Video of Tesla on Autopilot Crashing into Stopped Truck at 75mph
twitter.com
twitter.com
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.
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.
[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.
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.
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
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.
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]
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.
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)
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$.
They are gambling on the ability for AI to "just sort it out". Even if the system can drive in the dark, you can bet it's mostly just looking at white lines 10 ft ahead. It's not looking for obstacles or anything like that, it can't even see them.
So dangerous. A complete black box where every second of driving you never know whether there's an obscure edge case where the car decides today is your final day and it drives off the side of a cliff because the image processor saw something it didn't understand.
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.
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
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.
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.
In this video, the reporter attempts to illuminate the car so that it is more visible.
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.
Think about the times you've taken pictures on your phone at night and the clearly visible background details render as nearly invisible on camera.
I’d bet that truck was much more visible in person than it is in the video.
I believe cameras and software can drive cars safely. But these cameras and this software can't.
So many questions
1> Why isn't the back corner of the pickup truck reflecting any light? Were the reflectors somehow tinted?
2> Was there still power in the truck? The hazard blinkers weren't on.
3> How sure are we there is zero latency between the UI and the frames as they roll in from the camera?
4> Does Tesla have a red team working to generate/test these type of corner cases? If not, why not?
5> Does Tesla FSD actively watch for mammal eyes at night in the country, like Deer, Moose, Bear, etc? If not, why not?
If I was driving in that lane, I would probably have hit it, if I hadn't noticed the occlusion.
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.
Shocking.
rock fall, tree, building.
Given that the road is occluded, this needs to be in the training data, if only for autonomous lane following.
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?
Also, how hard is Tesla training their object-detection models for things that are only an intermittent silhouette against opposing traffic? I'm guessing not at all.
[ ] "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.
This literally feels like spy kids 2, this watch can do everything except show the time.
Second, in the video the range of the low beams on the tesla seems absurdly short.
Our eyes have much better dynamic range so a person would have much better grasp on what the road is like behind the wheels than what they can see from the recorded video.
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.
I managed to get on the brakes but slammed into the vehicle still.
(Was in my 20's, so still youthful reaction times, drove lot, did SCCA events so at least an avg driver)
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".
I know most people would of probably hit that truck. Don't blame Tesla here. The uncompressed video is slightly better, but unless you get lucky your still hitting that truck.
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.
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.
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.
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.
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."