A Tesla vehicle using ‘Smart Summon’ appears to crash into a $3.5mm private jet | Hacker News Reader

A Tesla vehicle using ‘Smart Summon’ appears to crash into a $3.5mm private jet

theverge.com

193 pointsaaronbrethorst197 comments

summon is the wildest energy. its like handing your keys to a feral racoon on bath salts.

i once had to file a claim after a tesla on summon backed out of its parking space and directly over my motorcycle, crushing it against a barrier wall with such unrelenting force it pushed the cylinder head covers into the gas tank. this claim was a real tit spanner because the owners cut-rate insurance company flat-out refused to cover it since the driver wasnt in the car. it took nearly a year of subrogation but mercifully they settled.

I wonder if they'd refuse to cover damage from their client's car rolling down a hill in neutral, hitting another car, only because the owner wasn't in it.

Well that's disturbing. It doesn't bode well for how they are building their self-driving systems. A simple set of ultrasonic reversing sensors would be enough to prevent this.

The part of the jet it hit was in the air, the ultrasonic sensors are the main way it avoids hitting non-car objects in Smart Summon, and they have those all around the car.

I feel like there must be some way to detect and avoid objects that occlude space the car intends to occupy but that also don't touch the ground inside that footprint.

Maybe a lidar would've prevented these types of issues.

But how often is a car at risk of not clearing something in a parking lot? I don’t think they should be adding cost (manufacturing complexity, material/part cost) to specifically handle driving around a runway.

there are no rules about what kind of objects can exist in a parking lot. if they can't account for hanging obstructions, which are actually quite frequent everywhere, they probably should err on the side of "less likely to kill people" and just not offer it.

Hitting a partially lowered garage door seems the most likely.

Not a Tesla, just a plain old fuckup on my part - I backed into a fan box mounted about 3’ off the ground in a parking garage. Was too reliant on cameras and the backup radar- didn’t look and boom - nice big crease in my rear panel. So yes this sort of thing is not some sort of crazy edge case only encountered on airport tarmacs.

Other vehicles can be pretty far off the ground sometimes, and while I don't know if summon can avoid those, autopilot sure doesn't.

I feel like if you need to think about a case like this for its own sake then there may be some more fundamental problem with your approach.

I bet it'd work great on Mars, though. Not a whole lot of airplanes there, now or in the foreseeable future.

More flying vehicles on Mars than people.

Signs, banners, forklifts, scaffolding. Lots of things that humans would know not to hit, and automated cars should know not to hit.

Tree branches.

Probably not outside spaces, but here such thing as ventilation shafts in parts of underground parking infra are not exactly rare.

Jet crushed the upper windshield right in the spot where Tesla hides its monocular camera cluster.

Tesla is known to play fast and loose with their driver assist and autonomous vehicle tech and insist on avoiding lidar. Waymo and Cruise are a lot more conservative.

As a society, we have not figured out who is liable in cases like this -- the owner or the car company that engineered the AI.

It borders on reckless than Tesla isn't using LiDAR or a similar sensor.

Relying on images only requires you to perfectly identify every object in the scene in order to infer its bounding box. And that's an issue like we see here where you don't expect to see private jets near a Tesla and so it may not be modelled.

With LiDAR, Radar etc. at least they know "something" is there even if they don't know what is.

You shouldn't need to identify objects using images, just that the parallax shift indicates something there.

IIRC their cameras are videoconference grade webcam FHD sensors. Lots of noise in that signal.

How much would a LiDAR add to the cost of a Tesla?

I don't know about the cost of the hardware itself, but there is also the cost to license the tech, with different lidar companies attempting grab a piece of the patentable space for lidar.

https://www.theverge.com/2021/5/24/22451404/tesla-luminar-li...

https://www.patentforecast.com/2021/07/22/lidar-competition-...

https://www.patentforecast.com/2021/09/02/waymo-is-driving-l...

Well my Robovac cost $300 and has LIDAR. Not the same grade, but it’s not like it’s fundamentally expensive. You could use RADAR too; sensors are not always expensive and when produced in bulk costs go down.

> As a society, we have not figured out who is liable in cases like this -- the owner or the car company that engineered the AI.

For the most part none of this is novel to the law.

So long as the owner didn’t do anything negligent themselves, then it seems a strong claim of product liability (defective product). It seems there is very little chance of contributory negligence of the owner with this function, unlike crashes in self-drive mode, where the driver was supposed to pay attention and maintain their hands on the wheel, but turns out they were sleeping or in the backseat live streaming.

I do think there should be legislation adopted to protect the victims by giving them a statutory private right of action against both the owner and the car manufacturer to hold them jointly and severally liable for any accident involving self driving vehicles, then in practice the owner/manufacturer (i.e. their insurance companies) to sort out liability.

I've read that Tesla's summon feature requires the user to continually hold down a button on their phone in order to have the car continue its slow progress in driving towards the user.

Assuming that's the case, I would argue that this specific sort of accident is squarely in the domain of the user.

Can confirm that this is actually how it works. As soon as you let go of the button, the car instantly stops.

Not even mentioning the explicit warnings to keep line of sight when you use smart summon, as well as the fact that summon is automatically limited to only within 20 meters of the car.

What about the copy on Tesla's website that says "have your car come to you while you deal with a fussy child"?

Do you happen to have a link to Tesla's webpage that says this?

Asking not because I don't believe it, but because the only results I get on google by searching for that quote are your comments on HN from this year in multiple threads that include it.

If the Tesla website indeed says it (which I am yet to find), I would write it off as just future-promise-marketing. You can call it misleading in that case, which would be fair.

At the end of the day, that's not how it currently works, and you have to hold the button for the car to move during summon, and you gotta be within 20 meter radius. As soon as you let the button go, the car stops. This isn't me just saying based on what internet says, this is what I have observed myself and what the car manual says[0] (both the online and the print version, in multiple spots).

0. https://www.tesla.com/ownersmanual/model3/en_us/GUID-6B9A1AE...

> the only results I get on google by searching for that quote are your comments on HN

It's interesting how Google narrows your world view.

> If the Tesla website indeed says it (which I am yet to find)

Here you go:

https://www.tesla.com/blog/introducing-software-version-10-0

> I would write it off as just future-promise-marketing

No, it's just lies and advertising.

Imagine holding a child in your arms and pressing a button with your fingers. This is basically a low bar to jump over in multitasking/human dexterity, however it’s the core value of summoning your vehicle.

You may say multitasking is hard or impossible but we humans manage to approximate it anyways because it’s necessary for life. For example, if you are the parent of two fussy children in a busy parking lot, you better know how to multitask or one of those kids will escape your sight!

Oof that's right up there with

"The person in the driver's seat is only there for legal reasons"

from the 5+ year old marketing video still featured on https://www.tesla.com/autopilot

I’m not sure it’s as cut-and-dry as you suggest. Suppose you have to hold down a button to keep a meat-slicer blade moving. Then, as you slice a piece of meat, the slicer blade suddenly flies out and strikes a nearby customer and injures them. Later, it’s discovered that a faulty design or a manufacturing defect cause the blade to fly out. There’d still be a very good case that the manufacturer would still be liable for the defect.

Now, here, maybe the user could have prevented the car accident by letting go of the button. But, what if the user had faith (however misplaced) that the car was going to avoid the collision until it was too late to let go? Or what if the user had suffered a stroke at that particular moment and couldn’t let go? It’s an interesting thought experiment.

It’s not an interesting thought experiment. The finger on the button is no different than the toe on the gas pedal.

It is different: in a normal car, the driver controls the angle of the vector via the steering wheel, but with the Summon feature, the car’s programming does.

Although there are consumer protection laws that puts the onus on the company -- in America, where it is caveat venditor -- there is also the question of how much decision-making power we want to give to automation in general, and who is ultimately responsible.

Some things, like the decision to trigger a weapon in an act of war may be something societies want to have humans, and only humans make the ultimate call. That is, there is a desire for a human-in-the-loop.

On the other end of the spectrum are decisions we want to take out of the human hands because we think automation is more reliable and better for our society.

And then there are the unintended consequences. Looking back the past 10 years, it's arguable that the 'share' and 'retweet' features of social media that gives the power of virality to ordinary people have had a profound effect on our society and culture, both for good and for bad (and for many cases where we can't even agree on what's good or bad).

I argue that, while some of this is not novel to the law, I don't necessarily think that those precedents cover all of the changes tech has wrought to how we live and govern ourselves.

>I do think there should be legislation adopted to protect the victims by giving them a statutory private right of action against both the owner and the car manufacturer

If my car is parked in a parking lot, and someone rams my car so that it rolls and hits another car, I am liable. This is why parking breaks are important

I don't understand why self driving cars would be any different. I am liable to make sure I am using my car properly. Even if it has fancy tech.

The person who hit your car and caused it to roll is also liable to the person who owns the car yo damaged.

Tesla says that the owner is supposed to be able to see the car during all parts of the summon procedure precisely to prevent situations like this. Tesla knows that this system is not perfect.

Fwiw, the best use of summon is simply getting out of a tight parking spot (e.g., some residential garages).

Expecting it to be a valet is ambitious in its current state.

Tesla also says that "your car will come to you while you deal with a fussy child".

So which is the definitive statement?

There seems to be a very very clear disconnect between Tesla car's actual capabilities, their marketing claims, and the stuff they say to cover their asses legally.

Disclosure: I own a Tesla and I love it. I wasn’t certain until I drove one, and then I knew immediately that it was the car for me (and it has been great).

That said, some of the Tesla marketing, like in the example you provided, is very far off the mark. I really wish they wouldn’t do that, since it makes people question the other things marketed about the car that are (more or less) accurate.

Fwiw, the summon feature is a nice, minor convenience. I would even say that it is well-implemented for simple use cases. That said, the existing feature does not match the dreams that Tesla is currently selling (imho).

The card could have realized it hit something based on the engine load.

I find disturbing how the car just keep going... This is not a sensor issue.

The owner controls it by holding the button. Like a gas pedal.

Your comparison doesn't hold.

Tesla stops when it sense a imminent collision even if you have your foot on the pedal.

But, based on the incident, it seems to rely on the cameras and radars. Not on engine performance data.

Safety mechanisms don’t remove liability. If a car doesn’t stop of it’s own volition to avoid a crash while the driver continues accelerating (and most cars don’t), the responsibility is still with the driver.

Fair.

> this claim was a real tit spanner because the owners cut-rate insurance company flat-out refused to cover it since the driver wasnt in the car.

I thought the official dogma was that when self-driving cars were in "auto-mode" the manufacturers would be the parties held responsible/liable and pay for any damages cause by their systems.

forced arbitration has entered the chat

Only in the US, in Europa that wouldn't pass consumer protection laws.

How can they force arbitration with a third party?

> I thought the official dogma was that when self-driving cars were in "auto-mode" the manufacturers would be the parties held responsible/liable and pay for any damages cause by their systems.

See my response here:

https://news.ycombinator.com/item?id=31129075

tl;dr — Tesla states the driver is responsible when using this feature.

Obviously Tesla is just trying to side-step liability for yet another potentially dangerous self-driving feature. It might even sound reasonable reading the disclaimer, but its not like regular self-driving mode where the driver is supposed to be behind the wheel, alert and ready to take over/manually override the system. At least in that situation a driver has at least the minimal training to receive a license for the activity and probably has general familiarity with potential dangers and hazards. Whereas people aren’t trained/licensed to operate their vehicles remotely from outside the vehicle. Sure it might sound dumb, because any observer can reasonably see a potential accident, except that a reasonable person outside the vehicle probably expects the Tesla to also identify the same object the observer sees and avoid it. More specifically, as in this case, the driver probably had zero clue the Tesla was unable to “see” the plane engine because why wouldn’t a Tesla be able to see an object at the same level as the vehicle (a obvious “blind spot” that I’m sure Tesla was well aware of, but didn’t exactly go out of their way to advertise/warn their car owners).

>this claim was a real tit spanner because the owners cut-rate insurance company flat-out refused to cover it since the driver wasnt in the car

WTF, I'm against self-driving cars now. Glad you're all right.

Even if the insurance refuses to pay, the tesla owner would still have been liable. You'd have to go after them directly.

Lawyer fees to sue over a motorcycle would be more than the cost of a new motorcycle, unless it's a Ducati or something similarly car-priced.

> since the driver wasnt in the car.

That's shocking. What does that have to do with liability? Hypothentically if my car's brakes failed while my car was parked, my insurance should cover any damages.

I’m surprised that it took so long.

Tesla says that the car should be in sight during the whole summon procedure precisely so things like this don’t happen.

https://www.tesla.com/ownersmanual/model3/en_us/GUID-6B9A1AE...

“You must maintain a clear line of sight between you and Model 3 and closely monitor the vehicle and its surroundings at all times.”

“Smart Summon is a BETA feature. You must continually monitor the vehicle and its surroundings and stay prepared to take immediate action at any time. It is the driver's responsibility to use Smart Summon safely, responsibly, and as intended.”

I’m surprised that this was not a slam dunk for your insurance company.

I mean, the reason the self-driving community is furious at Tesla (well, Musk) is because they publish half-assed features like this and then push on the responsibility to the owners, and you'll always end up with idiots.

(I'm sure the engineers at Tesla are just as dissatisfied about the state of things, but Musk is relentless. I'm glad I chose not to work there)

The main reason isn't that they push responsibility to the owners, it's that they're risking getting the entire sector crushed just because Elon needs to prop his lies to keep the stock up.

> “In the U.S., things are legal by default, and in Europe, they are illegal by default,” Musk said last month during the opening of Tesla's factory near Berlin. “We have to get approval beforehand, whereas in the U.S., you can kind of do it on your own cognizance, more or less.”

This sums it up. Elon is gambling the entire autonomous industry on the hopes that these crashes never go too viral. That it's never a slow news week when FSD happens to drive like a blind teenager and cremate the driver and murder innocent bystanders.

If this ever happens it's the end of autonomous driving. The regulation will be so punitive and the hoops so plentiful that it will be over, for everyone.

Sounds like a win to me. We can move straight to Elon's nightmare scenario of trains and high efficiency two wheelers.

I'm so glad autonomous cars are a scam, because the price its proponents want to impose on everyone else to make it a reality (legally enforced non-ownership of a car you bought, even further expansion of the road network and even greater sprawl, a privately owned panopticon, banning other road users, even proposals for mandatory self driving car beacons on pedestrians, and guarantee there'd be a 'premium subscription' down the line to go anywhere other than a major chain for dinner) is far too great.

“ you'll always end up with idiots.” - so hardcore ‘self driving’ cars believers, the people that actually put their money where their mouth is, those that purchase and are eager to use and test the Autopilot and the FSD are - and using your word here - always idiots? Like pipe dream religious cult world saving idiots?

And yet if you read their marketing material:

"Have your car come to you while you deal with a fussy child."

This is Tesla to a tee. What marketing promises, the fine print take away.

>between you and Model 3

Ugh, why do they insist on this weird corporate speak?

That's Tesla trying to absolve themselves of any blame. It has nothing to do with getting the Tesla owner's insurance to pay.

Their insurance says "driver wasn't driving, not covered." Meaning the Tesla owner is still to blame, but you need sue them directly; meaning your insurance needs to sue them directly.

Tesla sounds like the guy who takes money to let someone cut in line behind them.

That sounds very weird. Summon is based on the ultrasound sensors. It will refuse to approach curbs and cones and some speed bumps. If anything it's tuned too tightly to be actually useful in practice (it also has no visual navigation, it just tries to head in roughly the right direction at 3-4mph, it's basically a toy). It's very hard to imagine that it missed something as big as a motorcycle.

In the linked case, the car's windshield hit the prodruding tail of the jet, which was maybe 1.5m above the tarmac. I'd be more willing to believe that this is an edge case of the sensors (they aren't tuned for overhangs, and I've heard of people scraping garage doors for presumably similar reasons). But... a motorcycle, really? Are you absolutely sure it was on summon?

Likewise I don't understand why you were dealing with the other driver's insurance company anyway. This is what you pay your own insurer for; making you whole without having to do the work yourself is literally what you pay them for.

There are two forms of "summon". There's regular summon, where you have to hold down a button to make the car go forward or back. It does turn to avoid obstacles, but with a very nervous energy. Regular summon is super slow, 1-2 miles per hour. If you let go of the button the car stops, but not as quickly as you'd think. It takes about a full second.

Smart summon does have wild raccoon energy. In this case, you need to be within ~200 feet and the car uses its GPS location to plot a path to you. It will back out and turn around obstacles. But it goes kind of fast. Uncomfortably fast. I had my car whip around a turn and nearly hit a parked car. I don't think I've ever seen it work well. Or at least it has always ended with me letting go of the phone kill switch. Which again, takes around a full second to stop the car.

Like someone else commented, this is a very different system than "full self driving" (16 year-old that just got their license energy), which is also a different thing than "Autopilot" (works great on highways).

1 full second seems a lot of time, even at a very low speed, of 3.6 km/h or a little more than 2 miles/hour, that is 1 m/s, i.e. the car will travel a whole meter in that second of delay (to which you may actually need to add some delay for the reaction of the user), I would have thought that releasing the button would result in a nearly instant "emergency braking", after all there is no one in the car and even if it stops abruptly it wouldn't be an issue.

And why should it be insurance company's problem if the driver weren't in car operating it? They are still themselves liable and can be sued for damages. Which I think is the more correct route. My and other insurance payers shouldn't be responsible to funding Tesla's fuck-ups. Or their owners idiocy and laziness...

Surprising lack of technical knowledge here about the current state of Tesla self driving.

Tesla current has 3 wildly different software stacks:

-FSD (invite only beta): this is their latest and greatest that everyone talks about. Surface streets only at the moment.

-Highway autopilot: legacy autopilot stack, AFAIK the next big push is to merge the FSD system to take over for highway and surface streets. The existing/public autopilot is years behind the existing FSD closed beta code (again, to the best of my knowledge)

-Summon: Ancient tech demo held together with duct tape and bubblegum. It is most definitely not using the new FSD stack and is way behind, since it’s a gimmick feature with low utilization they don’t seem to be in a hurry

IMO Tesla should kill the legacy stacks and push hard to get the current (actually quite good) FSD beta stack out to all scenarios.

Can’t edit anymore: for what it’s worth the Tesla in the video was operating on the “summon” stack

I wonder all the different things that needed to fail for this to happen.

Smart summon requires someone to hold the button down as they use it. Did they not pay attention or did this fail to stop the car after they let go etc.

Why didn’t the car react to the aircraft? Did it not detect it, or did it fail to judge it’s location?

Did the car bit notice the collision, or is that ignored by this feature?

The Hold the button down is one way of using Smart Summon and it's not the default way. The default way has the car moving forward until you hit stop or it detects something in the way.

I think that's just regular 'summon' where it'll only move straight forward or back, which very well might be what's happening in this video, but "smart summon" does require holding a button down on the phone. The original reddit comment[0] only said "someone tried to summon their Tesla".

0: https://old.reddit.com/r/flying/comments/u95dqt/someone_just...

"Hold the botton down" IS the default for both summon and smart summon.

There is also (and has always been) a shit ton of warnings you have to ignore to use the feature. [1]

Here's a particularly applicable warning from the manual.

> Smart Summon may not stop for all objects (especially very low objects such as some curbs, or very high objects such as a shelf) and may not react to all traffic. Smart Summon does not recognize the direction of traffic, does not navigate around empty parking spaces, and may not anticipate crossing traffic.

I know a lot of people like to dig at tesla, but they really aggressively warn drivers when it comes to using the driver automation.

[1] https://www.tesla.com/ownersmanual/modely/is_is/GUID-6B9A1AE...

For sure. But warnings don't solve complacency. You summon your Tesla to the front of Whole Foods successfully for months, go to the airport and expect the same thing. They think "oh these warnings are them just covering their butt." Which the warnings are doing. It shouldn't be called just "summon" it should be called "remote control summon." Because everything about those warnings puts you in control of the situation. But Tesla loves to use branding and will take the hit for when that branding doesn't actually meet its metrics.

This idea that as long as Tesla puts a warning on it, they have no responsibility is insane. They know it's broken and dangerous, they just don't care.

There are plenty of dangerous products sold without or with light warnings.

Guns, knives, wood chippers, chainsaws etc.

Summon/Smart summon can be used safely if you monitor the car.

Tesla tells you that if you don't watch the car, you can cause an accident.

Yeah, what's the point of making anything safer then. Just throw a warning label on it and call it a day, right?

Yeah, we need to ban the sale of bleach, knives, draino, aspirin, cigarettes, alcohol, hammers, nails, saws, shovels, tractors, semis, trains, cars, fishing poles.

After all, we can't make any of those things 100% safe so they are obviously completely broken.

Tesla's smart summon operates at under 5mph. It's not perfect, but it's also not fast. If you are using it in a situation where you can't see the car then yes, you are at fault for using a tool in an unsafe manner.

Can it be improved and made safer? Certainly. Should it? Yes. Does it need to be completely shut off until it is? No. It can be safely operated by responsible car owners. Until we start having major issues with smart summon damage I see no reason to kill it. A rich person ramming a private plane is not a "major issue" any more than a nascar wreck is a "major issue" for automobiles in general.

Most of those have danger intrinsicly with how they function, making it safe mostly means not using it. Many of them are also are largely only unsafe to the person using them. Those unsafe to the general public have a _large_ amount of regulation and testing around them.

Tesla's summon thing is unsafe because they can't be assed to take the time to implement it well and QA it appropriately. Summon isn't intrinsicly unsafe, they just have a half-assed implementation of it. Like many other of their features, they're too interested in pretending it works better than it does, and just throwing a warning in some doc somewhere as cover.

Also really all of them are dangerous to the Tesla owner/operator, where an argument could be made that they're accepting the risk, but _also_ to the general public.

The general public should not be the victims of Tesla's shitty, unprepared beta test(s), especially not just to satisfy Musk's impatience.

Tractors on public roads get to skip out on basically all safety testing, yet one recently killed a friend. They are more dangerous than Tesla summon both in deaths per year and deaths per mile yet the general response response from regulations is meh. Most Pollution gets the same free pass until something extreme happens, it’s simply the default to allow just about anything new until it proves a significant issue.

> yet one recently killed a friend

Very sorry to hear that, that's horrible.

> They are more dangerous than Tesla summon both in deaths per year and deaths per mile

I don't think deaths per year is that compelling as a metric for this. Teslas are becoming more common, so it's going to get worse.

Deaths per mile, where are you getting data for that? I mistrust anything from Tesla, they have a history of coloring the truth, at best.

Anyway though, I fundamentally disagree with this entire thought process of "X is worse, therefore we can't fix Y". Let's just fix them all!

> Why didn’t the car react to the aircraft? Did it not detect it, or did it fail to judge it’s location?

Because these smart vehicles are dumb as fuck when you put them anywhere other than on a well maintained, sunny, first world country highway. Not so long ago teslas braked on the highway when going under signs because they thought the shadow was an obstacle. A few of them crashed in overturned trucks because they though their roof was the sky/a cloud, one even missed a damn firetruck and smacked into it at full speed

When you have thousands of cars out there your one in a billion edge case will happen, this is the perfect example. A ten years old would have braked, the tesla didn't

I think you overestimate a 10 year old driver - even a 15 year old on their learners might not realize that z height matters, unless they're learning on a tall vehicle that can't fit under some drive-thru overhangs.

The plane didn't even clear the hood, safe to assume that if an obstacle is at eye level you won't make it under.

The roof should be around a foot if not more above eye level. Eye to top of head, headroom, thickness of roof.

The roof yes, but the >hood< is under eye level

> even a 15 year old on their learners might not realize that z height matters

There are a fair share of adults that do not understand height while driving either. Look no further than the 12'4" (previously 11'8") bridge on YouTube.

> Why didn’t the car react to the aircraft? Did it not detect it, or did it fail to judge it’s location?

According to https://www.tesla.com/ownersmanual/modely/is_is/GUID-6B9A1AE...

> Smart Summon may not stop for all objects (especially very low objects such as some curbs, or very high objects such as a shelf) and may not react to all traffic.

The tail of the plane was quite high, so smart summon probably wasn't trained to recognize it.

So we can’t even get an AI to recognize basic obstacles in its path but self driving cars are coming any day now ?

I am puzzled at how a system with such basic bugs is rolled out to production. Even obstacle detection hasn’t been solved yet. Yes it might be hard but an 8 year old in a GoKart would do better than this.

I would expect some sensor to detect some kind of impact/resistance and just stop. I mean if my robot vacuum can do it, I'd think a machine capable of killing someone would too.

But maybe that's a feature that'll come with their Level 4 autonomous software that'll come out very soon now /s

> Why didn’t the car react to the aircraft? Did it not detect it, or did it fail to judge it’s location?

This is a design flaw in Teslas [0]. The radar it uses has trouble seeing large stationary or slow-moving objects. Tesla refuses to use state of the art safety technology while developing these autonomous cars ostensibly in the name of safety. Not one, but two people have lost their lives because a Tesla failed to see a tractor trailer and collided with it [1]. I would have thought after the first time that happened it would be the last, but no. Tesla didn't have to change a thing about their design, and they are still allowed to beta test these things among the unwitting general public. Most everybody hasn't agreed to be part of Tesla's beta program, and yet we're subject to these machines with flawed sensor systems that Tesla just refuses to update to a safer technology. Why? People have died because of this, they're causing damage all over the place, and all they want is to put more out there as fast as possible. Why haven't they fixed the flaw that killed a man in 2016, and then killed a man the same way in 2019. How many bodies and how much damage exactly is enough before Elon Musk is forced to fix this thing? It's absolutely ludicrous these cars are allowed on public roads.

[0]: https://www.msn.com/en-us/news/technology/why-teslas-keep-st...

[1]: https://electrek.co/2019/03/01/tesla-driver-crash-truck-trai...

This is why you should use radar and/or lidar, not just cameras...

are you saying obstacle avoidance is impossible with multiple cameras (yes probably not as robust)? Is it not possible to tell the difference between an image artifact / shadows and actual solid object?

Much harder. Animals manage it, most of the time.

Clearly they need to add whiskers to all teslas.

You should patent that idea. Might work better than "AI".

Smart Summon does use radar. Last I checked it's disabled on vision-only Teslas. I wouldn't use until it uses the new FSD stack, because even with radar it is not good.

Tesla has radar but it is disabled? why is that?

Removed from newer models due to part shortages / cost reduction. The software function removed from models with the hardware for consistency / to give the false impression that it wasn't a useful sensor and that the cars are not going backwards in capability.

Because those radars are low results and lead to false positive. The achieved better results with camera only.

Tesla bragged radar could bounce under the car ahead to react to the second car ahead stopping, even though the second ahead couldn't be seen. Removing radar was needed due to the electronic component shortage.

Musk mentioned this on twitter a number of years before, but it turned to not to be that practical with those radars.

That they were planning to remove radar was clear a while before but the electronic component shortage made it a higher priority problem.

So with camera only how do you determine if a paper bag on the highway is empty or full of metal objects?

The problem with radar is it can’t see up high right? Just low to the ground where the unit is?

Based on this video I’m not sure the radar would’ve had anything to catch. It seems like the plane hit the top of the car, not that the car hit the landing gear.

Without even knowing how their internals work, I can guarantee that the bug in their software is that they don't check for height clearance, because it's not needed in 99.99% of the situations.

However, I do think radar should have caught it, because radar has almost like a cone like view in front. If the car was far enough, the cone would've spotted the height.

Radar can see wherever it's pointed. Point it up and it'll see up.

Given. But I don’t think that’s where normal car radars are pointed and as far as I know no one has an aim-able radar in a car. They all point to look for a car dead ahead on the road. Not an ultra-low bridge (or landed jet).

Maybe. I honestly don't know. But I can't help but wonder what's going to happen at a toll gate, at a railroad crossing with the arms down, after a storm and a branch is half-broken and suspended over the road, etc. Doubtful any of those things would have a radar return worth mentioning but if nothing is looking slightly up, folks are going to have a bad time.

How well is a front-facing radar going to see a Vision Jet, considering it's higher up and mostly composites?

Neither of those two things seem relevant at first glace, I could be wrong. Why would lidar be impacted by composites? Why would lidar be impacted by object height? Presumably you'd have sensors on the nose of the vehicle.

I was thinking the same way a year ago and telling my fanboy tesla friend how stupid it was removing lidars etc., but being open minded and always looking to be challenged, I started researching what they are doing and how they are doing it, watch their AI day video (its old in AI terms but gives a good insight), and I eventually got what they’ve meant when saying lidar introduces noise compared to what they are doing right now. TLDR is they convert photons into a multistereoscopic camera and reconstructing the brain’s behavior towards objects and motion…the way they transform the image data into high quality 3d point clouds is really mindblowing, people outside of the 3d ecosystem cannot fully appreciate that, and its kind of sad.

The camera performance is like the newer starlight sensors in terms of seeing in the dark, and they seem to process from raw data. Essentially, once they crack all the quirks, I am now a firm beleiver that it will better than the average joe with limited angulars vision+rotating head vs. surround stereoscopic vision.

Not saying I 100% agree with launching a beta program in the wild like this, but to their defense, given the current government driving paperwork and legislation to insane levels of unproductivity, innovators and challengers kinda have to play it out like that if they want to disrupt sectors (in the end) for the better.

Not saying the end justifies the means, but I can understand the logic.

> TLDR is they convert photons into a multistereoscopic camera and reconstructing the brain’s behavior towards objects and motion…

I think maybe someone dazzled you with some excessive jargon. Extracting point clouds from stereoscopic cameras isn't a Tesla invention, even though they've invested a lot into trying to make theirs work. This is some computer vision 101 stuff, but nobody calls it "converting photons".

> Essentially, once they crack all the quirks, I am now a firm beleiver

Well obviously if they solve all of the problems then we'd all be believers too.

It wasn't even new when Microsoft Kinect did it... let's see... 12 years ago.

Moogly from amiga?

Kinect point cloud was a total mess, never worked with more than a few cameras and had a LOT of restriction and artifacts, what I meant by impressive, imagine a moving 6 camera kinect with clean signals, that doesn't fail on specific texture or reflection because it track in 4D + AI.

Probably the people downrating are the same thinking the Tesla FSD AI is a kinect contraption.

> TLDR is they convert photons into a multistereoscopic camera and reconstructing the brain’s behavior towards objects and motion…

I just spent the time watching the vision segment of that video, and while the presenter uses the brain as a simile to motivate neural network architecture, it actually is not modeled at all on how primate visual cortex works. Seems some tweeps also got caught up in this hype. I think those few slides were really just to show the complexity of the problem.

For parties interested in actual models of biological vision, https://serre-lab.clps.brown.edu/wp-content/uploads/2014/10/... is a good resource.

Like I said, that was a while ago, If you watch all the follow-up interviews, including the recent TED talk with prior interview with the host, he does mention the process of thinking having solved it, exponential gains, then hitting a ceiling and having to change the approach.

The biggest problem I see is not if they will solve it (that's just a question of time), its how they communicate their optimism and "almost there" but always pushing further. That's more than annoying (at best) and contributing extensively to the negativity. I guess they are trying to get people excited, attract talent, etc.

Humans don't just use vision when driving a car. Our eyes also aren't fixed in our heads. They can gimbal independently of our heads. Our torsos also give our heads more degrees of freedom than just our necks. Human senses are actually fairly good.

The main issue with vision-only FSD is that visual artifacts don't convey high degrees of certainty about the solid nature of things.

If the AI doesn't recognize an object as being an object it can't add it to the list of things to worry about. If it's augmented with LIDAR and radar it doesn't need to recognize an airplane to know there's a solid object in its path.

It's dumb edge cases that cause problems. Vision-only FSD leaves a lot of edge cases on the table with the hope that AI magic will save the day.

Wow lots of moderation hatred when writing a simple tsla comment ;)

So if a human were driving the car, they would have crashed into the plane? Humans don't have radar or lidar.... and probably worse vision than the cameras on a Tesla have... but probably wouldn't have hit the plane.

There may well be valid arguments for radar/lidar superiority over cameras, but the line of reasoning you're suggesting isn't in that category.

humans are mobile, with moving necks/torso/eyes/etc.

I don't think anyone would argue that cameras couldn't do the same job if capable of realizing depth and re-assessing from different angles when confronted with ambiguity... but show me what a system like that would look like.

A human that is parking somewhere ambiguously often gets out of their car to inspect the area with their binocular vision from multiple angles; this is common among trailer owners.

The vision system on a vehicle is tasked with a harder problem in many ways -- it's trying to solve the ambiguity of the scene purely visually with mostly-stationary cameras and prior knowledge.

> moving necks/torso/eyes/etc

> realizing depth and re-assessing from different angles

A car with three or four widely spaced cameras can get way better visual information than a human can. (At least in the forward direction. Add more cameras for 360.) To the point that you could probably give the human a VR set tied into the cameras and have them do a better job because you can make the pillars transparent. (Might need a special VR set with extra-wide field of view.)

That's all fine and I don't necessarily take issue with what you're saying, but now we're talking about implementation details.

> but show me what a system like that would look like.

I'm guessing not so much more wacky as the lidar based cars I see driving through my neighborhood. The accident in the original link may well have been caused by an inadequate vision based system (I don't know), but just because an accident happened with a single vision implementation doesn't mean that somehow Lidar/Radar would necessarily be superior. That conclusion, tossed off by the original comment, is just wrong as it stands.

> A human that is parking somewhere ambiguously often gets out of their car to inspect the area with their binocular vision from multiple angles;

In most cases, but not always. I have an old friend that was completely blind in one eye: still legally licensed to drive and successfully navigated the environment better than a lidar/radar/artificial vision system still can. Even in bad weather conditions such as snow and rain (and night). Will that hold true over time? Hard to say, but I'd bet right now that the larger problem is processing the received information regardless of how the information was acquired.

The problem with the original statement is that it reminded me of the line of reasoning that goes, "there are aspects of the world that science has no answer for, therefore God." or "there was an abnormally cold day in spring, therefore no global warming" Tesla hasn't implemented a sufficiently robust visual driving system to avoid running into parked airplanes, therefore Lidar/Radar. It simply doesn't follow. Much more is needed.

Ultimately, to date, the only sufficient input for fully independent driving that we have is vision based. Lidar/Radar may ultimately be superior, but a vision based system is so far the only generally accepted method for driving... even if it is biological vision rather than computer vision.

I just had a look at the code of this feature. The existing Javascript function for this is just missing an if/else block for airplanes. I should have a fix ready shortly and I should be able push an update to all Teslas in the next few minutes. Nothing to worry, the update will keep airplanes safe.

P.S: Anyone know a npm package which has a "isAirplane" function which returns true for airplanes and false for others?

Does it have to return false for paper airplanes?

How about Helicopters?

Zepplins?

Further discussion here:

https://news.ycombinator.com/item?id=31123074

Friend of mine has a nice scrape across the side of his model 3 where it rubbed up against a pole in a parking garage during summon

the only thing that i want self-driving is my roomba

Until your Roomba crashes into the landing gear of your $3.5mm private jet.

it will turn around and find something else to crash on

A common cause of hull loss in aircraft.

Are you sure about that?

Pooptastrophe - https://www.usatoday.com/story/news/nation-now/2016/08/15/po...

Not OP but yes. Newer vacuums also have cameras to avoid this exact issue.

Having an autonomous and internet connected vehicle driving through my apartment makes me pretty uneasy.

Mine is running sone firmware (valetudo) that makes it talk to itself and think it's talking to an external API.

Even before that, it worked fine without ever being online.

This this is a game changer. It's like going from washing dishes to having a dishwasher.

Looks like they didn't teach the model about airplanes.

Perhaps should be a warning on the label. Something like:

"Training corpus did not include airplanes, people in wheel chairs,..."

"...any animals, ..."

That is going to be an infinitely long warning label

They can post it on all of your doors as a TOS for leaving the house. If you want to sue, you'll have to prove you climbed out of the window.

Models need to generalize over what they've been shown though. Most people won't learn to drive near an airplane or a wheelchair either, but would have no problem avoiding those.

Humans generate general 3d representation of the world. You don't know what this object is, yet you see it's borders. Tesla seems to generate understanding only of the objects that it has already seen.

Aggravating edge cases: you use the feature in an environment where not even 0.01% (optimistically) of drivers will drive

If your ML was half as generalized as you make it out to be it would not be a problem.

There are a lot of 0.01% cases when you have 600k+ tesla driving out there... and it's only the beginning

I can imagine a PM meeting at Tesla triaging the issues - "severe bug, yet very rare, waiving it for this release."

One can also imagine their modelers populating the synthetic training world with all of the planes right now. Like a "Mria" coming right in front of you on 101, as well as F-35 passing you on the left, and of course Millennium Falcon dropping right on you, and all that happening on a flight deck of huge aircraft carrier.

PSA. “mm” is not a unit of currency. It is a unit of length - a millimetre. If you want to represent millions of dollars use ‘M’, e.g., $3.5M.

I don’t know why people have started willfully using incorrect units, but it has to stop. FFS, just Google it.

In accounting/finance, m/M has a very long history of being 1,000 units, and mm/MM of being 1,000,000 units (it's not a unit itself, you still need the actual unit, it's effectively a representation of magnitude.)

It is starting to be replaced in many contexts (e.g., AP style for financial headlines) by the newer still-not-SI approach where one thousand uses k, one million uses M, and one billion uses B.

> I don’t know why people have started willfully using incorrect units

They aren't incorrect, just different than SI, which doesn't have units of money.

> FFS, just Google it.

It's funny, because if you had bothered to take your own advice you would know that this isn't something people recently started doing. Do your research before telling people to do theirs.

ETA to fix the airplane: 7 weeks.

ETA to fix the Tesla: 37 weeks.

7 weeks of MX time plus parts will easily run into 2-or-3-teslas price territory, assuming it is reparable at all. SF50 is one giant fiberglass shape. If you crack it seriously enough, there is no repair possible.

One expensive luxury toy crashes into another expensive luxury toy...

A vision jet is not necessarily a toy. Plenty of companies using them instead of larger aircraft.

Can't imagine why. The range on those things is pitiful. A TBM or a PC-12 seems like a much better choice if that's the class of aircraft you're buying.

Are you comparing a jet to propeller aircraft?

I'd imagine the jet flies higher, smoother and quieter, although I'm not rich enough to own one so I haven't looked into it too much.

Those are pressurized turboprops. They're turbine-powered and have the same performance as the SF50, including service ceiling, but with more range. The SF50 is what we call a "near jet"—it's technically a jet, but its performance is turboprop performance. It's not at all the same class of airplane as what most people think of when you say "jet." It's what you buy when you want to say you fly a jet but you aren't multi-rated and can't afford a Falcon.

I Was thinking the same thing. Backing mine out the garage now, headed over to Jimmy Johns, want anything?

I think your solution is flawed if it can't notice all objects in the box directly front of the vehicle. I would expect that to be one of minimum requirements, does the vehicle try to enter in "holes" smaller than itself.

Insurance call: "Sorry you said you crashed on what?"

Why would anyone use the summon feature and risk running over someone and killing them?

Autonomous vehicles present a shining example of why early adoption is risky.

Rich people problems

Until it crashes into some poor person's car.

Tesla owner’s insurance or the Tesla owner will pay for it, just as would happen in this case. No different than you leaving the parking brake off and the vehicle rolling down the hill into another car or someone’s house.

A plane would be well over most folks' car insurance limits.

We were talking about other cars, not $3.5M jets, in the subthread.

> Until it crashes into some poor person's car.

With that said, Tesla or not, you’re going to have a bad time if you hit a private jet with your car at the airport (unless you live in a state where you can just go bankrupt and protect your assets if they attempt to pursue you beyond your insurance limits).

It’s not like the whole $3.5MM jet is a write-off.

The hull damage and re-certification should be a fraction of the plane cost.

Composite fuselage - and the jet engine ($$$) itself sits between the ruddervators ($$$), with the BRS ($$$) just in front - probably the worst place to hit that jet. Easily $1-1.5M.

A wing or the nose would have been far easier to inspect and replace.

The legally required car insurance here in Spain covers damages up to 50 million EUR against a 3rd party's property.. Same in most European countries AFAIK

If they don't drive off before anyone notices.

What if the poor person was injured?

Liability coverage.

Insurance is great, but it doesn't automatically mean everyone lives happily ever after.

or some poor person.

Yeah but think of the wealth that could have trickled down to the middle class and eventually the lower class because of that jet ferrying the rich around to distribute their money as they see fit. Any conscious working class person would have thrown themselves in the way of that Tesla in the name of economic justice. The hospital bill would surely have been a crushing financial blow to him or her, and probably left them bankrupt, and also crippled, but what a boon to their class they would have preserved. Perhaps sympathetic workers could have donated to their cause, and to repair and clean the Tesla.

/s?

That's gonna leave a mark...

Viral indeed

new feature: auto pushback

Send a few Teslas to Ukraine. The need air defence -- could work this way :o)

I call shenanigans. I have a Tesla. This would be REALLY hard to pull off. you have to be physically close to the car (like within 60 feet I think), you have to hold down a button, and as soon as the car senses something near it (inches) it shuts down. this is either fake, or set up intentionally done for lulz.

A quick search turns up that Smart Summon has a range of 200 feet, not 60.

How is distance between the vehicle and the phone measure, and then provided to the vehicle?

Is it possible that the vehicle sensors that are responsible for detecting objects near the vehicle failed? If the sensor didn't fail, how about the control link that shuts the vehicle down?

So if anything is near the vehicle, it automatically shuts it down, or just during smart summon?

yup, good looking out, that's new since I last looked (3-ish years ago), and you are right, the sensors are designed for obstacles on a road, so you could probably miss something suspended 5 feet in the air.

It's possible the tail of the aircraft was just far enough out of the way for the camera to not see it, but still strike the edge of the MY.

But I agree, there is still something dubious about this. At the very end of the clip you three people basically stroll into frame. If you're either the owner of a 3.5 million aircraft or a $60k Tesla and they impact each other ... wouldn't you be moving a little more expediently? An "oh shit!" sort of reaction.

The Tesla is also going the wrong direction to get to these guys, theoretically.

And yea, you gotta hold down the summon button. The Tesla manual even states " remaining ready to release the button to stop the vehicle if needed.". If was summoning my Tesla and saw it headed remotely close to a Plane, I'd probably release the button. Especially after impact. Not keep holding it down so it keeps pushing the plan for another few seconds.

you are correct, not exactly difficult to stop a summons. yeah the camera by the rearview mirror is not angled up, so could probably miss an overhead obstacle.

> The Tesla manual even states " remaining ready to release the button to stop the vehicle if needed.".

Tesla also states "have your car come to you while you deal with a fussy child."

Apparently that should read "while you deal with a fussy child but maintain full attention on the car"?

You know how at the airport, they have that tug vehicle that pushes the aircraft out from the gate and into the taxiway?

I think the person was just trying to use their Tesla in the same way. The plane needed to be turned around, and the Telsa did it.

This is actually a feature, not a bug.

Instead of having to spool up the jet engine which would burn a lot of fuel and release a lot of CO2, the Tesla with its electric motor, can do it with zero CO2 emissions.

Reply on news.ycombinator.com