Tesla FSD beta crashes into object [video]
youtube.com
youtube.com
This should not be something that is allowed on public roads by end-users, but rather on closed tracks by specialists. If they want to test it out on public roads, run the analysis and look at whenever it diverges from the drivers decisionmaking instead.
Do they? If they don't, would you admit that it's probably a good thing to let it continue?
We just saw a FSD car run a red light and nearly hit pedestrians if the driver hadn't intervened.
I think where you're going here is toward an assertion that "any failure at all is unacceptable". And that seems innumerate to me. Cars fail. We're trying to replace a system here that is already fatally (literally!) flawed. The bar is very low.
Failure is not the issue when it comes to Tesla FSD, accountability is.
For any mistakes human drivers makes, they have to pay up with money, have their license suspended, or with jail time, depending on the severity of their mistake.
You fuck up, you pay the price. That's the contract under which human drivers are allowed on the road. Humans drivers are indeed flawed, but with our law and justice systems, we have accountability to keep those who break the law in check, while allowing freedom for those who respect it. It's one of the pillars of any civilized society.
In my country, running a stop sign or a red light means you get your license suspended for a while. When a self driving Tesla does the same mistake, why doesn't Tesla's FSD AI have its "license" suspended as well? That's the issue.
Are there accidents where death is unavoidable? Yes, they happen every single day, but after the investigations and trials are over, the parties found responsible pay up for those deaths in either money or jail-time, or both.
Does that mean we should we allow machines to make deadly mistakes, especially when death IS avoidable? Absolutely not. We sentence humans for such mistakes. Machines (either their operator or their manufacturer) should also have the same liability.
Those are two different things which you're trying to spin into a strawman.
/me roundhouse kicks out of the thread
Let's say you are on an overpass above a train, and a very fat man is in front of you. The train, if it isn't stopped, will kill 10 people on the tracks. But, if you push the person in front of the train, it will kill 11 people, and one of those would be you committing homocide.
What do you do?
[edit] Transit is 10x safer than private cars. [1]
> The effects extend beyond individual trip choices, too: the report notes that transit-oriented communities are five times safer than auto-oriented communities. Better public transportation contributes to more compact development, which in turn reduces auto-miles traveled and produces safer speeds in those areas. On a national scale, too, the U.S. could make large advances in safety if each American committed to replacing as few as three car trips per month with transit trips.
[1] https://mobilitylab.org/2016/09/08/transit-10-times-safer-dr...
By all means, advocate for more transit friendly urban centers. I'm with you. Just don't take away autonomy out of spite. Better cars are still better, even if they're not the solution you want.
I'd actually disagree with this stance. Making passenger cars safer through autonomy is probably a good thing if we can actually make it safer than human drivers. I've yet to be convinced we are anywhere close to meeting the bar on that if. I assume we will eventually, but I'm not even sure I'll live to see it.
It also ignores potential knock on effects, sure in isolation safer cars are better, but the reality is nothing exists in isolation. Could we save more lives if instead of spending the money we are on self-driving cars we instead invested it into our transit systems?
As an example of knock on effects, affordable cars feels like an easy win right? Makes travel easier for everyone. But by and large affordable cars are what has allowed suburbs to exist, but there's an argument to be made that urban sprawl is far from ideal and that we'd be better of with denser communities and public transit.
What would convince you? Data from 60k cars isn't sufficient?
It would be if the data showed they were safer than human drivers, and was independently obtained. I have yet to see any data that suggests this or anything close to this.
You seem to have taken the opposite conclusion, which is exactly what the feeding frenzy over the paper wanted.
[1] The methodology is hugely suspect: you can't take an incomplete data set and then just "correct" it by inventing axes that you pull in from other incomplete data sets that weren't studied or measured in the original! That's rank P-Hacking. It seems reasonable, but I guarantee that a talented statistician can push any such data set 2x in either direction with that kind of trick.
It's not a question of taking away autonomy, and there's certainly no spite about it. If you want to get around town, you have bikes or e-bikes. If you want to get out of urban centers, you can always rent a car at the periphery.
I've lived in SF for 10 years with no car and have never felt unable to do anything I've wanted at any time.
> If you want to pick blackberries at the local farm, you need a passenger car.
Not really, the farm can have a bus with regularly scheduled pick-ups or routes like a lot of the Napa wineries do.
> If they want to test it out on public roads, run the analysis and look at whenever it diverges from the drivers decisionmaking instead.
It would be trivial to analyze the data after the fact to see where the AI model diverges from the human driver’s actions, decide which one was right, then implement the correct action. That wouldn’t slow down testing at all, as the only difference is who’s controlling the vehicle. In any beta test, someone still has to analyze the data.
Making intelligence out of silicon isn't easy. Let the computers learn this way during the transition period, and finally we can remove human drivers from the road.
More than 38,000 people die every year in crashes on U.S. roadways. And Tesla makes the safest cars:
The Insurance Institute for Highway Safety (IIHS), the main independent organization that conducts crash tests on vehicles in the US, released the result of its latest tests on the Tesla Model Y and confirmed that it achieved the highest possible safety rating.
>The Insurance Institute for Highway Safety (IIHS), the main independent organization that conducts crash tests on vehicles in the US, released the result of its latest tests on the Tesla Model Y and confirmed that it achieved the highest possible safety rating.
That doesn't mean that Telsa makes the safest cars. There are roughly ~100 cars with that rating, and nothing suggests the Model Y is safer than any of the others. It's also important to note that that rating isn't based of real world data such as how often drivers actually crash and hurt other (ex, how often FSD fails), but rather how well the occupant is protected in the event of a crash.
When does it plan to start doing that? Safety features in cars have come for decades without trying to kill people first. Just ask Volvo.
The general non-Tesla-owning public, including pedestrians and cyclists, have not given their consent to be part of Elon's public beta test.
I bet the risk of getting injured by Tesla beta version of FSD is miniscule compared to a risk of getting into accident caused by a human driver. I am not for banning either of them. Even when we get to the point where FSD is much safer than drivers I would be against of banning humans.
You can bet all you want, but human drivers, as flawed as they may be, are all fully liable by law for any mistakes they make at the wheel and have to pay with money or jail time plus losing their license.
Who is liable for the mistakes FSD makes? Who goes to jail if it runs down a pedestrian by mistake? Elon? The driver? Can the FSD loose its license like human drivers can for their mistakes?
You can't compare a human drivers to FSD safety when FSD has zero liability in front of the law and all the blame automatically goes to the driver.
Yup. The driver. Aside from image, there appears to be few if any incentives for FSD to improve beyond the “it does the right thing 80% of the time” mark.
I think it is simply not true. Creating a FSD that could replace human drivers eg in case of trucks is potentially highly lucrative and it would make the economy more efficient.
But that doesn’t mean that tesla will save lives. Maybe they do a bunch of this and learn that cameras alone aren’t sufficient, and then waymo wins. Tesla wouldn’t have saved any lives, only killed a couple people unnecessarily.
Medicine is the classic example of applying this kind of thinking. You could do all sorts of unethical medical testing to speed up medical research, saving countless lives down the line, but we don’t because it doesn’t make it right.
With another medical example, you could roll out snake oil without testing it thoroughly because it’ll save lives if it works. But maybe snake oil doesn’t work, and it’ll be some other thing that works, and by rushing the snake oil, you just made things worse.
Bringing up medicine for your stance might back fire. There are tactical examples where policies to "go slow and reduce risk" have cost lives. For example, testing some medicine on pregnant women is bad for the fetus - so there are policies to "not test on anyone who might be pregnant", and as a result, there are few studies on women between the ages of 20 - 80, and women's health has suffered as a result
Also, regarding the medicine, do you really believe we do not do "unethical" medical testings? I guess it depends on your ethical standards and how high they are :)
But let's get back to cost benefit trade off. COVID vaccines tests were rushed. So it is obviously sometimes worth it.
There is a risk in not taking a risk.
The FSD AI shouldn't be connected to the real-world controls until it very rarely substantially deviates from what the human drivers do in their cars while controlling a virtual car from the real-world sensor inputs. And in those cases where it deviates, the devs should manually dig in and run to ground if it would have hit something/someone/broken the law. Not until that list of cases stops growing, particularly in your dense urban areas full of edge cases, do you even start considering linking up the real car.
From what I'm seeing they're instead turning Tesla drivers into training supervisors with the real-world serving as the virtual one, while putting everyone else on/near roads at risk.
It's criminal, and I expect consequences. It's already illegal to let a child steer from your lap, and that's without even handing over the pedals. People operating "FSD Beta" on public roads should be at least as liable, where are the authorities and enforcement?
Continuing this line of logic, the most advanced FSD would save the most lives. Thus the argument could be made that Tesla should abandon their research and license Waymo FSD.
Also, why do we allow car manufacturers to test their own software? Shouldn't it be done by a third party? And shouldn't that third party be the only ones allowed to push updates to the car?
Sometimes, conspicuous lack of progress is a good thing. It isn't something you necessarily appreciate until you suddenly start having to confront the law of large numbers in a very real and tangible way. Some incremental changes simply are not feasible to take until they are complete. Level 3 automation is one of those...
LIDAR greatly reduces the "black box" necessity. It basically allows you to do things like "if object is in the way then hit brake/move elsewhere", where the sensor doesn't really fail in good weather.
Given its safety over DL-only solutions, this should be step 1 to getting to FSD. Not reckless beta-testing with black box techniques.
Tesla has chosen the cheap way, which is also the irresponsible way.
And in the meantime, I and other drivers, cyclists, pedestrians are subject to increased danger for what? Oh, Tesla's profits? Forgive us if we don't all see this as an acceptable tradeoff.
Furthermore if you really were so edge-of-your-seat scared of traffic fatalities then Tesla would be at the bottom of your list. Why don’t you go do something about the droves of people that stream out the back of bars and into their cars every night? They kill thousands every year meanwhile Tesla has killed roughly zero people.
I am very, very well aware of exactly what causes traffic fatalities. According to the software at my work, I have personally responded to 378 fatality MVAs as a paramedic. Please don't try to assume everyone is ignorant about realities - we are not blindered, and only physically capable of recognizing and responding to one danger at a time.
Having said that I am not sure most people are safer for cyclists and pedestrians. FSD is in such a bad state right now that the tesla is driving in the streets at the speed 80y old people do. What I saw in a video is a car that drive at a similar pace to a cyclist, it is even much slower in the crossing sections.
So to be clear you can have an always available hot button for Netflix, but not for climate control. All Tesla's handwaving is entirely bullshit. "The driver is in the seat for legal purposes only. The car is driving itself."
This is horseshit of the highest order.
Imagine Volvo, but instead of Volvo you have a company that distinguished themselves by their lack of interest in safety.
Hard to understand how people have affection for this brand.
I agree they could start there, but you’ll graduate quickly without having learned much.
The main question is, are we willing to put people in harms way today for the benefit of future humans? The answer seems pretty obvious to me. Drunk humans are considerably worse than this and are not going away anytime soon. If we can solve self driving just 1 year earlier it’s equivalent to saving 30,000 American lives.
Put another way, if you want rules that delay the advancement of self driving driving cars, you're effectively murdering 30,000+ American lives every year.
That strawman argument works if you completely ban all human drivers the moment we solve self-driving.
The big question is, when exactly do we consider self driving solved that it can ban replace drivers? All current evidence points it's very, very far away, if ever.
That's pretty much the nail in the FSD coffin.
Evolution has gifted us with a phenomenal device, the brain, but evolution is extremely cautious and conservative. There is very little reason to expect that human innovation won't catch up and surpass evolution. 10,000 years from now, I expect self-driving will work perfect. 10,000 years from now, our brains will (unless modified by our technology) largely be unchanged.
A massive problem, particularly with the more capable L4 style technologies is you get lulled into a false sense of security because it'll drive a decent amount of distance perfectly fine right up until it doesn't (and it's normally spectacularly bad when it fails).
Testing purely on closed track or from simulated analysis only goes so far, you definitely need to do public road testing before you can hand it to end-users. But the driver needs to do advance driver training so that they're more aware of the hazards, they need to always be ready to take over (as with an L3 system, even if you're developing L4), they need to fully understand the capabilities, ODD and behaviour of the software, you need to keep the durations/distances short to avoid driver fatigue and ideally you have a second person in the vehicle to keep them honest.
Letting end-consumers use this technology on public roads is insane. It feels to me like the reason Tesla do it is they've boxed themselves into a corner because they've already sold it.
The common pro Tesla argument of delaying self driving costing lives is absolutely valid. But it doesn't follow that allowing Tesla FSD to run red lights is the best way to accelerate the adoption as a whole.
If the whole industry was as cautious as Waymo, I think the risk of regulation would be minimal.
This is orthogonal to how well done the system is. I am extremely impressed it works as well as it does. But it's not good enough.
It would be if self driving Teslas were safer, but the evidence seems to suggest that they crash more often than regular drivers once you control for when it is used.
If Tesla’s approach ultimately succeeds in getting FSD working and into the mainstream, then for each year earlier that happens than it otherwise would, up to about 30,000 lives are saved, millions of injuries prevented, and roughly $1 trillion USD in total lifetime value is preserved as the worldwide accident rate goes to zero.
In that context, FSD should absolutely be civilization’s next moonshot.
That’s to say nothing of the other economic benefits besides simply killed and maiming fewer people and destroying less property.
What cost would be attributed to Covid, roughly?
That’s just the USA
The other, darker, possibility is that Tesla does so poorly in this moonshot that they poison the well for everyone else, some of whom might be more equipped to succeed.
If through their actions FSD is ultimately delayed a year, then they contributed to extended the carnage of human driven automobiles.
The point is the scale of the problem, and how valuable it is to solve it.
If Tesla's approach ultimately turns into a dead end (and that is the outcome of the overwhelming majority of very ambitious tech projects) then all the time and resources spent on Tesla's approach could be seen in hindsight to have been diverted away from potentially fruitful projects and are actually making mainstream FSD happen later than they otherwise would.
FSD being civilization's next moonshot doesn't mean Tesla should necessarily get a free pass to test beta FSD tech out on public roads without scrutiny.
Because I'm saying just because the goal is extremely worthwhile, it's not carte blanche to move fast and break things.
But lives killed by inaction should be counted!
- 2.20: Taking a left turn with a trajectory into the opposite lane. The youtuber's camera clearly was adjusting exposure from underneath the overpass and I wonder if the car's cameras had the same issues.
- 2.40: Right on red without stopping. I wonder if this is related to the stop sign rolling stop issues
- 3.30: Collision with pylon. It seems like the perception system didn't pick up the pylon at all (or if it did, its not rendering on the screen)
- 5.10: Some weird funky trajectory and then pulling into the sunken curb/sidewalk on the right turn
- 6.28: Trying to drive down the railroad tracks
I gave up after the above.
Despite Tesla's (Elon) insistence that FSD will be safer than human drivers soon enough (1 year), there is very little evidence to support that. At the very least if we are going to allow beta testing on roads then Tesla should be forced to submit their FSD incident data to the CA DMV like the rest of car manufacturers that are testing AVs.
maybe while parked…
I'm also in awe of how fickle its predicted path is, especially when turning. Why commit to a new path for a tenth of a second and then change its mind again? This suggests to me it's placing way too high weight on poor quality telemetry. Heck, at 5:14 it can't even decide which street it wants to go down. Maybe it's just a UI thing (always displays the most likely path) and the underlying model is actually keeping both options open.
At 6:45 it tries some insane passing maneuver -- I can't get over that one. Passing on the right, in a bus-only lane, planning to squeeze through the rapidly-closing gap of the car in front of it and a parked car, where there is currently a bicyclist. And it looks like the bicyclist is the only thing keeping it from doing that insane maneuver -- each time the bicyclist is occluded by the car in front (object permanence, hello?), it attempts to pass.
I've lost count of the number of things this car doesn't understand, that it should to be on a road:
* object permanence
* reading basic text in simple print
* logical connections between the map and road geometry
* patience and commitment
* understanding physics and geometry of other moving and non-moving vehicles
* seeing stationary bright orange objects immediately in front of it
Wonder if LIDAR would have spotted the pylon? It sounds like neat tech.
I would totally drive on those train tracks myself.. it's pretty common here, and there was no signage. Curious to see what the AI would do next if the driver hadn't intervened.
I would not have had the ability to take that license with me to the new car I purchased. I'm surprised this practice hasn't been investigated by any state justice or commerce dept (I'm not sure which it should be in the US?) as many consumers were sold something that was never provided.
They always have this "buy now, the price will go up later!" but I had to ask myself: how much can the price go up? It's not going to be a $20,000 feature any time soon. I figured I was risking a couple thousand dollars that I would need to pay to upgrade later.
Then all of these old Teslas with FSD for life will become rare, similar to the Teslas witch unlimited free supercharging.
I'm not sure there's a way to regulate consumer stupidity. People are conned into losing money everyday, not all of those cons are crimes.
However, consumers putting down $10,000 for a promised feature with no firm deadline and severe restrictions on the said feature's portability are in a league of their own when it comes to stupidity.
Or, perhaps they are more akin to an Investor ..
[1] https://www.findmyelectric.com/blog/does-full-self-driving-f...
AI does not step forward progressively. Normally, engineers spend ages improving things, testing, improving again. With AI, it's far more of a "let's just re-train this black-box on a totally different set of data, wind up with a totally different black-box, and see if its better!" The difference in driving quality and intelligence between even minor version updates is dramatic. I very intentionally _do not_ update my Tesla until the release is ~6 months old. This... sucks.
I am still very bullish on Tesla and EVs, and on technology in general - unlike many I am not full of hate that this didn't work out, nor do I particularly think FSD is _dramatically_ more dangerous behind the wheel than say my grand father was before we convinced him to stop driving, but I do think I have more or less zero patience for AI engineers these days. The failure by Google and others to deliver even a moderately usable voice assistant is all the proof I need that FSD is _years_ away.
I think the current state of FSD is extremely impressive. There are clearly some extremely talented engineers working on that system. But my guess would be that it simply can't be incrementally improved until it's road-safe, in much the same way that GPT-3 is insanely impressive but can't be incrementally improved until it can write a whole human-quality novel. I suspect a new wave of radical rethinking of machine learning might be necessary, and that might be decades away.
Not only Musk. I would love to see a brain imaging study of humans as their net worth gains orders of magnitude.
I have seen and felt the effects myself to a very limited extent.
I have heard Elon and others say in regards to the internet: "We are not evolved to handle this much information, and this much choice."
My intuition leads me to believe that the same could be said for having an absurdly large net worth. We just aren't evolved for it and one's behavior changes in weird ways.
disclaimer: While I am pointing fingers, it is at our entire species, not "billionaires."
I think Tesla may be using the price of FSD as a way to dissuade people from buying it, while enabling it to be a funding mechanism for continued work on autonomous driving.
As evident in this video, FSD is often acting like a nervous student driver by being unable to confidently make decisions on the path to take, running through red lights, driving unpredictably at slow speeds, and more. This is probably due to the fact that there is ONLY cameras being used to inform FSD, which can suffer from contrast issues.
I've seen videos in the past where Tesla camera input is in black/white/grayscale for processing. It seems like if you converted the video to B&W, the pylon and the road are similar darkness/color, so I'm not surprised this had issues. Tesla Autopilot was suffering from issues with contrast all the way back in 2016 when Autopilot failed to detect a semi crossing a highway in Florida which lead to the death of the Tesla driver. This was due to the white trailer against a bright sky.
Ultimately, as a consumer and former Tesla owner, I don't feel confident in Tesla's ability to get autonomous driving working well at a human-capable level, let alone the "10x safer" bar they've set for themselves.
This is just one of the many reasons why FSD is not only putting drivers at risk, but it is really turning its customers into real-life crash dummies.
A more accurate re-brand of this product should be something more like 'Fools Self Driving' than 'Full Self Driving', given the false advertising and lack of safety features for these drivers using the product.
4:42 it makes a right onto tram track. And again at 6:24
Jesus.
I'm not doubting you, but I find that surprising. Admittedly the dark green blends into the background a bit, but the retroreflectors catching the sunlight stood out for me even in the video, which would have muted the effect somewhat.
> I imagine I’d have a hard time driving those roads in bad weather or at night and maybe even under the lighting during the video.
In bad weather or at night you should have you headlights on and those retroreflectors will make them shine brightly. They'll actually probably be easier to spot. In the event the weather is bad enough the retroreflectors can't bounce your headlights back at you, it's going to be bad enough your not seeing the paint either.
This car can't even see orange pylons and lane markings. That's not even passable for a student driver on a closed course.
But WHY is a (I'm assuming) concrete post that short and painted DARK GREEN? Green seems like a horrible choice for a post in night driving. That's not every particularly well designed for human driving. San Jose? Why isn't this post yellow? Yellow seems to be the typical concrete post color.
Surface/local street FSD driving is 10-20 years away and will require a lot of convergence, construction standards, regulatory testing, common databases of "danger locations", etc. It's so much harder. So many "target types" so much variation, so many different local standards.
In particular I don't agree with Tesla's lack of a database or route-specific training or the like. Lidar I think is actually less of an issue than that approach.
But a good divided highway automation should be doable. That has a lot more standardization, and higher amounts of traffic would lead to better statistical models and datasets for proving out algorithms.
These are the problems in highway driving: traffic jams (can be prewarned)/ stopped cars (a current tesla issue), deer/animals wandering onto the road, construction (can be signaled in a standard way and prewarned). Debris/potholes?
Those pylons have reflective strips on them for the night.
Not just the customers, anybody else unfortunate enough to share the road with them.
As a frequent pedestrian/bicyclist in my city, Teslas scare the shit out of me. Even worse, when Tesla drivers cause accidents, the response is typically complete shock and disbelief that FSD failed them. And an utter lack of responsibility for the damage and/or injury they caused. Because obviously they weren't driving, it was the car!
Both Elon Musk and the driver are equally at fault and need both go to jail, they enable each other to make the public roadways unsafe.
Seriously injecting religion by posting irrelevant scripture in a discussion of legal liability is purely proselytizing and an attempt to convert people.
We can choose how to look at the shared responsibility. No reason to restrict discussion to current legal doctrine.
https://en.wikipedia.org/wiki/Judgement_of_Solomon
I wouldn't be surprised if a same or similar story is present in other cultures.
There's another place where it tries to turn onto the light rail trackway and then ends up in the bus lane, but that area of First Street in San Jose is confusing.
Tesla used to make a lot of PR about how they were training their AI collecting data from vehicles in the field, although nobody seems to have tapped in and found out what they're collecting. You'd think that if they did that, they'd log all the places where there was trouble or the human overrode the system. Then transmit cautions to cars getting close to trouble spots, so the fleet as a whole never made the same mistake twice. Apparently not.
[1] https://nyc.streetsblog.org/2021/12/21/incoming-dot-commissi...
That's kind of like what you're describing, but more dumb where the car tries to avoid the trouble spot entirely.
Intel is pushing them through an IPO and I’m pretty sure I’m going to put down a substantial investment out of all the players they seem the be the only ones that actually taking a serious evolutionary approach to consumer AVs.
Their upcoming product stack shows an impressive strategy they aren’t trying to sell L4/5 as something that is just around the corner.
They are still perfecting Level 2/3+ as a consumer targeted product (~$1000 OEM price) and focusing on full autonomy only for robotaxis where an OEM price of $10,000-20,0000 for the hardware won’t be an issue.
Fully
And Stupid
Their attitude to testing would be laughable if it wasn't so dangerous and reckless.
It’s called full self driving: even rudimentary crash avoidance from 10+ years could avoid this
The streets have not only tram/train lanes but tram specific signals, keep clear areas,etc. Not to mention a mixture of bike and park lanes.
Never driven a Tesla personally but I'd be pretty confident FSD could handle the 50 mile commute to my friends place that is just interstate, highways, and a roundabout.
My only real complaint in the video is how aggressive it seems but I think that's configurable. Perhaps they should limit the FSD speed/aggressiveness to the specific drivers own capability in the area.
Lucky it was a plastic bollard and not someone's leg or baby carriage.
Most probably the type of accidents the FSD MVP will be having is different than a human would and vice versa.
There are 60k testers [0] of FSD and only 1 accident since September. OK, not really average drivers but that's one accident for 15k years of (some) driving. Probably FSD is already safer than most drivers. Although FSD is not yet complete as it cannot handle some situations at all and hands over to the human.
[0] https://seekingalpha.com/article/4482970-teslas-ace-in-the-h...
Which goal is more difficult technically, FSD or Mars Settlement?
As in, which one just needs money and time versus true technical invention?
How often do these version number increment?
Was he using a version that was tested for more than a couple of months?
If not, how can that be justified?