be worse than humans in some cases - simply because it can pick up a lot of gains by never being tired, distracted, or drunk. If one just looks at the pathological cases and says “I could have avoided that accident, therefore self-driving is a worse driver than me”, I believe one would be making an error of reasoning.
There’s an interesting philosophical/legal question at play here. Does self-driving have to be strictly superior - ie no edge-cases where it is worse than a human - to be acceptable? Or can it be worse at some tasks if it’s killing fewer people overall? Many people will die if we are too strict here. And yet, people need to trust the system as well. I suspect this will be optimized based on expected costs of lawsuits, which is a poor proxy for morality…
People expect self-driving cars to be better than human drivers and not just statistically better. Any accident a self-driving car gets into that most humans would have avoided should be considered a fault of the self-driving system. There shouldn't be wholly new ways to get killed on the highway and us just shrug it off as yeah but fewer people are dying overall. That's not gonna fly.
Of course it is going to fly. If we can, overall, save lives with these systems, then we are letting people die if we wait too long to deploy them. A lot of people think that way.
> I’ve seen analysis that suggests the overall safety numbers for self-driving mode are better than human drivers, though I’ve not dug deeply enough to be confident in them.
All similar analysis I've seen compared all-road all-vehicle driving stats against autopilot stats. Of course, the correct comparison would be against luxury cars (with similar safety features) and only on the same kinds of roads that autopilot gets used on (i.e. primarily easier highway driving). As far as I can tell we are extremely far away from the hypothetical you raise where self-driving is better than human.
> Are you using population-level accident statistics to make this assessment?
I’m making a rather simple assessment. I’ve seen Tesla autopilot/FSD fail catastrophically in situations where I would have been fine. This does not inspire confidence that it will do better in situations where I would have failed. As such, I’m not willing to put my life in its hand. And, I really hope no effing Tesla hits me.
> I was driving on a Freeway at ~65 mph, 87 N for those from the Bay Area, and due to the curving of the Freeway I did not see the traffic stopped in front of me
I always shudder when I hear that people drive too fast to be able to stop in time based on visibility. Same with the crash in the video. Way to high a speed for that visibility.
i think driving tech will get there but right now i wouldn't trust it with my life. Humans are basically short sighted, they basically rely on their short term sense of induction and gain a sense of over confidence. Oh it works once? ok i feel a little better, works twice? ohh even better, works for the 10th time? Hell yea, let me whip out my phone and watch videos while the car drives itself. I think for my i will trust this amazing thing in my head.
While it might be the case that the human eye can see father than a camera at night, the car should have other sensors that should enable the car to see further. That is the whole point
I think the fact that the human didn’t intervene means that no, human eye can’t see further in these conditions. There’s more to eyesight than maximum possible sensitivity and no the driver’s night vision is not going to be particularly good thanks to oncoming headlights. What you see on the video is the result of transcoding the actual CCCR camera feed to RGGB, reducing resolution from image native to what YouTube wants, and then compression to minimise the file size of a video with lots of black and lots of blown out areas.
Swerving could have prevented the collision, but braking is always the safer option. Swerving would still require the object to be detected, which neither the AP nor the human did in this instance.
Regardless you are still correct that people trusting their lives to AP instead of monitoring it like they are nagged to do by Tesla software are being foolish. I would never have been in the inside lane in this scenario, most likely if using AP I would be happily cruising behind the coach, meaning in any collision with a stationary object the coach would absorb the bulk of the collision energy.
Making this incident about AP is putting the cart well and truly before the horse.
Visibility of the stopped truck aside, one could see even in the non-compressed version of the video how the truck obscures headlights of on-coming vehicles. Human drivers can pick up on that contextual information to see that something might be there to obscure the light even if it's not clearly visible on its own.
I think this is one of the big points about humans vs. computers for complicated problem solving like this: context. The autopilot's job is to drive the car safely when it can. It's job isn't to pick up the tiny minutia of offputting sensory input like obstructed headlights - it isn't programmed to _reason_ about things like that, something that would raise the hairs on a human and cause them to double check what's in front of them.
Yeah, I think this is really fascinating and challenging aspect of AI and self-driving. We have a model of the world in our brain and we're constantly evaluating what ought to be there or not and checking consistency between all the sensory information we receive but don't think AI is quite there yet.
I think that in our constantly changing world AI will never be able to "catch up" with us. It might be able to have some ability to "catch up" to where humans are now, but that will take many decades, and by that point there will be new obstructions and different things to watch out for and it will need new training again. I have an intuition that these machines will always be lagging behind. But maybe I'm wrong.
There's a legitimate fear that AI will be just good enough to be the scapegoat for bad or lazy decisions that devastate people's lives. To some extent that's already here, but imagine widespread drone policing, social credit, or nanny-state appliances.
There is; we find and prosecute their human creators if they commit a crime. Every intermediate step is just smoke and mirrors. You wouldn't say, "Well, we can't charge the gun with anything..."
While theoretically that would be the best solution but I don't imagine it working in practice. No one was sentenced to jail for manslaughter or homicide for the Boeing 737 MAX failures. Company gets fined -> lower level less wealthy employees get laid off -> execs retain wealth & start a new company
I love all the other people saying that they wouldn't have been able to avoid it, but the many people who were actually driving on the road managed that just fine.
The first accident seems to have been very fresh. This Tesla may have been the first car to happen upon it. Though at least from TFT we can't be certain.
Many people would have hit the pickup. But some would have been able to respond and manually brake before impact, and this is the exact sort of collision that Automatic Emergency Braking (AEB) is supposed to mitigate, and there are dozens of cars of different makes that have it installed [1]. It's not like Tesla is the only car on the road with sensors. This is a failure. The fact that a radar-equipped "full self-driving" car doesn't even tap the brakes in the two seconds before hitting a unavoidable solid object in the middle of the road is a travesty.
The Volvo system is based on radar + camera and it's pretty obvious that that is the rear of a vehicle in front of the truck. It might not necessarily pick up every kind of stationary object.
It would do exactly what you would expect. It would increase the ABS pressure up to a limit that doesn't generate excessive G force to the passengers necks.
The impact would 100% still occur but at a much lower speed exponentially reducing the extent of injuries to all involved.
> All radar based braking systems in the market filter out stationary objects because too many false positives.
Absolutely categorically NOT true.
It may not fit your narrative but Front Radar electronics/firmware filter out stationary object ONLY if they are outside the vehicle trajectory envelope.
The majority of false positives are due to overhead passthroughs which may fall within the envelope, if the radar calibration is insufficient.
EuroNCAP tests for AEB include both moving targets AND stationary targets. [0]
Even those low-cost Bosch front Radar assemblies in entry-level vehicles could have mitigated that impact.
> It may not fit your narrative but Front Radar electronics/firmware filter out stationary object ONLY if they are outside the vehicle trajectory envelope.
The first challenge is that cars do not always travel in a straight line, so the "vehicle trajectory envelope" is ill-defined. At minimum, the radar needs to be steered (electronically) to match the anticipated vehicle path. However, the anticipated path does not necessarily match the actual path.
The second challenge is the size of the target. If you re-watch the video, you'll notice that the pickup is sideways, and appears to have collided with a sedan prior to the Tesla arriving. In other words, the radar target is approximately 2x to 3x the angular width of a normal vehicle.
If you filter the radar returns to eliminate wide stationary objects (to, for example, reduce phantom braking caused by overpasses), you necessarily handicap it when detecting a truck parked sideways on the road.
Putting on my ML engineer hat: One of the common themes in examples where Teslas fail to brake is that the target is stationary and very wide: A semi truck crossing a not-limited-access highway, a crashed pickup turned sideways in the lane, etc. etc. etc.
> EuroNCAP tests for AEB include both moving targets AND stationary targets
Your link shows only scenarios where the target is exactly positioned along the centerline of the vehicle under test.
Additionally, the targets are only as wide as a normal vehicle--and if that weren't enough, the tests specify the use a combination of radar absorbing material and corner cube reflector to create what is essentially a point target.
My Volvo has a system like this. I had maybe 4-5 benign false alarms since buying the car 4 years ago, and once it did stop me read-ending stationary traffic once.
What does full self driving have to do with it? The car is running on Autopilot (called Autosteer in the UI of my Model S).
> his is the exact sort of collision that Automatic Emergency Braking (AEB) is supposed to mitigate, and there are dozens of cars of different makes that have it installed
Is it supposed to? And do they actually mitigate it?
1. Many people don't view the world through a shitty 480p camera with compression artifacts.
2. Almost nobody would have hit the pickup still going at 75 MPH. Even dropping 10 MPH at that speed reduces the energy of the crash by 25%.
3. As another poster pointed out, nobody else on the road has managed to hit that truck. So, either all the other drivers are incredibly lucky, or it is, in fact, more than possible to both see and avoid that truck.
I don't know how much is from knowing there was a truck in the road, but I noticed in my first viewing how something was blocking oncoming headlights.
If you're going 75mph, that should really get your spidey-sense tingling. At least let off the accelerator/cancel cruise control. Be prepared to brake. Once you're closer and lined up, it's harder to tell there's anything going on because you're no longer approaching a bend with headlights on the other side of the truck, so by then it's too late.
But this is, again, all hindsight talking. In the moment, it's much harder to assess how a "typical human" might have reacted, if at all.
Had the same feeling, incoming headlights are not visible = obstruction. pity the car has nothing that could be called intelligence.
I’ve spent a lot of time staring at the road thinking how I’d write a program to do it. Main thing I’d do differently is provide visual feedback of confidence. Having the autopilot disengage when it gets confused isn’t good enough, I want to know when the system is losing confidence /before/ it disengages, and I need it to let me know that it’s not 100% sure the next thousand feet of road is actually clear. Tesla seems to drive unperturbed until its classifier recognizes an obstacle.
As someone who is a decent-ish programmer that's recently worked on recognition from videos, I would be totally lost trying to figure out how to handle a situation like this. As a human person (I swear!), I could see in the video that the headlights were being obstructed by something but, as a human programmer, I'm completely at a loss for how I would even determine if the headlights are blocked by something up close vs. far away using just the camera's view. Headlights get blocked by things all the time. How in the hell would you figure out that that's what's going on just by the visual?
In other words, it's a good thing I don't make this stuff. My car would be phantom braking every 10 seconds, especially at night.
Yeah, I've wondered about this too. If cars in the adjacent lanes slightly ahead of me braked all of sudden, I would brake too even if the road ahead of me is completely clear because I would think there is something going on. I wonder know how this type of thing gets programmed in to self-driving vehicles.
I would definitely want night vision infrared, but my main strategy would be making sure the texture of the road surface is flowing toward me at the expected speed. I imagine I could use something like the motion vectors in mpeg compression. I would exploit the fact that the road 1000 feet away is coming at me slower than the road 100 feet away: if there is a visual texture in the middle of my field of view (where the horizon should be: road approaching relatively slowly), and it’s coming at me faster than 1000-feet-away asphalt ought to be, I assume the road is obstructed and slow down until I get a clear view again.
Granted, I am a lowly web programmer with a hobbyist interest in compression, but I have thought about getting the comma.ai kit and learning how it works, trying to implement my own world modeling, but, sounds like a full time job :)
EDIT: as for the headlights, I don’t want to act like I could whip this up in a weekend or anything, but having objects flicker in and out of existence would at least subtract a few points from my confidence score.
That's an interesting way to go about it but would the texture of the road surface be consistent enough and reliable enough to make those calculations? This is the kind of stuff I'm talking about. I definitely do not have the confidence to do that. It's why machine learning is so scary to me. We're basically saying "I can't figure out all the possible permutations of this so let's throw them at a computer and hope it can do a better job".
Can you tell me more about what kind of road (in)consistency you’re describing? It’s fun for me to think of how to improve this strategy.
I am similarly opposed/nervous around machine learning, seems impossible to debug weird decisions the black box makes, I would prefer a traditional, deterministic computer vision approach.
Basically I want to segment incoming images into a grid and do a cosine similarity score between each cell and its neighbors one frame ago, to detect which direction and how quickly all the textures are moving. So I think it should be robust to whatever the road texture is, provided its not pitch black like wet asphalt.
you want the opposite, if it's consistent, then it's more difficult to gauge it's motion just by finding similar zones. think on mouse on glass or metal, it's too uniform for the sensor to detect motion.
The biggest change I’ve made over 35 years of driving is slowing more and sooner when things are unclear. At 75, just taking your foot off the accelerator will shed 10-20% of stopping distance in a few seconds. Waiting until you are sure it is bad news is the difference between defensive driving and avoidable crashes.
>At 75, just taking your foot off the accelerator will shed 10-20% of stopping distance in a few seconds.
Even more in a Tesla as the regen braking will slow you down a lot. I drive almost exclusively with 1 pedal now because of it and, while I don't know that that actually is safer, it feels safer because I'm slowing down waaaay earlier than I normally would in an ICE-mobile.
You can use where the headlights disappear and reappear to draw an arc of occluded space. You should be able to measure how quickly that arc changes to estimate the distance
If you're 1000 feet from an object and move a foot closer, it barely blocks anything else from view. If you're 5 feet away, moving a foot closer blocks a lot more of your view.
The problem is that relies on nearly constant backlighting on the object.
The real solution here is LIDAR, which makes this trivial. Tesla won't use it though, so they have stupid problems like trying to hack a way to determine how far away dark objects are.
Exactly. I was just going off of Musk's statements that they don't want to use things like LIDAR since people don't have LIDAR but can make these types of assessments.
Even if you could make an arc of occlusion (sounds like a D&D spell), there's way too much variability in headlight sizes/designs to accurately determine distance in the dark in the situation this video presents. Again, I am admitting a ton of ignorance when making that statement but that seems nearly impossible to me with just cameras.
You have the advantage of knowing before hand something terrible is about to go down. The driver was probably driving for several hours at night and has his/her spidey-senses a bit fatigued.
This is a dark truck with no lights in a high-speed road at night, I would probably have crashed too. Yes, human eyes have better night vision, but people have different visual acuity, attention span, and reflexes. Also, headlights from other cars can blind you momentarily.
Still, I don't understand why Tesla is so opposed to LIDAR. The more information the autopilot has, the better decisions it can make.
Lidar is prohibitively expensive. There aren't a lot of manufacturers. Velodyne is pretty much the only game in town, they primarily sell to defense industry. GM via Cruise bought a startup that promised solid state lidar. It ended up being a complete waste of time.
man, until you said that, it wasn't something i had noticed, and I was staring straight at it with full attention. I can't imagine someone driving for hours picking that up. After all it could an obstruction in the median.
Sure radar would have detected that, but even with naked eyes you cannot see the black crashed car against pitch black background without any lights emitting form it.
Comparison with human ability aside, this would most definitely not have happened with Lidar. Musk's argument has always been that you don't need Lidar because humans don't need Lidar, but since the goal is to do better than humans this might not really work out in the end.
Please try not to draw general conclusions from specific examples. First we need to know what percentage of crashes are due to lack of lidar/radar instead of looking at anecdotal cases.
Apart from the higher price why would you not want additional sensors like radar and lidar? If all human driven cars were already equipped with radar and emergency stop systems, mass pileups would be a thing of the past.
Obviously there are situations where cameras or our eyes simply won't be able to see enough. Dark unlit objects at night aren't exactly an obscure edge case scenario, the crash from the video is something we already see on a regular basis with human driven cars, sadly. And even if not, I'd rather pay $1000 or so extra to protect me from that freak scenario that's likely never going to happen. Until it does.
I also have airbags in my car although I never needed them.
This is a well known and basic topic that you will be trained on before allowed to enter any physics lab with lasers in it. This can happen around shiny objects, windows on streets and pretty much anything that reflects light.
"So coherent is laser light that one may observe interference effects even when the path difference between the interfering rays is much greater than 109 wavelengths! ... If two waves are exactly “in phase” as shown in Figure 1a, they will reinforce each other; this is called constructive interference." [0]
I was skeptical as well but looking it up it seems it's a genuine concern. Also you're right about limiting the power but interference in general could lead to problems, not just with lidar but radar as well. Maybe not impossible to solve but it seems like difficult issue, you wouldn't want misleading sensor reading lead to the car making a fatal decision.
Even if it was only used for AEB instead of AutoPilot (i.e. acting as an actual backup to autopilot) that would be a vast improvement.
A Tesla vehicle, when under human control, has AEB but when under AutoPilot it essentially does not since it is the same logic used for both processes and therefore it doesn't act as an additional safety feature.
Yep, their bet on not using Lidar seems bizarre to me. Presumably it’s been done for cost purposes, but lidar will likely be cheap by the time AI driving is viable.
I think the issue at this point is that they've promised - i.e. taken money for - "full self driving" features that are supposed to be delivered using the car's hardware as-is. So they have to either work with the cameras they have, or retrofit a bunch of cars with LIDAR, or just fail I guess.
However, the car was equipped with radar. Twitter author also said that the truck was seen by radar. I don't see how something like this could have been missed by a radar-equipped car. I chalk it up to the black-box nature of these models and the general lack of awareness about what bizarre edge cases could be encoded in the model's parameters.
I think you're completely misunderstanding my comment. I was saying that the fact that the model missed the truck can be described as an edge case. It's an edge case in the sense that it could end up being a strange and rare outcome that results from an unfortunate compounding of error at the intersection of a few parameter in the model. Deep learning models (which are undoubtedly involved) are problematic in this way. Of course I agree that this sort of driving condition is not an edge case and should be in the training set. But that doesn't mean that anything like it will always be correctly categorized by the model. We've seen that it's possible to trick some computer vision models with well-placed, single pixels. If anything, I'm saying that we shouldn't be entrusting our roads to deep learning models over a certain level of complexity.
It also shouldn't have happened with a RADAR. From the second tweet:
the truck was seen by both the radar and also could be clearly observed by the narrow camera from quite a distance away (pictures are spaced 1 second apart)
This is more about beta quality robotics software being shipped to actual customers and used in production. If you're someone who won't even use a programming language or library in production before it reaches v1.0, this is about a million times worse than that.
By the way, as a roboticist this whole situation is completely baffling to me. We go through such great lengths in our research to make our robots safe. Safety is a huge concern to our community. How is it that this company can just ship beta quality software to customers for use in production amongst the rest of us? In industry, we paint our robots orange and put them behind impenetrable cages due to safety concerns. Entire facilities are designed around the robots to make their operation safe. Operators are trained to use them. People working around them are trained to work along side them and observe proper safety protocols. Proper insurance is carried in case of accidents. And when an accident does occur, the situation is taken apart, analyzed, and changes are made so they won't happen again.
Not in the marketplace though, it seems. Apparently it's okay to just code whatever you want; put the code on whatever machine you want, even if it's "beta" quality; ship it to whomever you want who can pay for it; provide little to no training; lie about the capabilities ("Full self driving" minus the full, and self, and driving. "Autopilot" that isn't.); and when people inevitably and predictably end up dead, apparently it's okay to just keep on pushing this software out without a shred of self reflection or humility. Because you're saving the world I guess? I just don't understand this culture.
> Musk's argument has always been that you don't need Lidar because humans don't need Lidar, but since the goal is to do better than humans this might not really work out in the end.
aren't auto manufacturers that don't have features like autopilot currently adding things like lidar and automatic breaking to improve safety?
why wouldn't you add the thing that will make your system as safe as possible?
what's the goal here? some aspirational tech challenge that competes on arbitrary merits in some humans vs. machines challenge problem or building the best and safest way to move vehicles around?
The issue is that the guy who runs the company and controls everything doesn't know what he doesn't know. He made the uninformed decision years ago to go all-in on camera technology because he thought that he would be able to throw a bunch of money and engineers at the problem, and solve Full Self Driving in essentially 5 years. That was more than 5 years ago and the technology seems to have plateaued at a place where it runs frequently into parked or stationary objects.
They have since taken untold sums of cash from people (millions? tens of millions?) who they promised would get access to "Full Self Driving" with the hardware on board. Now they're in the situation where they cannot admit that the problem requires more hardware. The hardware is a fixed property here, a design requirement; they have to achieve actual full self driving with the hardware in place, or they're going to face refunds and lawsuits.
So now it's a delay game. I think it's clear now that short of AGI, Tesla will never get to "full" self driving with the sensor suit they have. Nonetheless they can't switch to a new sensor suit because that will admit defeat. Instead, they have to show continual improvement of the current sensor suite, while promising that "Full" self driving is just around the corner so they can continue the charade.
so is tesla running cameras with 21+ stops range? idk why isn't this argument dismantled every time musk mentions it just by pointing at how actually good eyes are.
Next time you drive at night time, compare what was visible to your eyes to what your dashcam captured. Then upload the video on Twitter, and compare it again.
Watch it full screen on your desktop monitor (still wouldn't be as good as human field of vision) and remember this was recorded by an average quality dashcam, not a high-end camera.
313 comments
[ 5.1 ms ] story [ 284 ms ] threadSounds like the risks in using both either autopilot or FSD systems at night is even more dangerous but very surprising that these systems still have no night-vision system.
Shocking.
[0] https://news.ycombinator.com/item?id=28732566
[1] https://news.ycombinator.com/item?id=29515784
rock fall, tree, building.
Given that the road is occluded, this needs to be in the training data, if only for autonomous lane following.
Does hw2.5 not use the radar data?
It seems like it does use the radar. The generation after 2.5 (3.0) did appear to improve the "frames per second" processing from ~110-200/sec to ~2300/sec.
I'm guessing the comment means something like "if this was related to sensor processing throughput or latency, the 2.5 version might be meaningful".
https://twitter.com/haenschengross/status/147332613675511808...
Having been personally involved in a collision with one where my vehicle ended up under their trailer at highway speeds, when the trucker abruptly swerved into my lane to avoid rear-ending another truck they were tail-gating without paying enough attention to brake in time, I find this thread of discussion patently absurd.
It's so sad to see it, but then they claim "oh but it was empty". People use it as their resort when following too closely, I wish they'd just accept their mistake and crash into the back of the car in front of them. But that would be too clean and logical.
Human eyes have much higher dynamic range than typical video cameras. What the video shows is not likely representative of what your eyes would have seen.
1. Even a sober, awake driver glances down to confirm their speedometer and scan their fuel gauge, check engine light, tire pressure, etc.
2. A sober, awake driver should be scanning the road along and beside the projected vehicle path. But they doesn't necessarily mean the would be looking in the direction of the occluded headlights at the very first instant they tell of a crashed car.
3. A sober, awake driver has to process what they see.
4. A sober, awake driver still has a good 0.2 to 0.25 second reaction time.
5. I actually do look for occluded lights when driving at night, but my experience talking about this with others in person is that this is a totally foreign concept.
In other words, you know where and how and when to look for the obstruction because you've watched the video a few times.
This is sort of like the Air France 447 flight or the Boeing 737 Max trim problem. If someone tells you exactly what kind of scenario you're going to face, how to handle it and approximately when you'll face it: you can win every time. 100% of pilots who were briefed on the specifics of AF447 avoided crashing the sim. Same for the 737 Max.
Yet, hundreds of people are dead because pilots in actual cockpits, who were not given prior notification failed to act correctly.
A major difference is that (from watching the uncompressed footage), a human driver would likely have stomped on the breaks before impact. That last minute breaking could make a world of difference.
A collision at 75 mph has a lot more energy (4.5x ish) than a collision at 35 mph. So every 1/10s of breaking will do a lot to reduce the damage of the impact.
That's one of my concerns about driverless vehicles that I want to see the statistics on. I have a hunch that automation will reduce the absolute number of impacts, but may lead the impacts that remain could be higher energy impacts, because they result from a classification error.
in person one might not realize a truck is there, but brain would probably think "something not right" and alert senses maybe even reflexively release pressure on the accellerator, so that when higher level consciousness makes out the truck, body is primed to act.
Wife works in automotive engineering and I asked her why FCW and AEB didn’t detect it and her only thought is that because the truck is on the same angle as the central reservation and already in a bit of state it might not have detected it, although she thinks that more likely something just went wrong
I also thought I might have been able to take the 18 wheeler's reaction as a clue, but it passed the wreck in the other lane and only starting braking right as it passed it.
With that in mind, in this exact scenario, I would have hit it also. Only with full manual control of the car do I think I could have avoided it.
I'm pretty sure I would have hit the truck too[1], but there actually was an earlier clue: Headlights from oncoming traffic are occluded by the two crashed vehicles ahead.
When rewatching the video (so I know both that I should be looking for something and approximately where that something is), the occluded headlights are pretty obvious. An attentive driver may have noticed the vehicles, but I think most people saying the truck was obvious are deluding themselves.
[1] The first time I watched the video, my reaction was "Where is this tru-OHSHIT".
If I'm driving at 120km/h , at night, chances are I'm also tiered and reacting that quickly at that speed with limited vision would have been a coin toss for me.
I'm all for making these cars safer but we are watching that video, knowing the truck would be there but fail to consider all the extra context around it. I think accidents can and will happen with self driving cars, no matter what we do.
I'm reminder of this multi-vehicle accident where you can see some drivers clearly reacting 1-2 seconds before impact. https://www.youtube.com/watch?v=-9yqXzZ16ns
In the uncompressed video it seems very obvious that there is an obstacle in the road much earlier
Whilst its unlikely the Tesla autopilot would have fared much better in the centre lane, the impact could have been substantially less.
On the other hand I think it's at least plausible that a human driver would have been able to notice the truck in time, even being in the outer lane.
However human night vision is much better than camera night vision, so I wonder what an actual human would have seen in that scenario.
But it's certain the camera could not really see it.
In any case, the tweet claims there was no braking of any kind, and I'd expect braking at least in the split second before the crash, where the truck is visible to the camera.
[0]. https://twitter.com/haenschengross/status/147332613675511808...
(Oddly, I can't seem to find any such videos at the moment or know what terms I should be using to search, but I definitely remember people uploading such videos to YouTube.)
[1] https://fortune.com/2018/03/19/uber-self-driving-car-crash/
Either one of those would have detected this.
My Audi's radar works great in pea soup fog. Has prevented me from nearly rear-ending a gray vehicle with no lights on in heavy fog.
If removing LIDAR lowers the cost of the vehicles which means more of them are sold which results in a fewer accidents and fatalities overall, then I think it's the right call.
This argument needs to be fleshed out more after a video like this surfaces. I can't accept at face value that this style of autonomous car, which has demonstrated issues running into parked and stationary objects, is the platform which when mass distributed will reduce overall accidents and fatalities.
It's clear from this video that this vehicle doesn't know when it's blind, and will merrily carry on at 75 mph even though it has no idea what's going on around it. If you are driving in the dark or the fog and you can't see 12 seconds in front of you, you slow down until you can. Not Tesla apparently -- it will continue on at 75mph even though it can't break in time if something emerges in front of it. How is this the technological basis of a something that promises to lower overall fatalities? How does this get better with scale? Seems to me, based on this repeatedly demonstrated mode of failure, scale will only reveal how broken their self driving system is.
Why is it accepted here that filling the road with autonomous cars that don't know when they're blind and run into parked objects will somehow result in a net positive outcome?
Moreover, how is Tesla in particular the company which is going to deliver these safer vehicles that make our roads safer? Tesla has a history and culture of deception in their marketing. They haven't taken these accidents seriously in their engineering and product design. Teslas have been running into stationary objects for 5+ years now and they show no sign of fixing that. Call me crazy but if I were shipping a driverless car technology, I'd want to make sure it didn't run into things. Maybe that's just me?
Tesla recklessly released this flawed technology and product to actual consumers for use on actual roads. How are these the actions of a safety-focused company? From my perspective, if you flood the roads with Teslas, you're not going to get safer roads, you're going to get more Twitter threads with videos like this. The claim that this will necessarily lead to safer roads is yet to be proven.
I think you underestimate how bad human drivers are. In the US, something like 100 people per day are killed in car crashes. Depending on how many Teslas are on the road, seeing a video like this every day may not be that surprising. If bugs in self driving tech kills 10,000 people per year (30 a day) but the overall fatality rate drops from 35k to 25k per year, it could be seen as a win.
Tesla as an organization has demonstrated a level of corporate sociopathy that tells me they don't take safety seriously as a culture. Surely, they tout safety as a selling point but their actions tell me they are far more interested in getting driverless cars on the street quickly than doing so safely.
You can't achieve sweeping victories in safety without intentionality. Tesla may claim that safety is a priority for them, but if that were true they wouldn't lie to their customers about the capabilities of their own products. They wouldn't eschew industry best practices as far has hardware is concerned. They wouldn't beta test unproven software on their customers and charge them for it, leading to bodily harm of their customers and others. They would fully investigate problems as they occur and definitively fix them instead of allowing concerning issues to persist for half a decade now (running into large stationary objects). There are many organizations working on driverless cars, and yet Tesla is the only one that behaves this way.
I really hope we as a community have grown past the idea that simply throwing technology into the world will have positive net benefits as long as there are good or good-adjacent intentions behind them. In fact, it takes planning and execution to ensure good outcomes. Simply doing something that can be cast as a boon for safety does not guarantee a boon for safety if safety isn't a top goal. And by top I mean it has to be #1 or #2, because any lower and other priorities eat into it.
Except that every bit of available evidence shows that not only do self driving cars not reduce fatalities, there is no reason to expect that they will.
People are surprisingly good drivers - the reason there are so many accidents is simply that people drive so much. I ran some numbers many years ago showing this: https://news.ycombinator.com/item?id=12012166 check them if you like.
It's hard to see in the video even when primed for it, much less likely a human would see it when not expecting it.
The system seems it, at least the sensors will and then they send the image to the module that processes this and runs the algorithm to detect that its a stationary object and that it needs to hit the brakes. Then the signal needs to be sent to the module that brakes. Being a real time system, all of these need to happen within a time window for the brakes to be applied at the right time. Its possible that because of some resource constraints or poor programming/design one of the steps in the chain is not triggered and so the system did not brake.
Maybe I would have stopped. Maybe not.
For instance:
https://www.volvocars.com/lb/support/manuals/s60/2015w46/sta...
I think the issue is that those lights are on the main power circuit, which shuts down if the user turns off the car.
That appears to be what happened here, I don't see running lights or anything. I think the person turned their truck off.
I honestly don't know if you're supposed to shut your car off after an accident to lower the fire risk, or leave it on to make the car more visible.
Really shows how Tesla has probably made a big mistake going all-in on camera only vision. And their vision system couldn't make the inference of an obstruction based on the oncoming headlights.
Lidar would have been able to detect this easily.
Tom Andersen: Your eyes have much wider brightness range than any video camera ever made, and many many times better than the $1 camera on a Tesla.
That is not a representation of what human eyes would see. It's a cheap video camera.
The thing that would(or should) have alerted you as a driver is the reflecting of the truck's lights on the paint of the broken down pickup as it drove past. Although why the truck driver didn't slowdown as it passed a crash is a bit of a mystery.
However, Teslas have radar for this sort of things. On the other hand this does kinda underscore the problem of using cameras with global shutter control for self driving. The auto contrast couldnt cope with the oncoming lights, so stopped down the sensitivity of the camera, leaving it blind to that broken truck.
Found:
> Beginning with deliveries in May 2021, Model 3 and Model Y vehicles built for the North American market will no longer be equipped with radar.
from https://www.tesla.com/support/transitioning-tesla-vision, so radar-less is a new thing. That Tesla should have had a radar unless bought recently.
either way, having a different class of sensor is wise, to make up for corner cases in the main sensor array.
Tesla already have limited number of sensors, and they do most of the driving using monocular(single camera, or single viewpoint with three different zoom levels) inference. This is absolute madness, given the state of current tech.
For that night situation, I would be pushing for a different camera type. One that doesn't have a global exposure setting. This means that groups of pixels can be lumped together to have local ISO settings(its more complex than that, and involves integrating processing silicon directly under the sensor, on the same die.). This is a way to get vastly greater dynamic range in a situation where you have lots of point lights (ie headlights)
This would require developing and qualifying a new camera sensor from scratch (I don't think generic versions of this type of camera exist yet) Musk is unlikely to do this as he has signaled that everything should be done via ML.
Second, using a different wavelength, IR, or near IR.
Thirdly, higher wavelength radar. It gives you this information for free, they are cheap and already qualified for vehicles.
Another thing I don’t understand is why wont Tesla invest in a front only radar sensor to avoid such collisions.
Edit: *Specifically, to be on the lookout for a black sideways truck on an unlit road directly in ahead of where the car is traveling.
Maybe a human could have braked even a little bit, but the autopilot didn’t brake at all. That’s a problem for advocates for it
He has responded to so many night crashes where they have had trouble locating the vehicle, that he wanted to make the next emergency responder’s job easier by helping them see the flipped vehicle if that were ever to (unfortunately) occur.
[ ] "YOU are the beta-test of all features including collision avoidance when using this vehicle"
[ ] "YOU will be used to detect bugs before they can be fixed, including collisions at high speed"
Because these things are assumed in a non-computerized car but the opposite when it's "gee-whiz".
If your phone app crashes well you just reboot the phone. If the car has a bug, you crash and you reboot.
Wow. I don’t know how anything without night vision wouldn’t have hit that. It’s practically invisible.
I think it was a huge mistake for Tesla to remove radar from their vehicles but radar isn’t very good on stationary objects so not much help here.
This must have been terrifying to be there for.
[1] https://www.nature.com/articles/ncomms12172
I doubt I would have seen it or reacted in time.
There was a human in the Tesla that apparently didn't see it either.
- edit -
Sure... maybe a human could have "lessened the blow", maybe some humans, but now we're hindsight backseat driving and pulling things out of our ass.
As a human there's a lot that goes into it, it depends on how long you were driving, your vision at the time, whether you were talking, your driving skills, your reaction time, your handling of accidents, knowing where you can swerve to.
Some humans with full attention may have prevented it or weakened the crash, but that's a maybe, and assuming full attention.
Many people in this thread who think they would have noticed the subtle lighting effects on a pitch black pickup in the middle of the road. I've nearly hit blacked out cars on the road, you don't see them. You could go back and probably see it in a dashcam, but you know it's there then.
The uncompressed version you're still expecting a truck, it seems like it's much slower than what rounding a corner going 75mph on a freeway is, I don't think any of us could judge without being in the seat at the time.
All in all, this is the most speculative thread I've ever seen on HN. Very reaching, but anti-Tesla is hot news here so...
The big takeaway for me is how that Tesla just took that head-on collision at 75mph without compromising the cabin.
You might not have been able recognise it was a pickup, but you would have at least braked earlier.
Changing a 75mph collision into a 35mph collision is a world of difference in terms of the total energy.
I think it's important not to view "collision" as a binary event. With human drivers, it's pretty rare not to heal the squeal of breaks just before an impact. I suspect with automated driving (at least for a while) these kinds of "full speed" impacts will be more common than with human drivers.
When you watch the uncompressed version, there are tons of visual cues to a human that there is an object in the way. This is no where near as subtle as the janky video makes it out. You can say you might have missed this at night but either you don't pay enough attention when you're driving, or you're giving tesla a pass here they don't deserve.
But from that video this looks like the kind of thing an attentive driver could easily miss until it was too late. Maybe they’d start breaking but would they have stopped anywhere near in time? I doubt it.
Radar and Lidar work in the dark. But yes, this is a problem with optical-only, especially since the camera-systems Tesla uses are pretty badly spec'd (1.2MP, not even 1080p)
Tesla _used_ to have a RADAR system installed. But stopped installing them and rewrote the software to get rid of it earlier this year.
> ON Semi AR0136A with 3.75 um Pixel size at 1280×960 1.2Mp resolution
This is a pretty crap camera / sensor. The ON Semi documentation suggests its a 1/3" sensor size: https://www.onsemi.com/products/sensors/image-sensors/ar0136...
Your iPhone 13 has a 1/1.9" wide sensor, significantly larger... and multiples of them working as a team.
Even if we go back to 2018: the Google Pixel 2 has a 1/2.6" sensor, still much larger (and with better resolution) than what this Tesla is using to see the road.
----------
If you want better night vision, then use a bigger sensor. Its just basic physics. A sensor with 2x the area will capture 2x the photons.
Complaining that this car can't see the road at night... well.... yeah. Its got an tiny camera sensor. What did you expect? Its got bad night vision.
One of the problems is tenperature range. Driving continuiusly in Death Valley can heat up your camera to the point it runs away into thermal self destruction, if you don't take counter measures. This is not a problem for mobile phones.
Another issue is continuously processing uncompressed data. Mobile phones compress with hardware support, especially in video mode. I had one of my mobiles shut down during a long video conference. This must not happen to an assistance system.
The issue with RADAR was allegedly phantom braking, if Tesla marketing was to be believed. But other companies seem to have RADAR systems that work.
------
Its fine to use 1.2 MP 1/3" CMOS sensors as part of a general-purpose sensor network (which may include RADAR, ultrasonic, and cheap Cameras all working together). I think LIDAR is still needed to really make things work, but RADAR + Cameras seem to cover a wide variety of cases.
What's a problem is when Tesla starts removing RADAR and pretends its an advantage. (Look, they couldn't figure out the RADAR false-positive problem, and there's a chip shortage right now anyway. So they just... stopped using RADAR)
And the phantom braking hasn't improved without radar. As with so many things at Tesla, this was clearly a cost-cutting or supply-constraint inspired move masquerading as some technological advance.
Have you ever tried to take a picture on your phone at night? Or even dusk? Things that are clearly visible render as 'practically invisible.' It's a safe bet that this video isn't a good representation of how visible the thing is, and it's guaranteed that the truck was more visible than it appears. The question of how much more visible probably isn't easy to answer other than "much more."
It has issues during the day, at night it is downright dangerous.
I do agree that while watching the video the stopped truck is not visible till very late and based just on the video it is not clear if a human would have done any better. However, I'm not sure we can make that call since it's not clear what the quality/aperture of the video is and what the human eye was able to see.
Still, even then, a human would have slammed the breaks or swerved to the right to a free lane and depending on how good the Tesla breaks are, I don't own one, it may have stopped/slowed down substantially.
EDIT:
I think some people are underestimating their reaction time. I have had a case, once in my life, where I was driving on a Freeway at ~65 mph, 87 N for those from the Bay Area, and due to the curving of the Freeway I did not see the traffic stopped in front of me in the left most lane. It was very similar to this, but while the freeway was curving. I did not slam the breaks, but swerved to the right lane, which very luckily for me was empty till I could slam the breaks. It was purely reflex action driven, no thinking was involved.
In this case there are two things:
- human eye can see farther at night than a camera as others have also noted.
- swerving would/could have prevented a head on collision since there was no traffic in the middle lane (though the stopped truck is occupying a bit of it) and would have hit the Tesla on the left side. Much better than a full head on collision.
Thank you, it seems I'm not the only one. My foot hurts much more when I'm using adaptive cruise control since it's always hovering over the brake but not touching it - unless there is absolutely not traffic whatsoever. Due to this, I almost never use it.
Somewhat. The product costs $10,000 and has been enabled since 2015 (originally $5000). Its reasonable for people to think that this thing works. Sure, perhaps those individuals should do better research, but there's a huge degree of misinformation coming in from the top here.
The root-cause of the problem is Tesla. Its unreasonable for Tesla to pretend it works, charge significant sums of money for the "feature", and ignore the difficult questions when this thing crashes into stopped Firetrucks over-and-over again (https://www.consumerreports.org/autonomous-driving/nhtsa-saf...)
At some point, we as a society need to accept and admit that Tesla is __lying__ about this feature. That it is dangerous, and always has been dangerous.
There's a degree of blame we can assign to the Tesla-fanboys who accept this marketing without any level of critical thinking. But the root cause of the problem is Tesla's marketing.
Autopilot has been on cars for years, has driven multiple billion miles, and now comes free with every Tesla. In reality it is just lane and traffic aware cruise control which numerous cars come with today.
Full Self Driving doesn't have near the track record, is still in beta (and maybe should still be in alpha), and has that 5 figure price tag you mentioned.
When you blur the lines between the two, it also results in people thinking Autopilot is more capable than it is or that Full Self Driving is safer than it is.
I'd rather not use the crap marketing terms of Tesla. They're clearly misleading and just propaganda. You and I know that its emergency braking.
It doesn't matter if its on an intelligent-cruise control / lane assist mode, or if its in the hands of a driver. Emergency braking is supposed to kick in and work in these cases
Yes, this accident is a problem that Tesla needs to work to fix. But no system is going to ever be flawless and I would bet many humans and other manufacturer emergency braking systems would fail to avoid this accident.
In the meantime everyone, from Tesla fanboys to Tesla critics, should be mindful about how they talk about these things because confusion is dangerous regardless of its source.
https://www.consumerreports.org/autonomous-driving/nhtsa-saf...
Which is Tesla crashing into stopped trucks in the middle of the road. The fact is, we have a pattern here, and it is now documented.
As sad as it is, this Twitter-thread isn't "news". The NHTSA opened an investigation last year, but you know how slowly the government works. Everyone has known this.
There is a pattern here about Tesla's emergency braking, be it on the default package, the 2015-era Autopilot for $5000, or the modern 2021-era Autopilot for $10,000.
The emergency braking simply doesn't work (https://www.carscoops.com/2020/09/tesla-model-3-fails-an-aut...).
Are you just trolling with that comment. You are making the same mistake again that I corrected you on in your first comment. Autopilot is free with the car.
Emergency braking systems are not foolproof. Tesla is not a market leader in that category, but these systems from other manufactures also fail in similar situations[1].
[1] - https://www.caranddriver.com/features/a24511826/safety-featu...
We both know they're on the same sensor suite on largely the same technology. "FSD" may have more software and "AI" behind it, but the fundamental "is a stopped car in front of me" logic is the same neural-network training through the Dojo-supercomputer or whatever else Tesla wants to talk about. Its just a software switch that's locked behind a $10,000 paywall, nothing more.
But the customer, and the people involved in crashes, don't care about the names of supercomputers, the different features of the car, or whether or not its in Beta or not.
--------------
Can Tesla's sensor suite + computer see a stopped car in front of it? Well... no. It can't. Not only does this case prove it, so have many other cases.
That's why there's an NHTSA probe on this very issue. Again, here's the link: https://www.consumerreports.org/autonomous-driving/nhtsa-saf...
I care much less about a singular accident than I do about trying to help prevent future accidents by clearing up the confusion. You are continuing to add to that confusion, seemingly intentionally by continuing to blur the lines that I have already clarified in previous comments.
You already admitted that a product being on the market for years would lead to a consumer being more confident in it. By telling people that FSD has been tested for years on public roads rather than months your criticism is counterintuitively also contributing to a false confidence in the technology.
But we have an NHTSA complaint documenting 11 other accidents of this form. That is, a stopped car in front of a Tesla that was provably on Autopilot. And then the Tesla crashes into the stopped car (most commonly: emergency vehicles like Firetrucks or Police cars). EDIT: These other examples from NHTSA were in broad-daylight, with flashing lights, at night, and a variety of different situations. So something is definitely wrong here.
And who knows about all the other situations where a crash was going to happen but the driver took emergency action and saved the day.
-----
This accident, the one in the twitter post, as well as the 11 other documented cases in the NHTSA investigation, prove that there's a systemic problem with the technology.
There are probably other accidents too where the emergency braking should have worked but aren't documented yet.
------
This is no longer a Rusell's Teapot scenario. We've got 12-examples (11 from the NHTSA complaint, and +1 from this Twitter thread / Twitter video). We've got the examples, live on camera. We can absolutely say without a doubt that Teslas have difficulty in this case.
Autopilot has not been on cars for years, Adaptive cruise control has been.
Moreover, while I’ll accept that my comment above confuses between them, I’ll argue its naming it as “ Autipilot” is what is at fault here since people overestimate its capabilities.
Autopilot is a term most commonly used in aircraft and that's where almost everyone has head from it before. At the same time almost everyone should be fully aware that aircraft on autopilot do in fact need at least one human pilot. Almost all aircraft inevitably crash if left on autopilot with no further human interaction.
So where does this idea come from that the word "autopilot" somehow misleads people into believing that it does not require a human.
>I bet for most people autopilot means computer does driving.
That is the claim but there is no evidence at all. Everyone knows there is a pilot in almost every aircraft and that autopilot is just a computer that takes work away. If it would fly alone there would be no pilot.
There should be no people with a driver license who think autopilot means no human needed. Its completely absurd. That word never meant that and its not people misinterpreting the word that sleep in their cars and crash it. These people are not mislead by a word they are just irresponsible and fully aware of it or drunk.
Assuming wide spread misinterpretation is a bit far fetched especially for a word that the average teenager knows what it means.
When Tesla Marketing videos said "The person in the driver's seat is only there for legal reasons. He is not doing anything. The car is driving itself".
You can find this quote all over the internet. Since this is Hacker News: https://news.ycombinator.com/item?id=26858525, you can see people discussing this and proving that the video said that.
Quote end-quote. Tesla literally said that and posted it all over the internet. Elon Musk then came on after that video was aired and said they'll have "Coast to coast Full-self driving" in 2017, and repeatedly double-downed on that claim.
I have to imagine that your postings here have something to do with the accident that this entire topic is about.
But I can assure you even modern passenger plane on autopilot do not avoid crashes with other planes, buildings or ground (mountains etc.) It also wont start or land a plane.
In a worst case scenario the autopilot would crash the plane immediately because it was give the task to fly into something and in best case scenario the plane flies according to the programmed route and corrects for wind and all that and if nothing is in the way it eventually runs out of fuel and crashes. AFAIK it would not even slow down and certainly it would not do an emergency landing.
I understand your explanation of the term with regards to airplanes, but not everybody understands that. Even if the term ‘autopilot’ is technically correct, it might be better to just call it driver assist for now.
I’m not laying the blame on Tesla for these sorts of accidents though. Regardless of the name, I think Tesla makes it explicitly clear how the function should be used.
I do think there is some friction here between Tesla’s branding and marketing goals, and the long and complex road to safe autonomous driving.
This is a marketing page, directly from Tesla in 2016. It is clearly misleading and/or lying in several ways, and Tesla has yet to update their direct marketing on their direct website.
------
Tweets directly sent by the CEO: https://twitter.com/elonmusk/status/866482406160609280?lang=...
Are also marketing. And this was a lie of course, "coast to coast FSD" never happened, not in 2017, and probably never. It doesn't seem like the technology is capable of seeing stopped cars in front of it (let alone of a coast-to-coast drive without disengagements).
The CEO participates in marketing, on Twitter, on Facebook. Tesla markets on their website, and through their employees and service centers. There are clearly a ton of Tesla influencers who are also being paid to shill the car and their features online (much like any other car company, social media shilling strategies are very effective). There are many, many web-based marketing strategies Tesla has, from referral links, to search-engine-optimization, to blogging, to Youtube videos, etc. etc.
That Tesla doesn't participate in the traditional TV-ad marketing is a bit of a non-sequitur. They clearly have a huge online presence and leverage that to influence their buyers. If you've ever done SEO before, you'd clearly see how Tesla (and Elon Musk) optimize for SEO.
Yeah, Musk says that they don't do marketing. But since when was Musk a trustworthy source of information? Think for a bit.
Isn't that just basic, competent driving?
Also, if it's so dark that you can't stop before hitting a car in front of you, then you're going too fast. The speed limit is a max, not a min, and it changes depending on prevailing road conditions no matter what the posted signs say. For example, if it's raining and you get into an accident, you can be fined for speeding even if you were going the speed limit. Going 75mph in pea-soup fog is a recipe for disaster. The dead of night has the same calculus. You should only go as fast as you can see. A good rule of thumb is you should always be able to see 12 seconds clearly ahead of you. That's how some professional human drivers manage to drive literal millions of miles over a career without a single accident.
Frankly, all this talk about how Teslas will improve the safety of traffic because they perform at a superhuman level is belied by this video evidence. Apparently these machines only drive as well as the drivers who programmed them. I wonder if there are even any professional drivers on the engineering team consulting on this project.
If the car can't even operate within the envelope of its own sensing and computing limitations, then how can we make any claims about its safety? It's driving at top speed in situations where it's essentially blind. When I'm blinded, I pull over. When the Tesla is blinded, it doesn't even know. See also, the guy who ran full speed into a tractor trailer and lost his head. It's been years since that incident and Teslas are still driving full speed into things they can't see. How is that a basis for platform that will save us all due to how safe it is?
Also, a sideswipe is almost certainly preferable for everyone than plowing head on at full speed.
Take a guess where that truck is going to go after it gets hit at 75 mph. Probably into another lane, potentially into oncoming traffic.
Human reactions are poor. Houseplant reactions are worse. Tesla acted more like a houseplant than a human
I’m fairly sure this is not true. Modern cars have lots of energy-absorbing crumple zones in the front, plus airbags. Head-on is the most highly-optimized crash orientation. Side-swipes are I believe the worst case, because there is so little protection for the driver.
> till, even then, a human would have slammed the breaks or swerved to the right to a free lane
Most humans unfortunately do not have the experience of knowing what to do at the right time; they may swerve to the left here in this video and crash into a more dangerous object (another high-speed car or truck) and cause a more dangerous pileup. There is no right or wrong answer in this situation, we don't do drills (at least in US) that at least tell people what to do in bad situations like this. This may be a practical question but in reality, no one has the training or experience to deal with this _instantly_ and worse, in dark environments.
Also, most people is likely to be tired if they're driving at night such as coming home from work.
What this proves is that we need emergency lighting systems in the event of a crash. Like if an airbag has been deployed, lights around need to be visible to avoid this kind of scenario.
It would be awesome if we could provide VR training and use videos like this as examples to train folks.
This is the first time I've been introduced to this idea, and it now feels immediately obvious that we should've been doing this for a long time.
I’m sure other cars are doing the flashing lights now but few companies have invested as much in safety as Mercedes.
First, robust response to a crash. Maybe this is lighting (flood LEDs on the back of first responder vehicles), maybe this is emergency vehicles with crash shock absorbers (Scorpion Impact Attenuators is a well known brand) that sit behind the accident in lanes of travel to protect responders at the accident scene. Cars should attempt to detect this, but it should be expected this may fail. The crash absorber will take the brunt of the crash, but everyone is going home to their families with some tears shed by the insurers. This applies to all vehicles with traffic aware cruise control and lane keeping who can’t reliably detect stoped vehicles (which is all automakers).
Second, first responders (through dispatch) should be able to issue an alert to Google Maps and any other traffic aggregator to alert about an accident in a lat long “bubble” (think a TFR in air navigation, pushed out over ADS-B or a data link) with an expiration time. Vehicles would receive this event and temporarily apply a reduced speed limit in the area (one that the driver could override, but that’ll be logged in the black box when they push the accelerator past the temporary limit). This also provides alerting to the driver to be attentive about an upcoming event, just like Waze alerts you to an upcoming accident or speed trap. The vehicle integration is key, as the robotics is what needs the safety context and awareness.
Now, if you’re not paying attention as a driver, no amount of tech and effort helps, and these folks end up killing themselves and others every day with inattentive and distracted driving.
This is also vastly cheaper than LiDAR on every vehicle in the short term.
What's worrying isn't that the autopilot didn't "know what to do"; but that it didn't detect the oncoming parked car at all.
Also, most people is likely to be tired if they're driving at night such as coming home from work.
I'd say more like around 10 percent would be significantly impaired. Heck, I'll give you 20.
"Most" (meaning > 50 percent) is just off the page, and doesn't jibe with reality. Sounds like you're a bit too eager to give Tesla a pass, here.
This is not simply a parked car, this is a dark non-reflective crashed pick-up truck sideway on a highway lane that's CURVED at night. There are multiple variables all of which are bad; non-reflective, driving high speed (75mph) in a curved lane, and night-time.
Worse, this car's detection system is based on video camera only, it does not have LIDAR. That's where Tesla has a big problem with and it clearly shows here. Video camera-only system has blind spots and using auto-pilot or self-driving system with this only is not a good idea, period.
In addition, we're not supposed to rely on autopilot only regardless of how good anything is; this is all is supposed to be a complimentary tool just like every other safety feature within the car; the cruise control feature is not supposed to be used blindly on its own and excellent human drivers with cruise control can still hit that car.
There's no 100% foolproof algorithm or method to predict every single scenario.
If the driver was relying on the auto-pilot and some kind of detection system and didn't pay attention, then it is not the system that failed, it is the driver's fault entirely.
> "Most" (meaning > 50 percent) is just off the page, and doesn't jibe with reality. Sounds like you're a bit too eager to give Tesla a pass, here.
I'm not giving anyone a pass, I'm pointing out that humans have their own weaknesses and blind spots that machines can help with but humans on their own, will not be doing better in this specific scenario with all of these variables. I can actually understand why the system failed to do anything given the technology with limited sensor data but I don't understand why they didn't use more sensors like lidar, infrared or emergency breaking system or so on. I may be missing something here.
Regardless, we're not looking at a scenario where it happens every day, it only takes one bad accident and one tired human to turn an uncommon scenario into a disaster. So, it doesn't matter what percentage we deal with.
IMO it’s the fault of the engineers who did not take human factors into account. We all know that humans are terrible at following instructions or even using common sense — hence warnings about not using your toaster in the bathtub — so when engineers pretend that users will do the right thing, the engineers are at fault for not addressing the behaviours of the multitudes who will not do the right thing.
TLDR: the driver is at fault for being a dumbass using autopilot at night, and Tesla is at fault for not considering how real humans will use the product.
I think they know perfectly well, on some level.
As they are of the fact that large portion of their drivers will be perfect dumbasses.
But they're invested too deeply in the promise of an "Autopilot" product (which of course consumers will take to mean exactly what it says) by this point that they're not going to backpedal or change course -- unless forced to do so.
If I had a uninterrupted four hours of drive as part of my regular commute I wouldn't buy a Tesla, I would buy a gun and a single bullet with my name on it. Two hours one way every day were already enough to drive me almost insane.
I would really be interested in your source on this.
Anyone who's ever taken a photograph in low light conditions knows that in the dark, cameras fail to capture loads of detail that humans can make out just fine.
I'd wager a human driving in such circumstances would have seen the truck much more easily than we can see it in some dark, blurry dashcam footage.
In any case, our human eyes are actually really, really good. Far better than typical CMOS technology at capturing dynamic range and whatnot. Our Iris opens/closes depending on a variety of conditions, and our retina itself is 22 mm, far larger than typical camera CMOS sensors (but smaller than the top-of-the-line 35mm cameras).
Not only is our surface-area larger... our rods / cones seem to be superior in dynamic range than CMOS technology.
----------
That being said: I think a big problem for Tesla's system is that they use tiny and cheap 1/3" CMOS sensors at low resolutions, and then expect AI to magically fix all their problems.
Other companies are willing to not only buy more expensive cameras, but also more-expensive sensors, such as RADAR or LIDAR, to assist the camera.
A dark-adapted human eye may be able to detect light levels as low as a single photon:
https://www.nature.com/articles/ncomms12172
(Not reliably, mind you. But the fact that this works at all is incredible.)
Agreed; the human brain has special chemicals which significantly amplify our processing capabilities during moments of acute danger. I was once going up freeway, in the dead of night, and spotted a (darkly painted) car fully flipped over, about 2 seconds ahead of me. Managed to both decelerate (just enough) and bear gently to the right (not swerve), with room to spare.
Granted I was going 60, not 75. But I'm not the one telling people to drive on AP at 75, in the dead of night, no less
Are you using population-level accident statistics to make this assessment?
I’ve seen analysis that suggests the overall safety numbers for self-driving mode are better than human drivers, though I’ve not dug deeply enough to be confident in them.
I’d caution against extrapolating from individual incidents though. I think we should expect that at the point that self driving is safer than a human driver, it will be worse than humans in some cases - simply because it can pick up a lot of gains by never being tired, distracted, or drunk. If one just looks at the pathological cases and says “I could have avoided that accident, therefore self-driving is a worse driver than me”, I believe one would be making an error of reasoning.
There’s an interesting philosophical/legal question at play here. Does self-driving have to be strictly superior - ie no edge-cases where it is worse than a human - to be acceptable? Or can it be worse at some tasks if it’s killing fewer people overall? Many people will die if we are too strict here. And yet, people need to trust the system as well. I suspect this will be optimized based on expected costs of lawsuits, which is a poor proxy for morality…
We must demand better.
All similar analysis I've seen compared all-road all-vehicle driving stats against autopilot stats. Of course, the correct comparison would be against luxury cars (with similar safety features) and only on the same kinds of roads that autopilot gets used on (i.e. primarily easier highway driving). As far as I can tell we are extremely far away from the hypothetical you raise where self-driving is better than human.
I’m making a rather simple assessment. I’ve seen Tesla autopilot/FSD fail catastrophically in situations where I would have been fine. This does not inspire confidence that it will do better in situations where I would have failed. As such, I’m not willing to put my life in its hand. And, I really hope no effing Tesla hits me.
I always shudder when I hear that people drive too fast to be able to stop in time based on visibility. Same with the crash in the video. Way to high a speed for that visibility.
Swerving could have prevented the collision, but braking is always the safer option. Swerving would still require the object to be detected, which neither the AP nor the human did in this instance.
Regardless you are still correct that people trusting their lives to AP instead of monitoring it like they are nagged to do by Tesla software are being foolish. I would never have been in the inside lane in this scenario, most likely if using AP I would be happily cruising behind the coach, meaning in any collision with a stationary object the coach would absorb the bulk of the collision energy.
Making this incident about AP is putting the cart well and truly before the horse.
The video shows noone else in the lane, who are you saying avoided it?
That truck in the road was a recent wreck, its most likely the Tesla was the first to come upon it.
How do you know when the crash occurred? How do you know for sure other people managed to do just fine?
Watch the video; there was no other car in that lane for 10s at the very least.
Scary, but it happens all the time to humans too.
I'm more impressed how the Tesla looked at that 75mph accident, the cabin really was not compromised at all. Wow.
[1] https://www.consumerreports.org/car-safety/most-new-cars-hav...
To my knowledge there is no AEB system in production that could reliably avoided this.
The impact would 100% still occur but at a much lower speed exponentially reducing the extent of injuries to all involved.
Absolutely categorically NOT true.
It may not fit your narrative but Front Radar electronics/firmware filter out stationary object ONLY if they are outside the vehicle trajectory envelope.
The majority of false positives are due to overhead passthroughs which may fall within the envelope, if the radar calibration is insufficient.
EuroNCAP tests for AEB include both moving targets AND stationary targets. [0]
Even those low-cost Bosch front Radar assemblies in entry-level vehicles could have mitigated that impact.
[0] https://cdn.euroncap.com/media/17719/euro-ncap-aeb-test-prot...
The first challenge is that cars do not always travel in a straight line, so the "vehicle trajectory envelope" is ill-defined. At minimum, the radar needs to be steered (electronically) to match the anticipated vehicle path. However, the anticipated path does not necessarily match the actual path.
The second challenge is the size of the target. If you re-watch the video, you'll notice that the pickup is sideways, and appears to have collided with a sedan prior to the Tesla arriving. In other words, the radar target is approximately 2x to 3x the angular width of a normal vehicle.
If you filter the radar returns to eliminate wide stationary objects (to, for example, reduce phantom braking caused by overpasses), you necessarily handicap it when detecting a truck parked sideways on the road.
Putting on my ML engineer hat: One of the common themes in examples where Teslas fail to brake is that the target is stationary and very wide: A semi truck crossing a not-limited-access highway, a crashed pickup turned sideways in the lane, etc. etc. etc.
> EuroNCAP tests for AEB include both moving targets AND stationary targets
Your link shows only scenarios where the target is exactly positioned along the centerline of the vehicle under test.
Additionally, the targets are only as wide as a normal vehicle--and if that weren't enough, the tests specify the use a combination of radar absorbing material and corner cube reflector to create what is essentially a point target.
1 https://youtu.be/Qk4zIqOiOjc
Never did it not react to my late braking.
> his is the exact sort of collision that Automatic Emergency Braking (AEB) is supposed to mitigate, and there are dozens of cars of different makes that have it installed
Is it supposed to? And do they actually mitigate it?
1. Many people don't view the world through a shitty 480p camera with compression artifacts.
2. Almost nobody would have hit the pickup still going at 75 MPH. Even dropping 10 MPH at that speed reduces the energy of the crash by 25%.
3. As another poster pointed out, nobody else on the road has managed to hit that truck. So, either all the other drivers are incredibly lucky, or it is, in fact, more than possible to both see and avoid that truck.
If you're going 75mph, that should really get your spidey-sense tingling. At least let off the accelerator/cancel cruise control. Be prepared to brake. Once you're closer and lined up, it's harder to tell there's anything going on because you're no longer approaching a bend with headlights on the other side of the truck, so by then it's too late.
But this is, again, all hindsight talking. In the moment, it's much harder to assess how a "typical human" might have reacted, if at all.
I’ve spent a lot of time staring at the road thinking how I’d write a program to do it. Main thing I’d do differently is provide visual feedback of confidence. Having the autopilot disengage when it gets confused isn’t good enough, I want to know when the system is losing confidence /before/ it disengages, and I need it to let me know that it’s not 100% sure the next thousand feet of road is actually clear. Tesla seems to drive unperturbed until its classifier recognizes an obstacle.
In other words, it's a good thing I don't make this stuff. My car would be phantom braking every 10 seconds, especially at night.
Granted, I am a lowly web programmer with a hobbyist interest in compression, but I have thought about getting the comma.ai kit and learning how it works, trying to implement my own world modeling, but, sounds like a full time job :)
EDIT: as for the headlights, I don’t want to act like I could whip this up in a weekend or anything, but having objects flicker in and out of existence would at least subtract a few points from my confidence score.
I am similarly opposed/nervous around machine learning, seems impossible to debug weird decisions the black box makes, I would prefer a traditional, deterministic computer vision approach.
Basically I want to segment incoming images into a grid and do a cosine similarity score between each cell and its neighbors one frame ago, to detect which direction and how quickly all the textures are moving. So I think it should be robust to whatever the road texture is, provided its not pitch black like wet asphalt.
you want the opposite, if it's consistent, then it's more difficult to gauge it's motion just by finding similar zones. think on mouse on glass or metal, it's too uniform for the sensor to detect motion.
anyway: paper https://dspace.mit.edu/bitstream/1721.1/6298/2/AIM-476.pdf
lecture: https://www.youtube.com/watch?v=0BVZDyRrYtQ
usually circular motion is used, but linear works just as fine
Even more in a Tesla as the regen braking will slow you down a lot. I drive almost exclusively with 1 pedal now because of it and, while I don't know that that actually is safer, it feels safer because I'm slowing down waaaay earlier than I normally would in an ICE-mobile.
You can use where the headlights disappear and reappear to draw an arc of occluded space. You should be able to measure how quickly that arc changes to estimate the distance
If you're 1000 feet from an object and move a foot closer, it barely blocks anything else from view. If you're 5 feet away, moving a foot closer blocks a lot more of your view.
The problem is that relies on nearly constant backlighting on the object.
The real solution here is LIDAR, which makes this trivial. Tesla won't use it though, so they have stupid problems like trying to hack a way to determine how far away dark objects are.
Even if you could make an arc of occlusion (sounds like a D&D spell), there's way too much variability in headlight sizes/designs to accurately determine distance in the dark in the situation this video presents. Again, I am admitting a ton of ignorance when making that statement but that seems nearly impossible to me with just cameras.
Still, I don't understand why Tesla is so opposed to LIDAR. The more information the autopilot has, the better decisions it can make.
This literally feels like spy kids 2, this watch can do everything except show the time.
Obviously there are situations where cameras or our eyes simply won't be able to see enough. Dark unlit objects at night aren't exactly an obscure edge case scenario, the crash from the video is something we already see on a regular basis with human driven cars, sadly. And even if not, I'd rather pay $1000 or so extra to protect me from that freak scenario that's likely never going to happen. Until it does.
I also have airbags in my car although I never needed them.
Even if it is a real issue, the government regulators can simply set the max power limits lower for LiDAR units to ensure it's safe.
"So coherent is laser light that one may observe interference effects even when the path difference between the interfering rays is much greater than 109 wavelengths! ... If two waves are exactly “in phase” as shown in Figure 1a, they will reinforce each other; this is called constructive interference." [0]
[0] https://faraday.physics.utoronto.ca/IYearLab/intdif.pdf
A Tesla vehicle, when under human control, has AEB but when under AutoPilot it essentially does not since it is the same logic used for both processes and therefore it doesn't act as an additional safety feature.
By the way, as a roboticist this whole situation is completely baffling to me. We go through such great lengths in our research to make our robots safe. Safety is a huge concern to our community. How is it that this company can just ship beta quality software to customers for use in production amongst the rest of us? In industry, we paint our robots orange and put them behind impenetrable cages due to safety concerns. Entire facilities are designed around the robots to make their operation safe. Operators are trained to use them. People working around them are trained to work along side them and observe proper safety protocols. Proper insurance is carried in case of accidents. And when an accident does occur, the situation is taken apart, analyzed, and changes are made so they won't happen again.
Not in the marketplace though, it seems. Apparently it's okay to just code whatever you want; put the code on whatever machine you want, even if it's "beta" quality; ship it to whomever you want who can pay for it; provide little to no training; lie about the capabilities ("Full self driving" minus the full, and self, and driving. "Autopilot" that isn't.); and when people inevitably and predictably end up dead, apparently it's okay to just keep on pushing this software out without a shred of self reflection or humility. Because you're saving the world I guess? I just don't understand this culture.
aren't auto manufacturers that don't have features like autopilot currently adding things like lidar and automatic breaking to improve safety?
why wouldn't you add the thing that will make your system as safe as possible?
what's the goal here? some aspirational tech challenge that competes on arbitrary merits in some humans vs. machines challenge problem or building the best and safest way to move vehicles around?
They have since taken untold sums of cash from people (millions? tens of millions?) who they promised would get access to "Full Self Driving" with the hardware on board. Now they're in the situation where they cannot admit that the problem requires more hardware. The hardware is a fixed property here, a design requirement; they have to achieve actual full self driving with the hardware in place, or they're going to face refunds and lawsuits.
So now it's a delay game. I think it's clear now that short of AGI, Tesla will never get to "full" self driving with the sensor suit they have. Nonetheless they can't switch to a new sensor suit because that will admit defeat. Instead, they have to show continual improvement of the current sensor suite, while promising that "Full" self driving is just around the corner so they can continue the charade.
they since retrofitted the hardware twice since the claim was made, last retrofit being to the tune of 10k$.
You basically cannot see it until it's too late. No issue here, any other car on the road would of crashed into the truck as well.
https://streamable.com/7ixg7l
Same video, without Twitter compression https://streamable.com/7ixg7l
Watch it full screen on your desktop monitor (still wouldn't be as good as human field of vision) and remember this was recorded by an average quality dashcam, not a high-end camera.
My question - how did the driver do in this crash?
This isn't super clear cut, the vehicle was unlit in fast lane on the freeway. That could be a tough for a number of drivers.
That said, one lidar sensor maybe? Just to let at least brakes come on?