This is why I believe the approach of incremental improvement towards full self driving is fundamentally flawed. These advanced driver assist tools are good enough to lull users into a false sense of security. No amount of "but our terms and conditions say you need to always pay attention!" will overcome human nature building that trust and dependence.
Your reasoning doesn’t apply to the incremental improvements to self-driving, rather Tesla’s decision to allow all cars to use TACC/auto-steer. They haven’t even given people “the button” to enroll in FSD beta, likely because they know it would be extremely bad PR when a bunch of people use it without paying attention.
Today, there is at least one of the most advanced Neural Networks entering each car: A human being. If we could just implement the AI to add to this person and not replace it...
This is the bit nobody likes to realize. FSD at it's best...is still about as fallible as a human driver. Minus free will (it is hoped).
I will be amused, intrigued, and possibly a bit horrified if by the time FSD hits level 5, and they stick with the Neural Net of Neural Nets architecture if there isn't a rash of system induced variance in behavior as emergent phenomena take shape.
Imagined news:
All Tesla's on I-95 engaged in creating patterns whereby all non-Teala traffic was bordered by a Tesla on each side. Almost like a game of Go, says expert. Researchers stumped.
Then again, that's imply you had an NN capable of retraining itself on the fly to some limited degree, which I assume no one sane would put into service... Hopefully this comment doesn't suffer a date of not aging well.
What would such an AI even look like? If it spots every real danger but also hallucinates even a few dangers that aren’t really there, it gets ignored or switched off for needlessly slowing the traveler down (false positives, apparently an issue with early Google examples [0]); if it only spots real dangers but misses most of them, it is not helping (false negatives, even worse if a human is blindly assuming the machine knows best and what happened with e.g. Uber [1]); if it’s about the same as humans overall but makes different types of mistake, people rely on it right up until it crashes then go apoplectic because it didn’t see something any human would consider obvious (e.g. Tesla, which gets slightly safer when the AI is active, but people keep showing the AI getting confused about things that they consider obvious [2]).
I disagree. One feature my car has is to pull me back into the lane when I veer out of it (Subaru's lane keep assist). That is still incremental improvement towards "full self driving". I agree, however, that Tesla's Autopilot is not functional enough, and any tool designed to allow humans to remove their hands from the wheel should not require their immediate attention in any way.
I think people just assume Tesla's Autopilot is more capable than it really is.
My car has adaptive cruise control and lane keep assist, but I'm not relying on either for anything more complex than sipping a drink while on the highway.
Yep, if anything they’re just a way to make long drives or stop and go highway traffic more tolerable. When I got my first car with those features it seemed like a gimmick, but they really help to reduce fatigue.
I think that's actually a step towards a local maximum that makes us less likely to achieve actual FSD. The safer we can make AI-guided driving where a person is still in control, the higher the bar becomes for a solo AI to be significantly safer than the alternatives.
In a plane you also have far more time to react once the autopilot disconnects for whatever reason, than the fraction of a second that a car gives you.
This is a misrepresentation of the dumpster fire that was the WMATA train situation. Yes, the fatal crash was the last straw, but the root problem was not the automation system but rather the complete lack of maintenance that led to its inability to work properly. Congress refusing to fund maintenance and then falling behind 10-15 years on it lead to all kinds of systems failing. The fatal fire in the blue line tunnel under the river occurred with a human at the controls, but we’re similarly not blaming that incident on the perils of human operation.
I don’t blame the operator for the crash. The other train was behind a blind curve and she hit the emergency brake within a reasonable amount of time given what she could see. However the speeds of the system were set too high for the operator to safely stop because they assumed the ATC would work perfectly.
The difference is, they have traffic controllers and the train have their own dedicated rails, almost no obstructions and a train into train crash danger situation rarely arises.
The planes have a lot of maneuvering space to all sides.
Car traffic and streets are more dense and often have humans crossing them without regards to laws, bicycles, motorbikes, road construction and bad weather.
Not saying one auto pilot system is better than the other, however, they operate in different environments.
I have a Subaru Forester base model with lane keeping and adaptive cruise control.
I need to be touching the wheel and applying some force to it or it begins yelling at me and eventually brings me slowly to a stop.
I’ve had it for a year now and I cannot perceive of a way, without physically altering the system (like hanging a weight from the wheel maybe?) that would allow me to stop being an active participant.
I think the opposite is true: Tesla’s move fast and kill people approach is the mistake. Incremental mastering of autonomous capabilities is the way to go.
> physically altering the system (like hanging a weight from the wheel maybe?)
was exactly what people were doing. But it's also possible to be physically present, applying force, but being "zoned out", even without malicious intent.
> I need to be touching the wheel and applying some force to it or it begins yelling at me and eventually brings me slowly to a stop.
> I’ve had it for a year now and I cannot perceive of a way, without physically altering the system (like hanging a weight from the wheel maybe?) that would allow me to stop being an active participant.
That's exactly what people were doing with the Tesla. Hanging a weight to ensure the safety system doesn't kick in. [0][1]
If people are consciously modifying their car to defeat obvious safety systems, I have a really hard time seeing how the auto manufacturer should be responsible.
I guess the probe will reveal what share of fatal accidents are caused by this.
Well, it doesn't help when the CEO of the company publically states that the system is good enough to drive on its own and those safety systems are only there because of regulatory requirements.
GM's Supercruise (which is the actual king of the hill for L2 systems) uses cameras to track the driver's eye position to ensure they are paying attention. It's significantly harder to defeat, is geofenced to prevent use in incompatible situations like surface streets, and has a much more graceful disengagement process. Most of the time autopilot is smooth, but sometimes it just hands control back to the driver without warning.
Teslas can famously be tricked by wedging an orange between the rim and spoke of the steering wheel to produce enough torque on the wheel to satisfy the detection. There are enough videos of it on youtibe that tesla could easily be found negligent for not doing enough to prevent drivers from defeating a safety system given that alternate technology that more directly tracks attention is available and tricking tesla's detection method became common knowledge.
Personally I’ve found this to be sufficient in my Forester. Even holding the wheel but not being “there” isn’t enough. The car is really picky about it.
I own a Model Y and am a pretty heavy Autopilot user. You have to regularly give input on the steering wheel and if you fail a few times it won't let you re-engage until you park and start again.
Personally Autopilot has actually made driving safer for me... I think there's likely abuse of the system though that Tesla could work harder to prevent.
I personally think the issue boils down to their use of the term "Autopilot" for a product that is not Autpilot (and never will be with the sensor array they're using IMO.)
They are sending multiple signals that this car can drive itself (going so far as charging people money explicitly for the "self-driving" feature) when it cannot in the slightest do much more than stay straight on an empty highway.
They should be forced to change the name of the self-driving features, I personally think "Backseat Driver" would be more appropriate.
Autopilot is precisely the correct term -
An autopilot is a system used to control the path of an aircraft, marine craft or spacecraft without requiring constant manual control by a human operator. Autopilots do not replace human operators.
> the issue boils down to their use of the term "Autopilot" for a product that is not Autpilot
It is literally an autopilot. Just like an autopilot on an airplane, it keeps you stable and in a certain flight corridor. There's virtually no difference except for Tesla's Autopilot's need to deal with curved trajectories.
Well, they still need to avoid the collisions more reliably, apparently. Once they do it perfectly reliably I will add it into the list of the things it does differently from an airplane autopilot. ;)
You're literally describing how the Tesla system works. It requires you to keep your hand on the wheel and apply a slight pressure every so often. The cabin camera watches the driver and if they're looking down or at their phone, it does that much more often.
People causing these problems almost certainly are putting something over the cabin camera and a defeat device on the steering wheel.
It depends. As long as the resulting package (flawed self driving system + the average driver) isn't significantly more dangerous than the average unassisted human driver, I don't consider it irresponsible to deploy it.
"The average driver" includes everyone, ranging from drivers using it as intended with close supervision, drivers who become inattentive because nothing is happening, and drivers who think it's a reasonable idea to climb into the back seat with a water bottle duct taped to the steering wheel to bypass the sensor.
OTOH, the average driver for the unassisted scenario also includes the driver who thinks they're able to drive a car while texting.
> As long as the resulting package (flawed self driving system + the average driver) isn't significantly more dangerous than the average unassisted human driver...
Shouldn't that compared to "average driver + myriad of modern little safety features" instead of "average unassisted driver"? The one who has the means to drive a Tesla with the "full driving" mode certain has the means to buy, say, a Toyota full of assistance/safety features (lane change assist, unwanted lane change warning and whatnots).
Does that even matter? If the state doesn’t care to enforce its laws against reckless driving, why should the manufacturer be encumbered with that responsibility?
It almost certainly is, at least when combined with the intentional inattention that follows.
Making it a crime isn't an "obvious solution" to actually make it not happen. Drunk driving is a crime and yet people keep doing it. Same with texting and driving.
The problem is determining who is liable for damages, not prevention. Shifting the liability for willfully disabling a safety control puts them on notice.
Prevention as a goal is how we end up with dystopia.
The average driver breaks multiple laws on every trip. Most of the time no one gets hurt. But calibrating performance against folks violating traffic and criminal laws sets the bar too low for an automated system. We should be aiming for standards that either match European safety levels or the safety of modes of air travel or rail travel.
Yes, I agree. We should hold automated systems to a higher standard. Unless you’re proposing we ban automated systems until they’re effectively perfect because that would perversely result in a worse outcome: being stuck with unassisted driving forever.
Except that doesn't work if you're trying to produce a safe product. Investigations into crashes in the airline industry have proven that removing pilots from active participation in the control loop of the airplane results in distraction and an increased response time when an abnormal situation occurs. Learning how to deal with this is part of pilots' training, plus they have a co-pilot to keep an eye on things and back them up.
An imperfect self driving vehicle is the worst of all worlds: they lull the driver into the perception that the vehicle is safe while not being able to handle abnormal situations. The fact that there are multiple crashes on the record where Telsas have driven into stationary trucks and obstacles on roads is pretty damning proof that drivers can't always react in the time required when an imperfect self driving system is in use. They're not intrinsically safe.
At the very least drivers should be required additional training to operate these systems. Like pilots, drivers need to be taught how to recognize when things go awry and react to possible failures. Anything less is not rooted in safety culture, and it's good to see there are at least a few people starting to shine the light on how these systems are being implemented from a safety perspective.
> Perfect is the enemy of good, and rejecting a better system because it isn't perfect seems like an absurd choice.
Nothing absurd about thinking a system which has parity with the average human driver is too risky to buy unless you consider yourself to be below average at driving. (As it is, most people consider themselves to be better than average drivers, and some of them are even right!) The accident statistics that comprise the "average human accident rate" are also disproportionately caused by humans you'd try to discourage from driving in those circumstances...
Another very obvious problem is that an automated system which kills at the same rate per mile as an average human drivers will tend to be driven a lot more because no effort (and probably replace better-than-average commercial drivers long before teenagers and occasional-but-disproportionately-deadly drivers can afford it).
> drivers using it as intended with close supervision
Doesn't this hide a paradox? Using a self-driving car as intended implies that the driver relinquishes a part of the human decision making process to the car. While close supervision implies that the driver can always take control back from the car, and therefore carries full personal responsibility of what happens.
The caveat here is that the car might make decisions in a rapidly changing, complex context which the driver might disagree with, but has no time to correct for through manual intervention. e.g. hitting a cyclist because the autonomous system made an erroneous assertion.
Here's another way of looking at this: if you're in a self-driving car, are you a passenger or a driver? Do you intend to drive the car yourself or let the car transport you to your destination?
In the unassisted scenario, it's clear that both intentions are one and the same. If you want to get to your location, you can't but drive the car yourself. Therefore you can't but assume full personal responsibility for your driving. Can the same be said about a vehicle that's specifically designed and marketed as "self-driving" and "autonomous"?
As a driver, you don't just relinquish part of the decision making process to the car, what essentially happens is that you put your trust in how the machine learning processes that steer the car were taught to perceive the world by their manufacturer. So, if both car and occupant disagree and the ensuing result is an accident, who's at fault? The car? The occupant? The manufacturer? Or the person seeking damages because their dog ended up wounded?
The issue here isn't that self-driving cars are inherently more dangerous then their "dumb" counter parts. It's that driving a self-driving car creates it's own separate class of liabilities and questions regarding responsible driving when accidents do happen.
I actually disagree. (And before you respond, please read my post because it’s not a trivial point.)
The fact that an huge formal investigation happened with just a single casualty is proof that it may actually be superior for safety in the long-term (when combined with feedback from regulators and government investigators). One death in conventional vehicles is irrelevant. But because of the high profile of Tesla’s technology, it garners a bunch of attention from the public and therefore regulators. This is PRECISELY the dynamic that led to the ridiculously safe airline record. The safer it is, the more that rare deaths will be investigated and the causes sussed out and fixed by industry and regulators together.
Perhaps industry/Tesla/whoever hates the regulators and investigations. But I think they are precisely what will cause self driving to become ever safer, and eventually become as safe as industry/Tesla claims, safer than human drivers while also being cheap and ubiquitous. Just like airline travel today. A remarkable combination of safety and affordability.
This might be the only way to ever do it. I don’t think the airline industry could’ve ever gotten to current levels of safety by testing everything on closed airfields and over empty land for hundreds of millions of flight hours before they had sufficient statistics to be equal to today.
It can’t happen without regulators and enforcement, either.
I largely agree with you, but I just wish regulators would start by only allowing these assist programs for people that are already known to be poor drivers. The elderly and convicted drunk drivers, for example. That way we could have the best of both worlds.
Requiring people to buy a special car/system to be able to drive doesn't seem like an incentive - it seems similar to the interlock system we currently require drunk drivers to purchase to be able to drive.
If anything a driver monitoring system seems even better than the interlock system, for example you couldn't have your kids/friends blow for you to bypass it.
I disagree. I would not put people who shoved poor judgment in situation, where they can further hurt other or themselves it. People like that are more likely not to pay attention and do other irresponsible things.
>
In the 1st quarter, we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.05 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 978 thousand miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 484,000 miles.
I think the comparison should be Tesla with/without AI, not Tesla/not-Tesla; so roughly either x2 or x4 depending on what the other active safety features do.
It’s not nothing, but it’s much less than the current sales pitch — and the current sales pitch is itself the problem here, for many legislators.
> we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged [...] for those driving without Autopilot but with our active safety features, we registered one accident for every 2.05 million miles driven
This still isn't the correct comparison. Major selection bias with comparing miles with autopilot engaged to miles without it engaged, since autopilot cannot be engaged in all situations.
A better test would be to compare accidents in Tesla vehicles with the autopilot feature enabled (engaged or not) to accidents in Tesla vehicles with the autopilot feature disabled.
As was stated elsewhere, most accidents happen in city driving where autopilot cannot be activated so the with/without AI is meaningless. We need to figure out when the AI could have been activated but wasn't, if you do that then you are correct.
On the contrary to your overall point: The fatal crash rate per miles driven is almost 2 times higher in rural areas than urban areas. Urban areas may have more accidents, but the speeds are likely lower (fender benders).
Tesla's vehicles are almost 10X safer than the average vehicle. Whether their autopilot system is contributing positively or negatively to that safety record is unclear.
The real test of this would be: of all Tesla vehicles, are the ones with autopilot enabled statistically safer or less safe than the ones without autopilot enabled?
Tesla's safety report lacks data and is extremely misleading.
1. Autopilot only works on (or intended to work on) highways. But they are comparing their highway record to all accident records including city driving, where accident rate is far higher than highway driving.
2. They're also comparing with every vehicle in the United States including millions of older vehicles. Modern vehicles are built for higher safety and have a ton of active safety features (emergency braking, collision prevention etc). Older vehicles are much more prone to accidents and that skews the numbers.
The reality is Teslas are no safer than any other vehicles in its class ($40k+). Their safety report is purely marketing spin.
They also include miles driven by previous versions of their software in the “safe miles driven” tally. There’s no guarantee any improvement would not have resulted in more accidents. They should reset the counter on every release.
> The reality is Teslas are no safer than any other vehicles in its class ($40k+).
Would another way of saying this be that they are as safe as other vehicles in that class? And that therefore Autopilot is not more unsafe than driving those other cars?
I would probably agree, but I also think it’s a case of “need more data”.
We should really compare Autopilot with its competitors like GM’s Super Cruise or Ford’s Blue Cruise, both of which offer more capabilities than Autopilot. That will show if Tesla’s driver assist system is more or less safe than their competitors product.
What capabilities does GM or Ford have that Tesla doesn't? Neither GM nor Ford have rolled out automatic lane changing. Teslas have been doing that since 2019.
The reason GM's Super Cruise got a higher rating by Consumer Reports was because CR didn't even test the capabilities that only Tesla had (such as automatic lane change and taking offramps/onramps). Also, the majority of the evaluation criteria weren't about capabilities. eg: "unresponsive driver", "clear when safe to use", and "keeping the driver engaged".[1]
Do you know many vehicles $40K+ that don't have BLIS and rear/front cross traffic alerts? While a radar-based blind sport alert (one that warns if a car behind is moving too fast to safely merge) is probably irrelevant for the city driving, the cross traffic is extremely useful when pulling out of driveway obstructed by parked cars, I personally have seen several accidents just on my street that could have been prevented with cross traffic detection. I think the expensive models (S/X) still have the front radar so they may have the front alert but I don't think any model ever had the rear radar for the rear cross traffic alert.
Asking someone to pay attention when they are not doing anything is unrealistic.
I would be constantly bored / distracted. My wife would instantly fall asleep. Etc etc.
This argument is flawed, because when regulators investigate a Tesla crash, Waymo doesn't care the slightest. The technologies (emphasis on having skeuomorphic cameras vs a lidar), approaches (emphasis on generating as many situations as possible in simulated worlds and carefully transitioning to the business case vs testing as early in the real world with background data captation) and results are so different between the actors in this specific industry that one's flaws being fixed or improved won't necessarily translate into others benefitting from it.
Conversely, when Waymo iterates and improves their own safety ratios by a significant amount, that evidently does not result in Tesla's improving in return.
Then why not flip the scheme. Instead of have the human as backup to the machine, make the machine backup the human. Let the human do all the driving and have the robot jump in whenever the human makes a mistake. Telemetry can then record all the situations where the human and the machine disagreed. That should provide all the necessary data, with the benefit of the robot perhaps preventing many accidents.
Of course this is impossible in the real world. Nobody is going to buy a car that will randomly make its own decisions, that will pull the wheel from your hands ever time it thinks you are making an illegal lane change. Want safety? How about a Tesla that is electronically incapable of speeding. Good luck selling that one.
> * Let the human do all the driving and have the robot jump in whenever the human makes a mistake.*
Because when the human disagrees with the machine, the machine is usually the one making a mistake. It might prevent accidents, but it would also cause them, and you lose predictability in the process (you have to model the human and the machine).
Nobody is going to buy a car that will randomly make its own decisions, that will pull the wheel from your hands ever time it thinks you are making an illegal lane change.
That's almost exactly what my Honda does. Illegal (no signal) lane change results in a steering wheel shaker (and optional audio alert). And the car, when sensing an abrupt swerve which is interpreted as the vehicle leaving the roadway, attempts to correct that via steering and brake inputs.
But, I agree with your more general point - the human still needs to be primary. My Honda doesn't allow me to remove my hands from the steering wheel for more than a second or two. Tesla should be doing the same, as no current "autopilot" system is truly automatic.
Interesting, I assumed it didn't, given the prevalence of stories about driver watching movies on their phones. I guess they just leave one hand lightly on the wheel, but are still able to be ~100% disengaged from driving the car.
> Illegal (no signal) lane change results in a steering wheel shaker (and optional audio alert).
To be clear, tying the warning to the signal isn't about preventing unsignaled lane changes, it's gauging driver intent (i.e. is he asleep and drifting or just trying to change lanes). It's just gravy that it will train bad drivers to use their signals properly.
Is a lane change without signal always illegal? I know that it almost certainly make you liable for any resulting accident, but I'm not sure that it is universally illegal.
I have no idea, but the point wasn't so much that the lane change is illegal, but that lack of signal is used to indicate lack of driver attention. I shouldn't have used "illegal" in my original post.
> 142 (1) The driver or operator of a vehicle upon a highway before turning (...) from one lane for traffic to another lane for traffic (...) shall first see that the movement can be made in safety, and if the operation of any other vehicle may be affected by the movement shall give a signal plainly visible to the driver or operator of the other vehicle of the intention to make the movement. R.S.O. 1990, c. H.8, s. 142 (1).
That said there's zero cost to doing so regardless of whether other drivers are affected.
It could be one of those "if a tree falls in the forest" scenarios. If a cop is near enough to see you not signal, he could easily argue that he himself might have been affected by the turn or lane change.
Yes, failure to signal is a traffic violation. At least everywhere I've lived/traveled in the US. It's also a rather convenient excuse for police to "randomly" pull you over (I've been pulled over by Chicago PD for not signaling for a lane change, despite actually having done so).
> That's almost exactly what my Honda does. Illegal (no signal) lane change results in a steering wheel shaker (and optional audio alert). And the car, when sensing an abrupt swerve which is interpreted as the vehicle leaving the roadway, attempts to correct that via steering and brake inputs.
By the way, this is fucking terrifying when you first encounter it in a rental car on a dark road with poor lane markings while just trying to get to your hotel after a five hour flight.
I didn't encounter an obvious wheel shaker, but this psychotic car was just yanking the wheel in different directions as I was trying to merge onto a highway.
Must be what a malfunctioning MCAS felt like in a 737 MAX, but thankfully without the hundreds of pounds of hydraulic force.
Just to add, I have a 2021 Honda, and disabling this functionality is a 1-button-press toggle on the dash to the left of the steering wheel. Not mandatory.
> Want safety? How about a Tesla that is electronically incapable of speeding.
That would be unsafe in many situations. If the flow of traffic is substantially above the speed limit--which it often is--being unable to match it increases the risk of accident. This is known as the Solomon curve [1].
> Subsequent research suggests significant biases in the Solomon study, which may cast doubt on its findings
With the logic presented in the theoretical foundation section, it seems that the safer move would actually be slow down and match the speed of all the trucks and other large vehicles... which won't happen.
Matching speed sounds great, except there are always people willing to go faster and faster. In my state they raised the speed limit from 70 to 75, it just means more people are going 85-90. How is that safer?
To address your last paragraph, everyone going 85-90 is less safe than everyone going 70-75, you are correct.
However, you individually going 70-75 when everyone else is going 85-90 is less safe than you going 85-90 like everyone else in the exact same situation.
>there are always people willing to go faster and faster
That’s why no one says “go as fast as the fastest vehicle you see”, it is “go with the general speed of traffic”. That’s an exercise for human judgement to figure that one out, which is why imo it isn’t a smart idea to have the car automatically lock you out of overriding the speed limit.
> However, you individually going 70-75 when everyone else is going 85-90 is less safe than you going 85-90 like everyone else in the exact same situation.
And yet the roads are full of vehicles literally incapable of going 85. Many trucks cannot do more than 69mph.
People are going faster because they felt it's safer, not because of the speed limit. You can design roads that cause humans to slow down and be more careful.
A self driving car obviously needs to be aware of other cars on the road. I don't see any reason why the car couldn't observe other cars, see what speed they are going at, and refuse to go faster than the rest. A car that refuses to do 120mph when all the other cars are doing 60mph in a 50mph zone should be trivial.
(Trivial if the self driving tech works at all....)
You're getting downvoted for this comment apparently, but I'm still of the firm belief that we will never see full autonomous driving without some sort of P2P network among cars/infrastructure.
There's just too much shit that can't be "seen" with a camera/sensor in conjested traffic. Having a swarm of vehicles all gathering/sharing data is one of the only true ways forward IMO.
On most or all roads below 100km/h autopilot won’t allow speeding, and therefore I drive at the limit, which I know I would not have done if I controlled it. It also stays in the lane better than I do, keeps distance better, and more. Sometimes it’s wonky when the street lines are unclear. It’s not perfect but a better driver than I am in 80% of cases.
My insurance company gives a lower rate if you buy the full autopilot option, and that to me indicates they agree it drives better than I, or other humans, do.
On most or all roads below 100km/h autopilot won’t allow speeding, and therefore I drive at the limit, which I know I would not have done if I controlled it
If following the speed limit makes cars safer, another way to achieve that without autopilot is to just have all cars limit their speed to the speed limit.
Sometimes it’s wonky when the street lines are unclear. It’s not perfect but a better driver than I am in 80% of cases
The problem is in those 20% of cases where you'd lulled into boredom by autopilot as you concentrate on designing your next project in your head, then suddenly autopilot says "I lost track of where the road is, here you do it!" and you have to quickly gain context and figure out what the right thing to do is.
Some autopilot systems use eye tracking to make sure that the driver is at least looking at the road, but that doesn't guarantee that he's paying attention. But at least that's harder to defeat than Tesla's "nudge the steering wheel once in a while" method.
> just have all cars limit their speed to the speed limit.
The devil is in the details... GPS may not provide sufficient resolution. Construction zones. School zones with variable hours. Tunnels. Adverse road conditions. Changes to the underlying roads. Different classes of vehicles. Etc.
By the time you account for all the mapping and/or perception, you could've just improved the autonomous driving and eliminated the biggest source of humans driving: The human.
But you can still impose a max speed limit based on available data to cover most normal driving conditions but it's still on the driver to drive slower if appropriate. And that could be implemented today, not a decade from now when autonomous driving is trustable.
The parent post said that autopilot won't let him go over the speed limit and implies that makes him safer. My point is that you don't need full autopilot for that.
So this is not a technical problem at all, but a political one. As the past year has shown, people won't put up with any convenience or restriction, even if it could save lives (not even if it could save thousands of lives)
The single system you're describing, with all of its complexity, is a subset of what is required for autonomous vehicles. We will continue to have road construction, tunnels, and weather long past the last human driver. Improving the system here simply improves the system here -- you cannot forsake this work by saying "oh the autonomous system will solve it" -- this is part of the autonomous system.
Just add all those to the map system. It could be made incredibly accurate, if construction companies are able to actually submit their work zones and "geofence" them off on the map.
GPS is extremely accurate honestly. My garmen adjusts itself the very instant I cross over a speed limit sign to a new speed, somehow. Maybe they have good metadata, but its all public anyway under some department of transportation domain and probably not hard to mine with the price of compute these days. Even just setting a top speed in residential areas of like 35mph would be good and save a lot of lives that are lost when pedestrians meet cars traveling at 50mph. A freeway presents a good opportunity to add sensors to the limited on and off ramps for the car to detect that its on a freeway. Many freeways already have some sort of sensor based system for charging fees.
What would be even easier than all of that, though, is just installing speeding cameras and mailing tickets.
During the 3 years I’ve owned it there are 2 places where lines are wonky and I know to take over.
I have not yet struggled to stay alert when it drives me, and it has driven better than I would have - so it certainly is an improvement over me driving 100% of the time. It does not have road rage and it does not enjoy the feeling of speeding, like I do when I drive, nor does it feel like driving is a competition, like I must admit I do when I am hungry, stressed, or tired.
> just have all cars limit their speed to the speed limit
No way I’d buy a car that does not accelerate when I hit the pedal. Would you buy a machine that is not your servant?
I keep saying the same thing actually whenever people say that manual driving will be outlawed. Like, no, it won't be - because the computers will still save you in most cases either way, autopilot enabled or not.
>>How about a Tesla that is electronically incapable of speeding. Good luck selling that one.
From 2022 all cars sold in the EU have to have an electronic limiter that keeps you to the posted speed limit(by cutting power if you are already going faster) - the regulation does allow the system to be temporarily disabled however.
Your summary is incorrect. The ETSC recommends that Intelligent Speed Assistance should be able to be overridden.[1] It's supposed to not accelerate as much if you're exceeding the speed limit, and if you override by pressing the accelerator harder, it should show some warning messages and make an annoying sound. It's stupid, but it doesn't actually limit the speed of your car.
I think it's a silly law and I'm very glad I don't live in a place that requires such annoyances, but it's not as bad as you're claiming.
I hired a new car with Intelligent Speed Assistance this summer, though it was set (and I left it) just to "ping" rather than do any limiting. I drove it to a fairly unusual place, though still in Europe and with standard European signs. It did not have a GPS map of the area.
It could reliably recognize the speed limit signs (red circle), but it never recognized the similar grey-slash end-of-limit signs. It also didn't recognize the start-of-town or end-of-town signs, so it didn't do anything about the limits they implied.
I would certainly have had to disable it, had it been reducing the acceleration in the way that document describes.
> Of course this is impossible in the real world. Nobody is going to buy a car that will randomly make its own decisions, that will pull the wheel from your hands ever time it thinks you are making an illegal lane change.
Yeah, add to that the unreliability of Tesla's system means that it cannot pull the wheel from the driver, because it's not unusual for it to want to do something dangerous and need to be stopped. You don't want it to "fix" a mistake by driving someone into the median divider.
> How about a Tesla that is electronically incapable of speeding. Good luck selling that one.
Instead they did the exact opposite with the plaid mode model S, lol. It kind of works against their claims that they prioritize safety when their hottest new car - fully intended for public roads - has as its main selling point the ability to accelerate from 60-120 mph faster than any other car.
What? Cars with collision detection systems already exist, and they can handle both side on and head on collision avoidance when a human is driving.
People literally are buying cars that “make their own decisions”. Importantly though, these systems only activate in the case of an imminent collision IF the corrective measure won’t cause another collision.
> that will pull the wheel …
Yah, of course no one is going to buy a care that does what you describe because what you describe is insane and inherently unsafe. Unless a collision is imminent, nothing happens.
This line of thinking is flawed because it assumes a smooth surface over the safety space, where if you make incremental improvements you will head towards some maxima of safety. e.g. : the wing fell off; investigate; find that you can't use brittle aluminum; tell aircraft manf. to use a more ductile alloy. Self driving technology isn't like that -- you can't just file a bug "don't mistake a human for a plastic bag", fix that bug and move on to the next one. No number of incremental fixes will make self driving that works as any reasonable human would expect it to work.
To tell you the truth, I generally do and haven't used it for ages. Where I drive, the roads have some amount of traffic. I find (traditional) cruise control encourages driving at a constant speed to a degree that I wouldn't as a driver with a foot on the gas. So I don't "hate" regular cruise control but I basically never use it.
Maybe a fairly small sample size but I don't know the last time I've been in a car where the driver has turned on cruise control. But it probably varies by area of the country. In the Northeast, there's just enough traffic in general that it's not worth it for me.
In traffic is where traffic aware cruise control is most useful. A lot of people I knew who bought Tesla's in the bay area specifically bought it so their commutes would be less stressful with the bumper to bumper traffic. I drove 3000+ miles across the country last year with > 90% of it on AP and I was way less tired with AP on vs off and it allowed me to just stay focused on the road and look out for any issues.
Yes. I was (explicitly) talking about traditional "dumb" cruise control. I haven't used adaptive cruise control but I agree it sounds more useful than traditional cruise control once you get above minimal traffic.
One thing worth noting about Subaru's approach to this that is specifically relevant to bumper-to-bumper traffic, is that it will stop by itself, but it won't start moving by itself - the driver needs to tap the accelerator for that. It will warn you when the car in front starts moving, though.
That is the fatal flaw in anything but a perfect system - any kind of system that is taking the decisions about steering from the driver is going to result in the driver at best thinking about other things and worse getting into the back seat to change. If you had to develop a system to make sure someone was paying attention, you wouldn't make them sit in a warm comfy seat looking at a screen - you would make them actively engage with what they were looking at - like steering.
And ultimately it doesn't matter how many hundreds of thousands of hours of driving you teach your system with, it may eventually be able to learn about parked cars, kerbs and road signs, but there won't be enough examples of different accidents and how emergency vehicles behave to ever make it behave safely. Humans can cope with driving emergencies fairly well (not perfectly admittedly) no matter how many they've been involved in using logic and higher level reasoning.
We have an education problem. People have no idea what computers do because they're illiterate (literacy would mean knowing at least one language well enough to read and write in it) so they just take other people's word that they can do some magical thing with software updates. The most extreme examples of this were the iPhone hoaxes telling people that software updates provided waterproofing or microwave charging.
I mean there are videos of a vehicle's occupant sitting in the rear seats making food and drinks while the vehicles are tricked into operating off of the vehicles sensors.
It is not solely the trust and dependence but inclusive is the group of idiots with access to wealth without regard to human life.
I expect that to design self-driving you need to push the limits (with some accidents) a bit with a bunch of telemetry. Going from not-much to full-self-driving requires a lot of design increments.
All Level 2 systems need to better integrate with the driver. Upon engagement driver and driver assist are team where communication and predictability is crucial.
I think the fundamental flaw is indisputable. Everyone is aware of in-between stage problems. I don't think it's an insurmountable flaw.
These things are on the road already. They have issues, but so do human only cars. Tweaks probably get made, like some special handling of emergency vehicle scenarios. But, it's not enough to stop it.
Meanwhile, it's not a permanent state. Self driving technology is advancing, becoming more common on roads. Procedures, as well as infrastructure, is growing around the existence of self driven cars. Human supervisor or not, the way these things use the road affects the design of roads. If your emergency speed sign isn't being headed by self driven cars, your emergency speed sign has a bug.
Full self driving is one of those things where getting 80% of the way there will take 20% of the effort and getting the remaining 20% of the way there will take 80% of the effort.
Tesla auto-drive seems like it's about 80% of the way there.
Is it? Tesla is still alive because they're selling cars.
It's just that the companies that are NOT doing incremental approaches are largely at the mercy of some investors who don't know a thing about self-driving, and they may die at any time.
I agree with you that it is technically flawed, but it may still be viable in the end. At least their existence is not dependent on the mercy of some fools who don't get it, they just sell cars to stay alive.
That's one of the major problems of today's version of capitalism -- it encourages technically flawed ways to achieve scientific advancement.
Would be interesting to know how many buy based on the fsd hype(including the ones who don't pay for the package) and how many buy because of the "green" factor.
However many there are who buy because of the fsd promise, all that revenue is coming from vaporware (beta ware at best) and is possible due to lack of regulatory enforcement.
History shows that the longer the self regulatory entities take the p, the harder the regulatory hammer comes down eventually.
One problem that is often ignored in these debates is that people already don't always pay attention while driving. Spend some time looking at other drivers next time you are a passenger in slow traffic. The number of drivers on their phones, eating, doing makeup, shaving, or even reading a book is scary.
It therefore isn't a clean swap of a human paying attention to a human who isn't. It becomes a complicated equation that we can't just dismiss with "people won't pay attention". It is possible that a 90%/10% split of drivers paying attention to not paying attention is more dangerous when they are all driving manually than a 70%/30% split if those drivers are all using self-driving tech to cover for them. Wouldn't you feel safer if the driver behind you who is answering texts was using this incremental self-driving tech rather than driving manually?
No one has enough data on the performance of these systems or how the population of drivers use them to say definitively that they are either safer or more dangerous on the whole. But it is definitely something that needs to be investigated and researched.
I remember reading Donald Norman's books decades ago, and one of the prime examples of the dangers of automation in cars was adaptive cruise control- which would then suddenly accelerate forward in a now-clear off-ramp, surprising the heck out of the previously-complacent driver, and leading to accidents.
We've known for a very long time that this sort of automation/manual control handoff failure is a very big deal, and yet there seems to be an almost willful blindness from the manufacturers to address it in a meaningful way.
A car that beeps when I drift out of lane, or beeps when I go too fast before a curve, or beeps like hell if I cross over the center median would be hugely useful, because a record of every warning would be there, whether correct or not.
Conversely, if it didn't warn me right before an accident, then the absence of that warning would be useful too.
All of that information should be put back into the model based on crash reporting. Everything else can be ignored.
I would argue that the information should be available to all automakers (perhaps using the NHTSA as a conduit), so that each of them have the same safety information, but can still develop their own models. The FAA actually does this already with the FAA Accident and Incident Data Systems [0] and it has worked pretty darn well.
The new Toyota RAV4's have this feature-- if you go out of bounds in your lane they beep and the steer5ing wheel gives a bit of resistance.
It also reads the speed limit signs and places a reminder in the display. I think it can brake if it detects something in front of it, but I'm not certain.
My main point (perhaps buried more than it should have been) is that by centralizing accident data along with whether an alert went off (or not), and sharing that with all automobile manufacturers can help this process proceed better.
Right now the data is highly fragmented and there is not really a common objective metric by which to make decisions to improve models.
> "The involved subject vehicles were all confirmed to have been engaged in either Autopilot or Traffic Aware Cruise Control during the approach to the crashes,"
No mention of the deceptive marketing name "Full Self Driving" in the article.
Exactly. Some of these cars do not even have 'Driver Monitoring', which means the car doesn't even track if the driver has their eyes on the road at all times, which puts many other drivers at risk.
On top of that, FSD is still admittedly Level 2; Not exactly 'Full Self Driving'? And the controls can easily be tricked to think that the driver has their 'hands on the wheel' which is not enough to determine driver attentiveness while FSD is switched on.
I checked the website and they seem to be contextualizing "Full-self driving" with it coming at a future date:
> All new Tesla cars have the hardware needed in the future for full self-driving in almost all circumstances. [...] As these self-driving capabilities are introduced, your car will be continuously upgraded through over-the-air software updates.
I also personally would prefer they stuck to 'autopilot' and avoided the word full in 'full self-driving' and otherwise be more specific about what it means.
Other car companies typically productize the various features like lane assist, following cruise control, etc rather than bundle it into one. But that definitely makes communicating it more difficult.
Tesla probably doesn't want to call it 'limited self-driving' or 'partial self-driving'. Maybe 'computer assisted driving' but that doesn't sound as appealing. I can see the difficulty marketing here. But again not using 'full' as in it's complete and ready-to-go would help.
This is just more evidence of the confusion that Tesla marketing has created. "Full Self-Driving Capability" is the literally quoted option they've been selling for years now.
I don’t know much about trading, but it appears to be down nearly 5% this morning as of right now. Regardless, I think you’re conflating trading price with whether something is material in general.
Unfortunately the article doesn't mention anything about how common it is for human drivers to crash into first responder vehicles during the night. I'm not trying to downplay these cases, as hitting emergency vehicles is very bad indeed, yet ~4 such crashes per year might be in the same ballpark or even better than "unassisted" drivers that cause such crashes.
Also doesn’t mention that this is an inherent limitation in TACC+ systems, and is specifically called out as such in Volvo, BMW, and Cadillac vehicle manuals as a limitation. Much ado about nothing unless regulators are going to outlaw radar based adaptive cruise control (which, of course, they’re not).
frankly, it's hard to even crash into an emerfency in my opinion while actually driving snd paying attention given tgeir lights have gotten so darn bright it's damn near blinding. frankly, I have to slow to a crawl not out of rubbernecking fascination, butout of self preservation to adapt to the dang lighting searing my retinas at night.
now running into unlit emergency vehicles? still think tgat's rather difficult sans inebriation or sleep dep.
the lights are actually what cause the crash. Some drivers just drive straight into the lights. This is part of why police, at least around here, have particular protocols around how far they stop behind a car, and never standing between the cop car and the car they stopped.
Considering that less than 0.3% of cars in the US are Teslas and that - I would guess - less than 10% of them are using autopilot at any given time - they are likely 100s of times more likely to hit first responders.
In Canada the Highway Act states that you must move over (change lanes) for stopped emergency vehicles. It seems to solve that problem gracefully, leaving an empty lane between the stopped vehicles and traffic.
Long overdue. We're going to need a more rapid and iterative way to do this if we got to have even a chance of autopilot type technologies succeeding over the long run. Companies and regulators need to be collecting feedback and pushing for improvement on a regular basis. I still don't think it's likely to succeed, but if it did, this would be the way.
Predictable outcome of a sensing system fully based on deep learning. Rare unusual situations don't have enough training data and lead to unpredictable output.
I still think that Tesla's approach is the right one, I just think they need to gather more data before letting this product be used in the wild unsupervised.
Current TACC/auto-steer doesn’t use deep learning except on the newest Model 3/Y vehicles with “TeslaVision”. All cars with radar use the radar and radar only to determine if they should stop for the following car.
How does TeslaVision work with stationary objects at night? Like say a big ass truck with its lights off? Do you just pray the vision system recognizes “something” is there? I know they want to pursue a pure-vision system with no radar input, but it seems like there will be some crazy low light / low visibility edge cases you’d have to deal with.
How does a human detect a big ass truck with its lights off at night? This is solvable with computer vision. Tesla's dataset is almost nothing but edge cases, and they keep adding more all the time. My money says they'll get there.
The thing is we don’t. Many people die from rear ending broke down trucks. I’m fine with that, I’m not so sure if regulators will be fine with a TSLA killing someone and going “welp, a human wouldn’t have seen it either, let’s just add this to our edge case dataset”.
I think it is necessary for the crashes to occur, to gather the data required to re-train the auto-pilot. We as a society need to decide whether we want to pay this cost of technological advancement.
No. Gathering data of human drivers braking in those circumstances would result in a perfectly fine dataset. This idea of needing human sacrifice is bonkers.
Oopsie. This strikes me as perhaps one of the growing pains of not-quite-self-driving: common sense would dictate that manual control would be taken by the driver when approaching an unusual situation like a roadside incident, but we just can't trust the common sense of a minority of people.
Tesla perhaps isn't being loud enough about how autopilot isn't self-driving, and shouldn't even be relied upon to hit the brakes when something is in front of you.
What on earth makes you think it's the minority of people who stop paying attention when machinery is handling some task all by itself the vast majority of the time? There's plenty of research that says otherwise even among highly-trained individuals.
I remember when Google was first presenting on driverless technology about 10 years ago, and they mentioned how you have to go right to full self driving, because any advanced driver assistance will clash with human risk compensation behavior.
Risk compensation is fascinating; driving with a bike helmet causes the biker and drivers around the biker to behave more dangerously.
Is society sophisticated enough to deal with advanced driver assistance? Is it possible to gather enough data to create self driving ML systems?
Risk compensation probably also works the other way, looking forward to all news cars standard supplied with a new safety device that cuts traffic accidents to a small fraction of what they used to be, the only ones remaining are all fatal for the driver.
A nice and very sharp 8" stainless steel spike on the steering wheel facing the driver.
I cannot open the study in the first link but the second on seems to actually refute the claim instead of supporting it.
> There is a body of research on how driver behaviour might change in response to bicyclists’
appearance. In 2007, Walker published a study suggesting motorists drove closer on average when
passing a bicyclist if the rider wore a helmet, potentially increasing the risk of a collision. Olivier
and Walter re-analysed the same data in 2013 and claimed helmet wearing was not associated with
close vehicle passing.
> We then present a new analysis of the original dataset, measuring directly the extent to which drivers changed their behaviour in response to helmet wearing. This analysis confirms that drivers did, overall, get closer when the rider wore a helmet.
Glad the regulators are looking into this. It bothers me that now Tesla seems to have no liability at all for the system not working, since it's always the driver's fault for not paying enough attention.
As Teslas get better at driving the drivers will be paying less attention inevitably, Tesla needs to start being responsible at some point
Every year young drivers die because they were inexperienced and didn't realize they were going too fast to too slow for a certain situation.
Once full self driving is statistically safer than humans how will you not let people use it? It is like saying you would rather have 10 children die because of bad driving skills rather than 1 child die because they were not paying attention at all times.
>Once full self driving is statistically safer than humans how will you not let people use it?
I'm fine with self-driving if/when it works (though I'm pretty sure from watching FSD Beta videos shot and edited by their biggest fans with a few interventions every 5 minutes, this is many many many years away for Tesla). But the company selling the self driving has to be responsible to some degree for the mistakes it makes.
As far as that goes if we want to save lives we can just regulate that semi-autonomous cars have to enforce the speed limit + 0 visibility from fog and heavy rain is an automatic pull over and wait for conditions to improve.
I don't think cars enforcing the law is a good idea at all. You have the moral and legally defensible right to break the law where not doing so would create substantial and immediate risk to yourself of others.
Just statistically safer won’t cut it - it will have to me many orders of magnitude safer. Instead of drunk people and mobile phone users dying it will be random accidents that humans would easily have avoided but is some weird edge case for the ML model. It’ll be a cars plowing down kids on trikes on a clear day, all captured in perfect HD on the cars cameras and in the press the next day with the crying driver blaming the car.
That’ll be a hard thing to overcome for the public. The drunk person “had it coming”, but did little Timmy?
The crash in the city of Woodlands, Texas, was pretty terrifying. After hitting a tree, the car caught on fire. The driver was found in the back seat, presumably because he couldn't figure out how to open the door to get out.
I just read about that after your post. Even with the typical lawyer hyperbole it's pretty bad.
It seems to me that Tesla door handles (in a world where they've been designing door latches for some time) are just plain ridiculous and likely unreliable but are a side effect of the market the company has been selling into. Gadgets go a long way with Tesla owners.
Obviously, things like a latch should not only work under all conditions including no-power, but they should probably be the same under all conditions. 'Emergency' latches aren't going to be used during an emergency as muscle memory is too important.
Interesting! When HN discussed that story a few months ago[1], the common notion was that the driver had enabled autopilot and climbed into the back seat.
That discussion was because the local sheriff's office said that there was nobody in the drivers seat at the time of impact. No idea why they would have said that. Tesla says the steering wheel was deformed in a manner consistent with a person being in the driver's seat when it crashed.
I drive a lot across Europe: as in, really a lot, long trip across several countries, several times a year. I drive enough on the highways to know a few scary situations, like the truck driver in a big curve slightly deviating out of his lane and "pushing" me dangerously close to the median strip for example.
To me driving requires paying constant attention to the road and being always ready to act swiftly: I just don't understand how you can have a "self driving car but you must but be ready to put your hands back on the steering wheel and your foot on the pedal(s)".
I have nothing against many "recent" safety features, like the steering wheel shaking a bit if the car detects you're getting out of your lane without having activated your blinker. Or the car beginning to brake if it detects an obstacle. Or the car giving you a warning if there's a risk when you change lane, etc.
But how can you react promptly if you're not ready? I just don't get this.
Unless it's a fully self-driving car, without even a steering wheel, a car should help you focus more, not less.
> But how can you react promptly if you're not ready? I just don't get this.
You cannot, that's the simple truth. You're supposed to focus on the road anyways and should be able to take over once any sort of autopilot or assist system starts working erroneously, yet in practice many people simply assume that those systems being there in the first place mean that you can simply stop focusing on the road altogether.
It feels like the claim of "fully self driving vehicle" is at odds with actual safety, or at least will remain so until the technology actually progresses far enough to be on average safer than human drivers, moral issues aside. Whether that will take 15, 50 or 500 years, i cannot say, however.
That said, currently such functionality could be good enough for the driver to take a sip from a drink, or fiddle around with a message on their phone, or even mess around on the navigation system or the radio - things that would get done regardless because people are irresponsible, but making which a little bit safer is feasible.
It's nothing (well certainly not everything) to do with people's assumptions. There's a ton of research around how people simply stop paying attention when there's no reason for them to pay attention 99% of the time. It doesn't even need to be about them pulling out a book or watching a movie. It can simply be zoning out.
Maybe, as you say, it's feasible today or soon to better handle brief distractions but once you allow that it's probably dangerous to assume that people won't stretch out those distractions.
We have empirical data showing how safe actual level 2 self driving cars are in practice. So there’s no reason to work from base assumptions. Yes, level 2 self driving cars cause avoidable accidents, but overall rate is very close to the rate people do. The only way that’s happing is they are causing and preventing roughly similar numbers of accidents.
Which means people are either paying enough attention or these self driving systems are quite good. My suspicion is it’s a mix of both, where people tend to zone out in less hazardous driving conditions and start paying attention when things start looking dangerous. Unfortunately, that’s going to cause an equilibrium where people pay less attention as these systems get better.
> We have empirical data showing how safe actual level 2 self driving cars are in practice.
Do we? Where does that come from? The data Tesla provides is hopelessly non-representative because it makes the assumption that the safety of any given road is independent of whether a driver chooses to switch on the system there.
Only overall numbers actually mater here, if self driving is off then that’s just the default risk from human driving in those conditions. Talk to your insurance company, they can give you a break down by make, model, and trim levels.
These are exactly my arguments to my girlfriend on why she shouldn’t use the Autopilot on our Tesla. Your mind will stray, the feature is exactly meant to do that to you. The feedback loop goes the wrong way. Then boom you don’t see emergency vehicles at a wreck apparently. I do blame Elon, he did the Silicon Valley thing of just promise a lot of untested stuff before the laws have solidified. Uber, Lime scooters, etc. The Tesla is a great car, but self-driving is orders of magnitude harder than he thinks.
Agreed. I'd also add that other car manufacturers have made tradeoffs on safety issues for decades.
So I wonder if it's more about Telsa capitalizing on the hype of self driving cars (with the expensive self-driving add-on) in the short term and less about him misunderstanding the magnitude of difficulty.
Telsa is using the proceeds from that add-on to make them seem more profitable and fund the actual development. It's smart in some aspects, but very risky to consumers and Telsa.
If you go back a few years, there were clearly expectations being set around L4/5 self-driving that that have very clearly not been met.
I still wonder to what degree this was a collective delusion based on spectacular but narrow gains mostly related to supervised learning in machine vision, how much was fake it till you make it, and how much was pure investor/customer fleecing.
I drive a Tesla and don't use the self-steering feature exactly because of this. What I do instead is enable the warnings from the same software like the ones you describe. That is actually a large gain. I'm already paying attention as I'm driving the car at all times and the software helps me catch things I haven't noticed for some reason. Those features seem really well done as the false positives are not too frequent and just a nuisance but the warnings are often valuable.
Fully automated Self driving cars is either a pipe dream or decades away in which many more people will be killed on the road in the name of technological progress.
Fully autonomous cars are already a reality with Waymo in AZ and AutoX, Baidu in China. I don't know how safe the Chinese companies are, but Waymo's safety record [1] is nothing short of stellar.
Waymo selected the one state willing to entirely remove any safety reporting requirements for self-driving cars as the place to launch their service. Regardless of what they claim to the contrary, if they had confidence in their safety record, they would've launched it in California, not Arizona.
Waymo has lied about the capabilities of their technology regularly, and for that reason alone, should be assumed unsafe. A former employee expressed disappointment they weren't the first self-driving car company to kill someone, because that meant they were behind.
> Regardless of what they claim to the contrary, if they had confidence in their safety record, they would've launched it in California, not Arizona.
California only months ago opened up permits for paid robotaxi rides. So no, they couldn't have launched it in CA. If you've noticed, they actually are testing in SF with a permit.
> Waymo has lied about the capabilities of their technology regularly, and for that reason alone, should be assumed unsafe.
What lies? Their CA disengagement miles are for everyone to see, their safety report is open, they have had 0 fatalities in their years of operation. Seems like you just made this up.
I recall a particular incident where Waymo was marketing their car being able to drive a blind man to a drive-thru, way before the thing could safely drive more than a mile on it's own. My understanding is that in 2021, it still can't navigate parking lots (which would preclude using it for drive-thrus).
Later, they were talking about how sophisticated their technology was: It can detect the hand signals of someone directing traffic in the middle of an intersection. Funny that a few months later, a journalist got an admission out of a Waymo engineer that the car wouldn't even stop at a stoplight unless the stoplight was explicitly mapped (with centimeter-level precision) so the car knew to look for it and where to look for the signal.
The article is seven years old at this point, but it's also incredibly humbling in how much bull- Waymo puts out, especially compared to the impressions their marketing team puts out. (Urmson's son presumably has a driver's license by now.)
In at least one scenario, the former Waymo engineer upset he had failed to kill anyone yet ("I’m pissed we didn’t have the first death"), caused a hit-and-run accident with a Waymo car, and didn't report it to authorities, amongst other serious accidents: https://www.salon.com/2018/10/16/googles-self-driving-cars-i... Said star Waymo engineer eventually went to prison for stealing trade secrets and then got pardoned by Donald Trump. Google didn't fire him for trying to kill people, they only really got upset with him because he took their tech to Uber.
I'd say Waymo has a storied history of dishonesty and coverups, behind a technology that's more or less a remote control car that only runs in a narrow group of carefully premapped streets.
> I recall a particular incident where Waymo was marketing their car being able to drive a blind man to a drive-thru, way before the thing could safely drive more than a mile on it's own.
How is a marketing video relevant from 2015 relevant to their safety record? They weren't even operating a public robotaxi service back then.
> My understanding is that in 2021, it still can't navigate parking lots (which would preclude using it for drive-thrus).
Completely false. Here is one navigating a Costco parking lot (can't get any busier than that) [1]. If you watch any videos in that YouTube channel, it picks you up and drops you off right from the parking lot. Yes, you can't use it for drive-thrus, but it doesn't qualify as "lying about capabilities".
> Later, they were talking about how sophisticated their technology was: It can detect the hand signals of someone directing traffic in the middle of an intersection. Funny that a few months later, a journalist got an admission out of a Waymo engineer that the car wouldn't even stop at a stoplight unless the stoplight was explicitly mapped (with centimeter-level precision) so the car knew to look for it and where to look for the signal.
Here is one recognizing a handheld stop sign from a police officer while it stopped for an emergency vehicle [2].
Good for Waymo and hopefully Google keeps up this science project. But it's a very limited and almost as perfect an environment as you could have outside of a controlled test area. Those who were saying L4/5 would be decades at least away seem to be those who were on the right track. Kids growing up today are going to have to learn to drive.
L5 may be decades away. I think we will see L4 in some major metro areas in the US by end of this decade. SF is heating up with Cruise and Waymo's heavy testing. Their progress will be a great indicator for true city driving.
>we will see L4 in some major metro areas in the US by end of this decade
I think you're far more likely to see L4 on limited access highways in good weather. A robotaxi service in a major city seems much more problematic given all the random behavior by other cars, pedestrians, cyclists, etc. and picking up/dropping off people in the fairly random ways that taxis/Ubers do. (And you'll rightly be shut down 6 months for an investigation the first time you run over someone even if they weren't crossing at a crosswalk.)
And for many people, including myself, automated highway driving would actually be a much bigger win than urban taxi rides which I rarely have a need for.
Will changes such as machine-readable road markings, car to car communications, and traffic management systems make this happen quicker.
For example, couldn't emergency vehicles could send out a signal directly to autonomous vehicles or via a traffic managagemnt system to slow down or require the driver to take over when approaching. An elementary version of this is Waze which will notify you of road hazards or cars stopped on the side of the road.
which many more people will be killed on the rise in the name of technological progress.
Seeing as car crashes are the leading cause of deaths from people aged 1-54, it may be an improvement from the status quo
More than 38,000 people die every year in crashes on U.S. roadways. The U.S. traffic fatality rate is 12.4 deaths per 100,000 inhabitants. An additional 4.4 million are injured seriously enough to require medical attention. Road crashes are the leading cause of death in the U.S. for people aged 1-54.
I'd say it depends on how many of those deaths are caused by the driver doing something unsafe. I'd be more comfortable with higher traffic deaths that primarily affect bad drivers than a lower number of deaths randomly spread across all drivers by a blackbox algorithm.
If you are texting while driving and hit a stopped car or run a red light, you are very lightly to kill others. Actually more likely, as a side impact is more dangerous than a frontal one.
> Road crashes are the leading cause of death in the U.S. for people aged 1-54
This isnt true according to the CDC. Cancer and heart disease lead for the 44-54 group, and while "accidental injury" does lead from 1-44, if you break down the data, in many cases vehicle based accidents are not not the largest single source. For example:
I agree with everything you said. I do hope that eventually the tech gets to the point where it can take over full time. We recently took a road trip for our vacation and the amount of road rage we witnessed was ... mind boggling. Don't get me wrong, not everyone is a raging asshole, but there were enough to make me wonder just why so many people are so freaking angry.
To me they are really aids. Of course you keep being concentrated, but I found that it takes out a lot of mental load like keeping the car straight, constantly tweaking the accelerator, etc..
It just makes the trips easier on the brain, and thus, for me, safer overall: its easier to see the overall situation when you've got free mental capacity
Yes. This also makes me kind of nervous when just using normal car adaptive cruise control. I feel as though my foot needs to be hovering near the pedal anyway and that's often less comfortable than actually pushing on the pedal and controlling it myself.
It can be really jarring too when a car behaves differently than you expect: I regularly use cruise control on my Kia, which makes driving much less stressful. It keeps the car centered in the lane, more or less turns the car with the road, and of course, matches the speed of the car in front of it with reasonable stopping distance. I wouldn't call it "self-driving" by any means, but if not for the alert that gets ticked off if your hands are off the wheel too long, it'd probably go on it's own for quite a long time without an incident.
However, I also once so far have experienced what happens when this system experiences a poorly-marked construction zone. Whilst most construction sites on the interstate system place temporary road lines for lane shifts, this one solely used cones. While I was paying attention and never left the flow of traffic, the car actually fought a little bit against me following the cones into another lane, because it didn't see the cones, it was following the lines.
It doesn't surprise me at all that if someone gets too comfortable trusting the car to do the work, even if they think they're paying attention, they could get driven off the roadway.
I was thinking about this the other day - driving in construction. The town I live in is currently doing water main replacement. So, lots of torn up roads, closed lanes and even single-lane only with a flagger alternating directions. No amount of safety cones will make it obvious what's going on.
How do automated systems deal with flaggers? Visibility of the stop/slow sign isn't sufficient to make a determination on whether it's safe to proceed (not to mention "stop" changes meaning here, entirely, from a typical stop sign). Often, whether or not you can proceed comes down to hand gestures from the flagger proper.
Not that I expect any reasonable driver to be using something like autopilot through such a situation, but we've also seen plenty of evidence that there are unreasonable drivers currently using these systems, as well.
Conceivably in the somewhat-near future (10 years+), most cars on the road will have some sort of ADAS system, in which I'd presume it'd start to make sense for construction to use some sort of digital signalling. Something like a radio signal broadcast that can send basic slow/stop flagging signals to a lane of traffic.
Of course, the problem is, if we haven't developed it today, the ADAS systems of today won't understand it in ten years when there's enough saturation to be practical to use it. Apart from Tesla, very few car manufacturers are reckless enough to send OTA updates that can impact driving behavior.
Lane-following ADAS systems of today, mind you, can work relatively fine in construction areas... provided lane lines are moved, as opposed to relying solely on traffic cones.
It is my belief that the most ideal form of truly self driving vehicles will not happen until a time when vehicles can talk to each other on the road to make each other aware of position and speed data. I don't think this has to be full GPS coordinates at all. This is about short range relative position information.
A mesh network of vehicles on the road would add the ability for vehicles to become aware of far more than a human driver can ever know. For example, if cars become aware of a problem a few km/miles ahead, they can all adjust speed way before encountering the constriction in order to optimize for traffic flow (or safety, etc.).
Of course, this does not adequately deal with pedestrians, bikes, pets, fallen trees, debris on the road, etc.
Not saying cars would exclusively use the mesh network as the sole method for navigation, they have to be highly capable without it. The mesh network would be an enhancement layer. On highways this would allow for optimization that would bring forth some potentially nice benefits. For example, I can envision reducing emissions through traffic flow optimization.
Remember that electric cars still produce emissions, just not necessarily directly while driving. The energy has to come from somewhere and, unless we build a massive number of nuclear plants, that somewhere will likely include a significant percentage of coal and natural gas power plants.
The timeline for this utopia is likely in the 20+ year range. I say this because of the simple reality of car and truck ownership. People who are buying cars today are not going to dispose of them in ten years. A car that is new today will likely enter into the used market in 8 to 10 years and be around another 5 to 10. The situation is different with commercial vehicles. Commercial trucks tend to have longer service lives by either design or maintenance. So, yeah, 20 to 30 years seems reasonable.
Also, these cars know the speed limits for the road but let you set cruise control/self driving above the speed limit. Seems like for safety purposes that should not be allowed. Not only are people paying significantly less attention but they also are speeding.
I'm not sure how the cars know the speed limit. Maybe someone else knows? My guess is combo of GPS/camera to position correctly on road and the lookup of known speed limit data. Perhaps it reads signs though?
My car shows the speed limit of roads it knows. It uses GPS and stored limits. It also doesn't know the limits of non-major roads and doesn't attempt to show a limit then. My car is a 2013, and I've not paid the $$ to update the maps in that time (seriously, they want $200-$400 to update the maps).
Since I bought my car, Illinois (where I live) has raised the maximum limit on interstates by 10 MPH. My car doesn't know about it. If my car limited me to what it thought the limit was, I'd probably be driving 20 MPH slower than prevailing traffic, a decidedly unsafe situation.
The rental car I am using now certainly a) reads road signs for speed limit information, b) is definitely fooled by signs on off ramps etc.
It’s hard to imagine how speed limit systems would work without some sort of vision capabilities — a database of speed limits would never be up to date with roadworks and so on.
Different manufacturers probably use different systems but no. BMW attempts to read the speed limit signs using the frontal camera with a mix of some sort of stored info - it knows that the speed limit is about to change (Mobileye?), but it is very often that it won't catch a sign in the bend or when the weather is bad. Also, it does not recognize time restricted speed limits, for example 30kph from 7:00 to 17:00 Monday to Friday so it would keep driving 30kph outside of those hours while 50kph is allowed. In some places in Germany, it does not recognize the city limits and carries on showing 70kph for a kilometer longer than it should.
That’s a feature accommodating realities of driving on public roads, not a bug.
If you drive on a 60mph speed limit highway, no one is driving 60mph, everyone is going around 70mph. If you decide to use autopilot and it limits you to 60mph, you singlehandedly start disrupting the flow of traffic (that goes 70mph) and end up becoming an increased danger to yourself and others.
Not even mentioning cases when the speed limits change overnight or the map data is outdated or if a physical sign is unreadable.
Over the road trucks often have speed governors, some companies limit their trucks to 60 mph because it saves a lot of fuel and leads to a much (50%) lower risk of collisions.
Apples to oranges. Stopping distance of a 16-wheeler is magnitudes larger than that of a typical sedan, so in their case it makes sense.
For specific numbers (after subtracting reaction distance being the same for both):
55mph: car 165ft, 16-wheeler 225ft.
65mph: car 245ft, 16-wheeler 454ft.
As you can see, the gap between a car's stopping distance and a 16-wheeler's stopping distance increases with speed increasing, and non-linearly at that. Not even mentioning the destructive potential of a car vs. a 16-wheeler.
I would agree with your point if majority of the roads were occupied by 16-wheelers, but it isn't the case (at least in the metro area that I commute to work in).
Note: I agree that it would be safer if everyone drove the exact speed limit, as opposed to everyone going 10mph above the speed limit. However, in a situation where everyone is driving 10mph above the speed limit, you are creating a more dangerous situation by driving 10mph slower instead of driving 10mph above like everyone else.
I see a lot of comments here postulating how autopilot is a terribly designed feature from people who appear not to be speaking from first hand experience and now I feel compelled to comment, exactly following that HN pattern someone posted about how HN discussions go. That said thanks for keeping this discussion focused & framed as a system design one, doesn't feel like a Tesla hate train so I feel comfortable hoppin' in and sharing. This is a little refreshing to see.
Anyway, perhaps I'm in a minority here, but I feel as though my driving has gotten _significantly safer_ since getting a Tesla, particularly on longer road trips.
Instead of burning energy making sure my car stays in the lane I can spend nearly all my time observing drivers around me and paying closer attention farther down the road. My preventative and defensive driving has gone up a level.
> I just don't understand how you can have a "self driving car but you must but be ready to put your hands back on the steering wheel and your foot on the pedal(s)".
I've not hit animals and dodged random things rolling/blowing into the road at a moment's notice. This isn't letting autopilot drive, it's like a hybrid act where it does the rote driving and I constantly take over to quickly pass a semi on a windy day, not pass it on a curve, or get over some lanes to avoid tire remnants in the road up ahead. I'm able to watch the traffic in front and behind and find pockets on the highway with nobody around me and no clumping bound to occur (<3 those).
To your suspicion, it is a different mode of driving. Recently I did a roadtrip (about half the height of the USA) in a non-Tesla, and I found myself way more exhausted and less alert towards the end of it. Could be I'm out of habit but egh.
Anyway, so far I've been super lucky. I don't think it's possible to avoid all car crashes no matter how well you drive. But I _for sure_ have avoided avoidable ones and taken myself out of situations where they later occurred thanks to the extra mental cycles afforded to me by auto-pilot. My safety record in the Tesla is currently perfect and I'll try and keep it that way.
I don't think autopilot is perfect either but I do think it's a good tool and I'm a better driver for it. Autopilot has definitely helped me spend better focus on driving.
This expresses the mindset I find myself in when I use Autopilot. It's like enabling cruise control, you're still watching traffic around you but now you don't need to focus on maintaining the correct speed or worry about keeping your car perfectly in a lane. You can more or less let the car handle that (with your hands on the wheel to guard against the occasional jerky maneuver when a lane widens for example) while you focus on the conditions around you.
Exactly. It frees the driver from increasingly advanced levels of mundane driving (cruise control manages just speed, adaptive cruise also deals with following distance, lane keeping deals with most of the steering input, etc) allowing the driver to focus more on monitoring the situation and strategic portion of driving rather than the tactical. Of course, this relies on the driver to actually do that. They could just use devote that extra attention to their phone.
Exactly this. I treat AP like I'm letting a learner drive. Constantly observing to make sure it's doing the right thing. I've been on long road trips and with AP my mind stays fresh for much longer compared to with other cars.
Some people activate cruise control and then rest their right foot on the floor. I activate cruise control whenever possible because while it is activated, I can drive with my foot resting on the brake pedal. I like being marginally more responsive to an event that requires braking since I don’t need to move my foot from the accelerator.
There's zero Tesla hate here and certainly zero EV hate here, on the contrary: I just feel the interior build quality on the Tesla could be a bit better but I'm sure they'll get there.
I wouldn't want my, strangely enough upvoted a lot, comment, to be mistaken for Tesla hate. I like what they're doing. I just think the auto-pilot shouldn't give a false sense of security.
> I've not hit animals and dodged random things rolling/blowing into the road at a moment's notice.
> I don't think it's possible to avoid all car crashes no matter how well you drive.
Same here... And animals are my worst nightmare: there are videos on YouTube just terrifying.
For I do regularly watch crash videos to remind me of some of the dangers on the road.
I think you two are talking about different things.
You're talking about Autopilot which is just driver assistance technologies; lane keep assistance, adaptive cruise control, blind spot monitoring, etc. It's not to replace driver attention, it's just monitor sections of the road the the driver can't pay attention to full time. The driver is still remaining in control and attentive to the road.
The person you're responding to seems to be talking talking about the Full Self Driving feature who's initial marketing implied that the driver need not be mentally engaged at all or too impaired to drive normally. Which was later back pedal led to say that you need to pay attention.
The problem is, even if your subjective idea of how Autopilot affects your own driving is correct, it appears not to be the case for a significant subset of Tesla drivers, enough that they've been plowing into emergency vehicles at such an elevated rate as to cause NHTSA to open an investigation.
Also, your subjective impressions may be what they are simply because you have not yet encountered the unlucky set of conditions which would radically change your view, as was surely the case for all the drivers involved in these sorts of incidents.
I learned this playing Gran Turismo video games way back when. The game has long endurance races (I seem to remember races that ran about 2 hours, but there may have been longer ones). Eventually you get hungry or thirsty or have to use the bathroom, so you pause the game, take care of business, and resume. It's really easy to screw up if the game was paused while your car was doing anything other than stable, straight travel. A turn that I successfully handled 100 times before can suddenly feel foreign and challenging if I resume there with little context.
Obviously that's not exactly the same thing as taking over for a real car when the driver assistance features give up, but seems similarly challenging to take over the controls at the most precarious moment of travel, without being sort of "warmed up" as a driver.
> like the truck driver in a big curve slightly deviating out of his lane and "pushing" me dangerously close to the median strip for example
This is a situation that you simply shouldn’t put yourself in. There is no reason to ever drive right next to a large vehicle, on either side, except for very short periods when overtaking them on a straight road.
This just isn't realistically possible on most highways except in the lightest traffic conditions. You are gonna spend some time beside trucks whether you like it or not.
Spending time right next to a truck is completely optional. You can either speed up or slow down, either of them will put you in a position where you are no longer right next to them.
Exactly, the only driving assistance feature I use is adaptive cruise control, and I don't have plans to use anything more. If I trust autonomous systems too much, I would not be ready when it matters.
Did I not just say this before? [0] [1] [2], seems like the safety probe formally agrees with my (and many other's) concerns over the deceptive advertising of the FSD package and its safety risks towards other drivers.
NHTSA is lumping TACC with Autopilot to increase number of incidents and make the case sound more serious:
"The involved subject vehicles were all confirmed to have been engaged in either Autopilot or Traffic Aware Cruise Control during the approach to the crashes," NHTSA said in a document opening the investigation.
> In one of the cases, a doctor was watching a movie on a phone when his vehicle rammed into a state trooper in North Carolina.
Doesn't autopilot require you to put your hands on the wheel fairly regularly? Are these incidents just a matter of people using this feature outside of its intended use case?
My thought would be that we hear lots of anecdotes because they claim a full self driving capability, so it’s particularly interesting when they crash and are using these assist features. Indeed when I bought my Model Y I paid extra for that capability.
Who is going to pay for these dedicated stretches of highway that only, presumably, relatively wealthy owners of self-driving cars are going to be allowed to use?
Any entity (individuals or corporate) could use it of course. Rich or not, since Electric vehicles such as buses could be a public form of transportation on these specially adapted roads.
Let me preface by saying that I hold no strong opinions on this matter and my comments are purely speculative.
This is kind of a position I've held for a long time but a different aspect of the problem. I think a system similar to IFF in aircraft would solve all of these issue. If every car knew where every other car was at all times, you could easily devise a system that would be nearly flawless. The issue is, there is no incremental path to this solution. You would essentially have to start over with the existing transportation network.
The problem is that you don’t just need to know about every other vehicle, you still need all the perceptual stuff for pedestrians, bikers, baby carriages, trash, road closures, traffic cops in the middle of the road, etc. All those things are arguably harder to detect reliably than a somewhat standard sized box of metal with pairs of lights in the front and back. I think shooting for superhuman perception of all these things is still where Tesla is failing.
True. I guess I was thinking that if you build totally new infrastructure for these new overhauled cars, you'd keep it completely separate from other modes of transportation. My sci-fi inclinations had me imagining tubes like Logan's Run.
>The question should be - how many lives were saved by this system vs how many would die if driven "normally"?
It is also necessary to project this into the future, i.e. looking at the integral of expected lives lost 'rushing' self driving cars vs. 'waiting-and-seeing' (as Americans die at a rate of 40,000 per annum).
If twice as many people die for a fixed number of years to create a self driving system that results in half the fatality rate of the status quo, that becomes worth it very, very quickly.
When I was learning to drive, my grandmother drilled into me to never swerve for an animal that jumps out in front of the car. This saved me when I was driving by the Grand Canyon and jack rabbits kept jumping, out of nowhere seemingly, in front of my car. I drilled "never swerve" into my son and it saved him on a mountain road when he hit a deer. He didn't go into the trees. When I drove in Alaska I asked why the forest was cut back from the road. They said that moose like to step out in front of cars.
I have no idea how self-driving fits into this. I don't have a feel how self-driving responds to emergencies. I'd have to experience an emergency in one. For that reason, I don't see myself ever trusting self-driving.
1. always pay full attention to where you are because there might be a truck or a family of 5 coming from the opposite direction, 2. never lift, 3. always look in the direction you want to travel in, not in the direction you currently travel
When my wife learned to drive in Norway she was instructed never to swerve to avoid a collision regardless of whether it was an elk, a dog or a human being in front of the car, just to stamp hard on the brake.
The rationale being that swerving most likely puts more people at risk more of the time. Especially true here where leaving the road often means either colliding with the granite cliff wall or ending up in the fjord or lake.
I don't know how Tesla ever presumes to achieve FSD when they cannot detect stopped objects in the road, especially emergency vehicles. This is incredibly disappointing.
Does anyone know if the FSD Beta has this ability?
All these crashes happened with the radar-controlled auto pilot. A radar system basically cannot detect static obstactles, as it doesn't have the spacial resolution to distinguish an obstacle in your lane from something right beside or above it (bridges). They can only use the radar to follow other cars, because these are not stationary objects.
Recently, Tesla switched from radar-based to pure optical obstacle recognition. This should vastly improve this kind of behavior. Ironically that the investigation starts at a moment when they basically got rid of the old system.
Look on youtube for videos of the FSD beta. It is amazingly good at recognizing the surroundings of a car, including parked vehicles at the road side.
I believe we should have safety probes. Lots of people who have taken money from the auto industry want this specifically for Tesla. There is a strong possibility this is political punishment for wrongthink.
506 comments
[ 4.8 ms ] story [ 306 ms ] threadP.S. /s. Obviously, Mr. Poe.
This is why 'attention' and 'driver monitoring' was not included.
It's okay. I do it too. Really need to work on seeing yourself making that argument as a starting point and not an endpoint.
I will be amused, intrigued, and possibly a bit horrified if by the time FSD hits level 5, and they stick with the Neural Net of Neural Nets architecture if there isn't a rash of system induced variance in behavior as emergent phenomena take shape.
Imagined news: All Tesla's on I-95 engaged in creating patterns whereby all non-Teala traffic was bordered by a Tesla on each side. Almost like a game of Go, says expert. Researchers stumped.
Then again, that's imply you had an NN capable of retraining itself on the fly to some limited degree, which I assume no one sane would put into service... Hopefully this comment doesn't suffer a date of not aging well.
[0] https://theoatmeal.com/blog/google_self_driving_car
[1] https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg
[2] https://youtube.com/watch?v=Wz6Ins1D9ak
My car has adaptive cruise control and lane keep assist, but I'm not relying on either for anything more complex than sipping a drink while on the highway.
Car traffic and streets are more dense and often have humans crossing them without regards to laws, bicycles, motorbikes, road construction and bad weather.
Not saying one auto pilot system is better than the other, however, they operate in different environments.
I need to be touching the wheel and applying some force to it or it begins yelling at me and eventually brings me slowly to a stop.
I’ve had it for a year now and I cannot perceive of a way, without physically altering the system (like hanging a weight from the wheel maybe?) that would allow me to stop being an active participant.
I think the opposite is true: Tesla’s move fast and kill people approach is the mistake. Incremental mastering of autonomous capabilities is the way to go.
> physically altering the system (like hanging a weight from the wheel maybe?)
was exactly what people were doing. But it's also possible to be physically present, applying force, but being "zoned out", even without malicious intent.
I've occasionally noticed myself zoning out behind the wheel of my non-self-driving car as well.
It's actually very common. [0]
[0] https://www.actuarialpost.co.uk/article/quarter-of-fatal-cra...
> I’ve had it for a year now and I cannot perceive of a way, without physically altering the system (like hanging a weight from the wheel maybe?) that would allow me to stop being an active participant.
That's exactly what people were doing with the Tesla. Hanging a weight to ensure the safety system doesn't kick in. [0][1]
[0] https://edition.cnn.com/2021/04/28/cars/tesla-texas-crash-au...
[1] https://twitter.com/ItsKimJava/status/1388240600491859968/ph...
I guess the probe will reveal what share of fatal accidents are caused by this.
Teslas can famously be tricked by wedging an orange between the rim and spoke of the steering wheel to produce enough torque on the wheel to satisfy the detection. There are enough videos of it on youtibe that tesla could easily be found negligent for not doing enough to prevent drivers from defeating a safety system given that alternate technology that more directly tracks attention is available and tricking tesla's detection method became common knowledge.
Personally Autopilot has actually made driving safer for me... I think there's likely abuse of the system though that Tesla could work harder to prevent.
They are sending multiple signals that this car can drive itself (going so far as charging people money explicitly for the "self-driving" feature) when it cannot in the slightest do much more than stay straight on an empty highway.
They should be forced to change the name of the self-driving features, I personally think "Backseat Driver" would be more appropriate.
It is literally an autopilot. Just like an autopilot on an airplane, it keeps you stable and in a certain flight corridor. There's virtually no difference except for Tesla's Autopilot's need to deal with curved trajectories.
Well, and it actively avoids collisions with other vehicles (most of the time). Airplane (and boat) autopilots don't do that.
"But you're using the word autopilot wrong!"
Damn, this 'drink verification can' is going to get us all killed.
People causing these problems almost certainly are putting something over the cabin camera and a defeat device on the steering wheel.
"The average driver" includes everyone, ranging from drivers using it as intended with close supervision, drivers who become inattentive because nothing is happening, and drivers who think it's a reasonable idea to climb into the back seat with a water bottle duct taped to the steering wheel to bypass the sensor.
OTOH, the average driver for the unassisted scenario also includes the driver who thinks they're able to drive a car while texting.
Shouldn't that compared to "average driver + myriad of modern little safety features" instead of "average unassisted driver"? The one who has the means to drive a Tesla with the "full driving" mode certain has the means to buy, say, a Toyota full of assistance/safety features (lane change assist, unwanted lane change warning and whatnots).
Making it a crime isn't an "obvious solution" to actually make it not happen. Drunk driving is a crime and yet people keep doing it. Same with texting and driving.
Prevention as a goal is how we end up with dystopia.
I'm not saying improvements should stop there, but once the system has reached parity, it's OK to deploy it and let it improve from there.
An imperfect self driving vehicle is the worst of all worlds: they lull the driver into the perception that the vehicle is safe while not being able to handle abnormal situations. The fact that there are multiple crashes on the record where Telsas have driven into stationary trucks and obstacles on roads is pretty damning proof that drivers can't always react in the time required when an imperfect self driving system is in use. They're not intrinsically safe.
At the very least drivers should be required additional training to operate these systems. Like pilots, drivers need to be taught how to recognize when things go awry and react to possible failures. Anything less is not rooted in safety culture, and it's good to see there are at least a few people starting to shine the light on how these systems are being implemented from a safety perspective.
Nothing absurd about thinking a system which has parity with the average human driver is too risky to buy unless you consider yourself to be below average at driving. (As it is, most people consider themselves to be better than average drivers, and some of them are even right!) The accident statistics that comprise the "average human accident rate" are also disproportionately caused by humans you'd try to discourage from driving in those circumstances...
Another very obvious problem is that an automated system which kills at the same rate per mile as an average human drivers will tend to be driven a lot more because no effort (and probably replace better-than-average commercial drivers long before teenagers and occasional-but-disproportionately-deadly drivers can afford it).
Doesn't this hide a paradox? Using a self-driving car as intended implies that the driver relinquishes a part of the human decision making process to the car. While close supervision implies that the driver can always take control back from the car, and therefore carries full personal responsibility of what happens.
The caveat here is that the car might make decisions in a rapidly changing, complex context which the driver might disagree with, but has no time to correct for through manual intervention. e.g. hitting a cyclist because the autonomous system made an erroneous assertion.
Here's another way of looking at this: if you're in a self-driving car, are you a passenger or a driver? Do you intend to drive the car yourself or let the car transport you to your destination?
In the unassisted scenario, it's clear that both intentions are one and the same. If you want to get to your location, you can't but drive the car yourself. Therefore you can't but assume full personal responsibility for your driving. Can the same be said about a vehicle that's specifically designed and marketed as "self-driving" and "autonomous"?
As a driver, you don't just relinquish part of the decision making process to the car, what essentially happens is that you put your trust in how the machine learning processes that steer the car were taught to perceive the world by their manufacturer. So, if both car and occupant disagree and the ensuing result is an accident, who's at fault? The car? The occupant? The manufacturer? Or the person seeking damages because their dog ended up wounded?
The issue here isn't that self-driving cars are inherently more dangerous then their "dumb" counter parts. It's that driving a self-driving car creates it's own separate class of liabilities and questions regarding responsible driving when accidents do happen.
The fact that an huge formal investigation happened with just a single casualty is proof that it may actually be superior for safety in the long-term (when combined with feedback from regulators and government investigators). One death in conventional vehicles is irrelevant. But because of the high profile of Tesla’s technology, it garners a bunch of attention from the public and therefore regulators. This is PRECISELY the dynamic that led to the ridiculously safe airline record. The safer it is, the more that rare deaths will be investigated and the causes sussed out and fixed by industry and regulators together.
Perhaps industry/Tesla/whoever hates the regulators and investigations. But I think they are precisely what will cause self driving to become ever safer, and eventually become as safe as industry/Tesla claims, safer than human drivers while also being cheap and ubiquitous. Just like airline travel today. A remarkable combination of safety and affordability.
This might be the only way to ever do it. I don’t think the airline industry could’ve ever gotten to current levels of safety by testing everything on closed airfields and over empty land for hundreds of millions of flight hours before they had sufficient statistics to be equal to today.
It can’t happen without regulators and enforcement, either.
If anything a driver monitoring system seems even better than the interlock system, for example you couldn't have your kids/friends blow for you to bypass it.
Go with safest drivers first.
https://www.tesla.com/en_CA/VehicleSafetyReport
I think the comparison should be Tesla with/without AI, not Tesla/not-Tesla; so roughly either x2 or x4 depending on what the other active safety features do.
It’s not nothing, but it’s much less than the current sales pitch — and the current sales pitch is itself the problem here, for many legislators.
This still isn't the correct comparison. Major selection bias with comparing miles with autopilot engaged to miles without it engaged, since autopilot cannot be engaged in all situations.
A better test would be to compare accidents in Tesla vehicles with the autopilot feature enabled (engaged or not) to accidents in Tesla vehicles with the autopilot feature disabled.
https://www.iihs.org/topics/fatality-statistics/detail/urban...
Yes it can. The only time it can't be activated is if there is no clearly marked center line.
The real test of this would be: of all Tesla vehicles, are the ones with autopilot enabled statistically safer or less safe than the ones without autopilot enabled?
Makes me laugh, especially with the “geeks are immune to marketing” trope that floats around here equally as regularly.
1. Autopilot only works on (or intended to work on) highways. But they are comparing their highway record to all accident records including city driving, where accident rate is far higher than highway driving.
2. They're also comparing with every vehicle in the United States including millions of older vehicles. Modern vehicles are built for higher safety and have a ton of active safety features (emergency braking, collision prevention etc). Older vehicles are much more prone to accidents and that skews the numbers.
The reality is Teslas are no safer than any other vehicles in its class ($40k+). Their safety report is purely marketing spin.
Would another way of saying this be that they are as safe as other vehicles in that class? And that therefore Autopilot is not more unsafe than driving those other cars?
We should really compare Autopilot with its competitors like GM’s Super Cruise or Ford’s Blue Cruise, both of which offer more capabilities than Autopilot. That will show if Tesla’s driver assist system is more or less safe than their competitors product.
The reason GM's Super Cruise got a higher rating by Consumer Reports was because CR didn't even test the capabilities that only Tesla had (such as automatic lane change and taking offramps/onramps). Also, the majority of the evaluation criteria weren't about capabilities. eg: "unresponsive driver", "clear when safe to use", and "keeping the driver engaged".[1]
1. https://www.consumerreports.org/car-safety/cadillac-super-cr...
AP has the luxury of being able to be turned off in less than ideal conditions. Human drivers can't do that.
It's the reason why only Tesla touts these numbers. They're inaccurate and misleading.
Asking someone to pay attention when they are not doing anything is unrealistic. I would be constantly bored / distracted. My wife would instantly fall asleep. Etc etc.
Conversely, when Waymo iterates and improves their own safety ratios by a significant amount, that evidently does not result in Tesla's improving in return.
Until it leads to something systemic, that then regulator mandates for all vehicles
Not the case at all between lidars and cameras.
Of course this is impossible in the real world. Nobody is going to buy a car that will randomly make its own decisions, that will pull the wheel from your hands ever time it thinks you are making an illegal lane change. Want safety? How about a Tesla that is electronically incapable of speeding. Good luck selling that one.
I really don't think that would give many data points, because all of the instances would be when a human fell asleep or wasn't paying attention.
Because when the human disagrees with the machine, the machine is usually the one making a mistake. It might prevent accidents, but it would also cause them, and you lose predictability in the process (you have to model the human and the machine).
Nothing more annoying then a car that thinks i don't know how to drive (warning beeps etc.).
That's almost exactly what my Honda does. Illegal (no signal) lane change results in a steering wheel shaker (and optional audio alert). And the car, when sensing an abrupt swerve which is interpreted as the vehicle leaving the roadway, attempts to correct that via steering and brake inputs.
But, I agree with your more general point - the human still needs to be primary. My Honda doesn't allow me to remove my hands from the steering wheel for more than a second or two. Tesla should be doing the same, as no current "autopilot" system is truly automatic.
To be clear, tying the warning to the signal isn't about preventing unsignaled lane changes, it's gauging driver intent (i.e. is he asleep and drifting or just trying to change lanes). It's just gravy that it will train bad drivers to use their signals properly.
> 142 (1) The driver or operator of a vehicle upon a highway before turning (...) from one lane for traffic to another lane for traffic (...) shall first see that the movement can be made in safety, and if the operation of any other vehicle may be affected by the movement shall give a signal plainly visible to the driver or operator of the other vehicle of the intention to make the movement. R.S.O. 1990, c. H.8, s. 142 (1).
That said there's zero cost to doing so regardless of whether other drivers are affected.
https://www.ontario.ca/laws/statute/90h08#BK243
Usually, it makes you liable because it is illegal. CA law for instance reqires signalling 100ft before a lane change or turn.
By the way, this is fucking terrifying when you first encounter it in a rental car on a dark road with poor lane markings while just trying to get to your hotel after a five hour flight.
I didn't encounter an obvious wheel shaker, but this psychotic car was just yanking the wheel in different directions as I was trying to merge onto a highway.
Must be what a malfunctioning MCAS felt like in a 737 MAX, but thankfully without the hundreds of pounds of hydraulic force.
As a simple example, ABS (https://en.wikipedia.org/wiki/Anti-lock_braking_system) only interferes with what the driver does when an error occurs.
More related to self-driving, there’s various variants of https://en.wikipedia.org/wiki/Lane_departure_warning_system that do take control of the car.
And it is far from “incapable of speeding”, but BMW, Audi and Mercedes-Benz “voluntarily” and sort-of limit the speed of their cars to 250km/hour (https://www.autoevolution.com/news/gentlemens-agreement-not-...)
That would be unsafe in many situations. If the flow of traffic is substantially above the speed limit--which it often is--being unable to match it increases the risk of accident. This is known as the Solomon curve [1].
[1] https://en.wikipedia.org/wiki/Solomon_curve
With the logic presented in the theoretical foundation section, it seems that the safer move would actually be slow down and match the speed of all the trucks and other large vehicles... which won't happen.
Matching speed sounds great, except there are always people willing to go faster and faster. In my state they raised the speed limit from 70 to 75, it just means more people are going 85-90. How is that safer?
However, you individually going 70-75 when everyone else is going 85-90 is less safe than you going 85-90 like everyone else in the exact same situation.
>there are always people willing to go faster and faster
That’s why no one says “go as fast as the fastest vehicle you see”, it is “go with the general speed of traffic”. That’s an exercise for human judgement to figure that one out, which is why imo it isn’t a smart idea to have the car automatically lock you out of overriding the speed limit.
And yet the roads are full of vehicles literally incapable of going 85. Many trucks cannot do more than 69mph.
(Trivial if the self driving tech works at all....)
There's just too much shit that can't be "seen" with a camera/sensor in conjested traffic. Having a swarm of vehicles all gathering/sharing data is one of the only true ways forward IMO.
My insurance company gives a lower rate if you buy the full autopilot option, and that to me indicates they agree it drives better than I, or other humans, do.
If following the speed limit makes cars safer, another way to achieve that without autopilot is to just have all cars limit their speed to the speed limit.
Sometimes it’s wonky when the street lines are unclear. It’s not perfect but a better driver than I am in 80% of cases
The problem is in those 20% of cases where you'd lulled into boredom by autopilot as you concentrate on designing your next project in your head, then suddenly autopilot says "I lost track of where the road is, here you do it!" and you have to quickly gain context and figure out what the right thing to do is.
Some autopilot systems use eye tracking to make sure that the driver is at least looking at the road, but that doesn't guarantee that he's paying attention. But at least that's harder to defeat than Tesla's "nudge the steering wheel once in a while" method.
The devil is in the details... GPS may not provide sufficient resolution. Construction zones. School zones with variable hours. Tunnels. Adverse road conditions. Changes to the underlying roads. Different classes of vehicles. Etc.
By the time you account for all the mapping and/or perception, you could've just improved the autonomous driving and eliminated the biggest source of humans driving: The human.
The parent post said that autopilot won't let him go over the speed limit and implies that makes him safer. My point is that you don't need full autopilot for that.
So this is not a technical problem at all, but a political one. As the past year has shown, people won't put up with any convenience or restriction, even if it could save lives (not even if it could save thousands of lives)
What would be even easier than all of that, though, is just installing speeding cameras and mailing tickets.
I have not yet struggled to stay alert when it drives me, and it has driven better than I would have - so it certainly is an improvement over me driving 100% of the time. It does not have road rage and it does not enjoy the feeling of speeding, like I do when I drive, nor does it feel like driving is a competition, like I must admit I do when I am hungry, stressed, or tired.
> just have all cars limit their speed to the speed limit
No way I’d buy a car that does not accelerate when I hit the pedal. Would you buy a machine that is not your servant?
I mean... that's an odd thing for someone to say who has bought a vehicle with over-the-air firmware updates.
I wish I could hack but car, but also wouldn’t trust others if they did
>>How about a Tesla that is electronically incapable of speeding. Good luck selling that one.
From 2022 all cars sold in the EU have to have an electronic limiter that keeps you to the posted speed limit(by cutting power if you are already going faster) - the regulation does allow the system to be temporarily disabled however.
I think it's a silly law and I'm very glad I don't live in a place that requires such annoyances, but it's not as bad as you're claiming.
1. https://etsc.eu/briefing-intelligent-speed-assistance-isa/
It could reliably recognize the speed limit signs (red circle), but it never recognized the similar grey-slash end-of-limit signs. It also didn't recognize the start-of-town or end-of-town signs, so it didn't do anything about the limits they implied.
I would certainly have had to disable it, had it been reducing the acceleration in the way that document describes.
Yeah, add to that the unreliability of Tesla's system means that it cannot pull the wheel from the driver, because it's not unusual for it to want to do something dangerous and need to be stopped. You don't want it to "fix" a mistake by driving someone into the median divider.
Instead they did the exact opposite with the plaid mode model S, lol. It kind of works against their claims that they prioritize safety when their hottest new car - fully intended for public roads - has as its main selling point the ability to accelerate from 60-120 mph faster than any other car.
People literally are buying cars that “make their own decisions”. Importantly though, these systems only activate in the case of an imminent collision IF the corrective measure won’t cause another collision.
> that will pull the wheel …
Yah, of course no one is going to buy a care that does what you describe because what you describe is insane and inherently unsafe. Unless a collision is imminent, nothing happens.
That is the fatal flaw in anything but a perfect system - any kind of system that is taking the decisions about steering from the driver is going to result in the driver at best thinking about other things and worse getting into the back seat to change. If you had to develop a system to make sure someone was paying attention, you wouldn't make them sit in a warm comfy seat looking at a screen - you would make them actively engage with what they were looking at - like steering.
And ultimately it doesn't matter how many hundreds of thousands of hours of driving you teach your system with, it may eventually be able to learn about parked cars, kerbs and road signs, but there won't be enough examples of different accidents and how emergency vehicles behave to ever make it behave safely. Humans can cope with driving emergencies fairly well (not perfectly admittedly) no matter how many they've been involved in using logic and higher level reasoning.
It is not solely the trust and dependence but inclusive is the group of idiots with access to wealth without regard to human life.
These things are on the road already. They have issues, but so do human only cars. Tweaks probably get made, like some special handling of emergency vehicle scenarios. But, it's not enough to stop it.
Meanwhile, it's not a permanent state. Self driving technology is advancing, becoming more common on roads. Procedures, as well as infrastructure, is growing around the existence of self driven cars. Human supervisor or not, the way these things use the road affects the design of roads. If your emergency speed sign isn't being headed by self driven cars, your emergency speed sign has a bug.
Tesla auto-drive seems like it's about 80% of the way there.
It's just that the companies that are NOT doing incremental approaches are largely at the mercy of some investors who don't know a thing about self-driving, and they may die at any time.
I agree with you that it is technically flawed, but it may still be viable in the end. At least their existence is not dependent on the mercy of some fools who don't get it, they just sell cars to stay alive.
That's one of the major problems of today's version of capitalism -- it encourages technically flawed ways to achieve scientific advancement.
It therefore isn't a clean swap of a human paying attention to a human who isn't. It becomes a complicated equation that we can't just dismiss with "people won't pay attention". It is possible that a 90%/10% split of drivers paying attention to not paying attention is more dangerous when they are all driving manually than a 70%/30% split if those drivers are all using self-driving tech to cover for them. Wouldn't you feel safer if the driver behind you who is answering texts was using this incremental self-driving tech rather than driving manually?
No one has enough data on the performance of these systems or how the population of drivers use them to say definitively that they are either safer or more dangerous on the whole. But it is definitely something that needs to be investigated and researched.
We've known for a very long time that this sort of automation/manual control handoff failure is a very big deal, and yet there seems to be an almost willful blindness from the manufacturers to address it in a meaningful way.
Conversely, if it didn't warn me right before an accident, then the absence of that warning would be useful too.
All of that information should be put back into the model based on crash reporting. Everything else can be ignored.
I would argue that the information should be available to all automakers (perhaps using the NHTSA as a conduit), so that each of them have the same safety information, but can still develop their own models. The FAA actually does this already with the FAA Accident and Incident Data Systems [0] and it has worked pretty darn well.
[0] - https://www.asias.faa.gov/apex/f?p=100:2:::NO:::
It also reads the speed limit signs and places a reminder in the display. I think it can brake if it detects something in front of it, but I'm not certain.
My main point (perhaps buried more than it should have been) is that by centralizing accident data along with whether an alert went off (or not), and sharing that with all automobile manufacturers can help this process proceed better.
Right now the data is highly fragmented and there is not really a common objective metric by which to make decisions to improve models.
I agree that would be reasonable and desirable :-)
No mention of the deceptive marketing name "Full Self Driving" in the article.
On top of that, FSD is still admittedly Level 2; Not exactly 'Full Self Driving'? And the controls can easily be tricked to think that the driver has their 'hands on the wheel' which is not enough to determine driver attentiveness while FSD is switched on.
> All new Tesla cars have the hardware needed in the future for full self-driving in almost all circumstances. [...] As these self-driving capabilities are introduced, your car will be continuously upgraded through over-the-air software updates.
https://www.tesla.com/en_CA/autopilot
I also personally would prefer they stuck to 'autopilot' and avoided the word full in 'full self-driving' and otherwise be more specific about what it means.
Other car companies typically productize the various features like lane assist, following cruise control, etc rather than bundle it into one. But that definitely makes communicating it more difficult.
Tesla probably doesn't want to call it 'limited self-driving' or 'partial self-driving'. Maybe 'computer assisted driving' but that doesn't sound as appealing. I can see the difficulty marketing here. But again not using 'full' as in it's complete and ready-to-go would help.
[0] https://news.ycombinator.com/item?id=27996773
TSLA is down almost 2% in pre-market trading at the time of this comment, though.
now running into unlit emergency vehicles? still think tgat's rather difficult sans inebriation or sleep dep.
Edit: I think the page count at the bottom of that list is off, it seems to repeat the last page so it might be less.
[1]https://www.respondersafety.com/news/struck-by-incidents/?da...
In Canada the Highway Act states that you must move over (change lanes) for stopped emergency vehicles. It seems to solve that problem gracefully, leaving an empty lane between the stopped vehicles and traffic.
I still think that Tesla's approach is the right one, I just think they need to gather more data before letting this product be used in the wild unsupervised.
Tesla perhaps isn't being loud enough about how autopilot isn't self-driving, and shouldn't even be relied upon to hit the brakes when something is in front of you.
Risk compensation is fascinating; driving with a bike helmet causes the biker and drivers around the biker to behave more dangerously.
Is society sophisticated enough to deal with advanced driver assistance? Is it possible to gather enough data to create self driving ML systems?
A nice and very sharp 8" stainless steel spike on the steering wheel facing the driver.
Didn't we have those in the 50s and 60s? Maybe not sharp, but collapsable steering columns are a significant improvement to survivability.
Do you have a truly reliable source for that? Because I hear this statement once in a while, and it feels flawed.
A helmet protects you from severe head injury if you are in an accident. There are more reasons for accidents than reckless car drivers. For example:
- Bad weather
- Driver not seeing the biker at all (no matter with or without helmet)
- Crash between 2 cyclists
https://www.bicycling.com/news/a25358099/drivers-give-helmet...
This result is both weakly supported and small, and it shouldn't be considered actionable.
Source please
https://www.sciencedirect.com/science/article/abs/pii/S00014...
https://psyarxiv.com/nxw2k
> There is a body of research on how driver behaviour might change in response to bicyclists’ appearance. In 2007, Walker published a study suggesting motorists drove closer on average when passing a bicyclist if the rider wore a helmet, potentially increasing the risk of a collision. Olivier and Walter re-analysed the same data in 2013 and claimed helmet wearing was not associated with close vehicle passing.
> We then present a new analysis of the original dataset, measuring directly the extent to which drivers changed their behaviour in response to helmet wearing. This analysis confirms that drivers did, overall, get closer when the rider wore a helmet.
there are some sources and studies linked. i.e. countries with the highest rate of helmet use also have the highest cyclist fatality rates
As Teslas get better at driving the drivers will be paying less attention inevitably, Tesla needs to start being responsible at some point
Once full self driving is statistically safer than humans how will you not let people use it? It is like saying you would rather have 10 children die because of bad driving skills rather than 1 child die because they were not paying attention at all times.
I'm fine with self-driving if/when it works (though I'm pretty sure from watching FSD Beta videos shot and edited by their biggest fans with a few interventions every 5 minutes, this is many many many years away for Tesla). But the company selling the self driving has to be responsible to some degree for the mistakes it makes.
https://www.tesla.com/videos/autopilot-self-driving-hardware...
"… HE IS NOT DOING ANYTHING. THE CAR IS DRIVING ITSELF."
Online since 2016, debunked as a lie. Still on Tesla’s website.
Maybe Elon can add a prompt in the car: one time payment to unlock 100+ mph
That’ll be a hard thing to overcome for the public. The drunk person “had it coming”, but did little Timmy?
It seems to me that Tesla door handles (in a world where they've been designing door latches for some time) are just plain ridiculous and likely unreliable but are a side effect of the market the company has been selling into. Gadgets go a long way with Tesla owners.
Obviously, things like a latch should not only work under all conditions including no-power, but they should probably be the same under all conditions. 'Emergency' latches aren't going to be used during an emergency as muscle memory is too important.
Your take seems a lot more plausible.
[1] https://news.ycombinator.com/item?id=26869962
To me driving requires paying constant attention to the road and being always ready to act swiftly: I just don't understand how you can have a "self driving car but you must but be ready to put your hands back on the steering wheel and your foot on the pedal(s)".
I have nothing against many "recent" safety features, like the steering wheel shaking a bit if the car detects you're getting out of your lane without having activated your blinker. Or the car beginning to brake if it detects an obstacle. Or the car giving you a warning if there's a risk when you change lane, etc.
But how can you react promptly if you're not ready? I just don't get this.
Unless it's a fully self-driving car, without even a steering wheel, a car should help you focus more, not less.
You cannot, that's the simple truth. You're supposed to focus on the road anyways and should be able to take over once any sort of autopilot or assist system starts working erroneously, yet in practice many people simply assume that those systems being there in the first place mean that you can simply stop focusing on the road altogether.
It feels like the claim of "fully self driving vehicle" is at odds with actual safety, or at least will remain so until the technology actually progresses far enough to be on average safer than human drivers, moral issues aside. Whether that will take 15, 50 or 500 years, i cannot say, however.
That said, currently such functionality could be good enough for the driver to take a sip from a drink, or fiddle around with a message on their phone, or even mess around on the navigation system or the radio - things that would get done regardless because people are irresponsible, but making which a little bit safer is feasible.
Maybe, as you say, it's feasible today or soon to better handle brief distractions but once you allow that it's probably dangerous to assume that people won't stretch out those distractions.
Which means people are either paying enough attention or these self driving systems are quite good. My suspicion is it’s a mix of both, where people tend to zone out in less hazardous driving conditions and start paying attention when things start looking dangerous. Unfortunately, that’s going to cause an equilibrium where people pay less attention as these systems get better.
Do we? Where does that come from? The data Tesla provides is hopelessly non-representative because it makes the assumption that the safety of any given road is independent of whether a driver chooses to switch on the system there.
Comma.ai makes the monitoring more strict when the system is less certain or when in denser traffic.
So I wonder if it's more about Telsa capitalizing on the hype of self driving cars (with the expensive self-driving add-on) in the short term and less about him misunderstanding the magnitude of difficulty.
Telsa is using the proceeds from that add-on to make them seem more profitable and fund the actual development. It's smart in some aspects, but very risky to consumers and Telsa.
I still wonder to what degree this was a collective delusion based on spectacular but narrow gains mostly related to supervised learning in machine vision, how much was fake it till you make it, and how much was pure investor/customer fleecing.
[1] https://waymo.com/safety
Waymo has lied about the capabilities of their technology regularly, and for that reason alone, should be assumed unsafe. A former employee expressed disappointment they weren't the first self-driving car company to kill someone, because that meant they were behind.
California only months ago opened up permits for paid robotaxi rides. So no, they couldn't have launched it in CA. If you've noticed, they actually are testing in SF with a permit.
> Waymo has lied about the capabilities of their technology regularly, and for that reason alone, should be assumed unsafe.
What lies? Their CA disengagement miles are for everyone to see, their safety report is open, they have had 0 fatalities in their years of operation. Seems like you just made this up.
Later, they were talking about how sophisticated their technology was: It can detect the hand signals of someone directing traffic in the middle of an intersection. Funny that a few months later, a journalist got an admission out of a Waymo engineer that the car wouldn't even stop at a stoplight unless the stoplight was explicitly mapped (with centimeter-level precision) so the car knew to look for it and where to look for the signal.
https://www.technologyreview.com/2014/08/28/171520/hidden-ob...
The article is seven years old at this point, but it's also incredibly humbling in how much bull- Waymo puts out, especially compared to the impressions their marketing team puts out. (Urmson's son presumably has a driver's license by now.)
In at least one scenario, the former Waymo engineer upset he had failed to kill anyone yet ("I’m pissed we didn’t have the first death"), caused a hit-and-run accident with a Waymo car, and didn't report it to authorities, amongst other serious accidents: https://www.salon.com/2018/10/16/googles-self-driving-cars-i... Said star Waymo engineer eventually went to prison for stealing trade secrets and then got pardoned by Donald Trump. Google didn't fire him for trying to kill people, they only really got upset with him because he took their tech to Uber.
I'd say Waymo has a storied history of dishonesty and coverups, behind a technology that's more or less a remote control car that only runs in a narrow group of carefully premapped streets.
How is a marketing video relevant from 2015 relevant to their safety record? They weren't even operating a public robotaxi service back then.
> My understanding is that in 2021, it still can't navigate parking lots (which would preclude using it for drive-thrus).
Completely false. Here is one navigating a Costco parking lot (can't get any busier than that) [1]. If you watch any videos in that YouTube channel, it picks you up and drops you off right from the parking lot. Yes, you can't use it for drive-thrus, but it doesn't qualify as "lying about capabilities".
> Later, they were talking about how sophisticated their technology was: It can detect the hand signals of someone directing traffic in the middle of an intersection. Funny that a few months later, a journalist got an admission out of a Waymo engineer that the car wouldn't even stop at a stoplight unless the stoplight was explicitly mapped (with centimeter-level precision) so the car knew to look for it and where to look for the signal.
Here is one recognizing a handheld stop sign from a police officer while it stopped for an emergency vehicle [2].
[1] https://www.youtube.com/watch?v=p5CXcJD3mcU
[2] https://www.youtube.com/watch?v=MpDbX1FViWk&t=75s
Well, yeah, that's the logic of an established business. Disruptive startups flout laws rather than following them.
I think you're far more likely to see L4 on limited access highways in good weather. A robotaxi service in a major city seems much more problematic given all the random behavior by other cars, pedestrians, cyclists, etc. and picking up/dropping off people in the fairly random ways that taxis/Ubers do. (And you'll rightly be shut down 6 months for an investigation the first time you run over someone even if they weren't crossing at a crosswalk.)
And for many people, including myself, automated highway driving would actually be a much bigger win than urban taxi rides which I rarely have a need for.
For example, couldn't emergency vehicles could send out a signal directly to autonomous vehicles or via a traffic managagemnt system to slow down or require the driver to take over when approaching. An elementary version of this is Waze which will notify you of road hazards or cars stopped on the side of the road.
Seeing as car crashes are the leading cause of deaths from people aged 1-54, it may be an improvement from the status quo
More than 38,000 people die every year in crashes on U.S. roadways. The U.S. traffic fatality rate is 12.4 deaths per 100,000 inhabitants. An additional 4.4 million are injured seriously enough to require medical attention. Road crashes are the leading cause of death in the U.S. for people aged 1-54.
This isnt true according to the CDC. Cancer and heart disease lead for the 44-54 group, and while "accidental injury" does lead from 1-44, if you break down the data, in many cases vehicle based accidents are not not the largest single source. For example:
Drowning is the largest single cause in 1-4
Cancer is the largest single cause in 5-9
Suicide is the largest single cause 10-14
https://wisqars-viz.cdc.gov:8006/lcd/home
It just makes the trips easier on the brain, and thus, for me, safer overall: its easier to see the overall situation when you've got free mental capacity
However, I also once so far have experienced what happens when this system experiences a poorly-marked construction zone. Whilst most construction sites on the interstate system place temporary road lines for lane shifts, this one solely used cones. While I was paying attention and never left the flow of traffic, the car actually fought a little bit against me following the cones into another lane, because it didn't see the cones, it was following the lines.
It doesn't surprise me at all that if someone gets too comfortable trusting the car to do the work, even if they think they're paying attention, they could get driven off the roadway.
How do automated systems deal with flaggers? Visibility of the stop/slow sign isn't sufficient to make a determination on whether it's safe to proceed (not to mention "stop" changes meaning here, entirely, from a typical stop sign). Often, whether or not you can proceed comes down to hand gestures from the flagger proper.
Not that I expect any reasonable driver to be using something like autopilot through such a situation, but we've also seen plenty of evidence that there are unreasonable drivers currently using these systems, as well.
Of course, the problem is, if we haven't developed it today, the ADAS systems of today won't understand it in ten years when there's enough saturation to be practical to use it. Apart from Tesla, very few car manufacturers are reckless enough to send OTA updates that can impact driving behavior.
Lane-following ADAS systems of today, mind you, can work relatively fine in construction areas... provided lane lines are moved, as opposed to relying solely on traffic cones.
A mesh network of vehicles on the road would add the ability for vehicles to become aware of far more than a human driver can ever know. For example, if cars become aware of a problem a few km/miles ahead, they can all adjust speed way before encountering the constriction in order to optimize for traffic flow (or safety, etc.).
Of course, this does not adequately deal with pedestrians, bikes, pets, fallen trees, debris on the road, etc.
Not saying cars would exclusively use the mesh network as the sole method for navigation, they have to be highly capable without it. The mesh network would be an enhancement layer. On highways this would allow for optimization that would bring forth some potentially nice benefits. For example, I can envision reducing emissions through traffic flow optimization.
Remember that electric cars still produce emissions, just not necessarily directly while driving. The energy has to come from somewhere and, unless we build a massive number of nuclear plants, that somewhere will likely include a significant percentage of coal and natural gas power plants.
The timeline for this utopia is likely in the 20+ year range. I say this because of the simple reality of car and truck ownership. People who are buying cars today are not going to dispose of them in ten years. A car that is new today will likely enter into the used market in 8 to 10 years and be around another 5 to 10. The situation is different with commercial vehicles. Commercial trucks tend to have longer service lives by either design or maintenance. So, yeah, 20 to 30 years seems reasonable.
Does it always get this correct, or does it sometimes read a 30mph sign on a side road and then slow the car on the motorway down to that speed?
Since I bought my car, Illinois (where I live) has raised the maximum limit on interstates by 10 MPH. My car doesn't know about it. If my car limited me to what it thought the limit was, I'd probably be driving 20 MPH slower than prevailing traffic, a decidedly unsafe situation.
It’s hard to imagine how speed limit systems would work without some sort of vision capabilities — a database of speed limits would never be up to date with roadworks and so on.
If you drive on a 60mph speed limit highway, no one is driving 60mph, everyone is going around 70mph. If you decide to use autopilot and it limits you to 60mph, you singlehandedly start disrupting the flow of traffic (that goes 70mph) and end up becoming an increased danger to yourself and others.
Not even mentioning cases when the speed limits change overnight or the map data is outdated or if a physical sign is unreadable.
For specific numbers (after subtracting reaction distance being the same for both):
55mph: car 165ft, 16-wheeler 225ft. 65mph: car 245ft, 16-wheeler 454ft.
As you can see, the gap between a car's stopping distance and a 16-wheeler's stopping distance increases with speed increasing, and non-linearly at that. Not even mentioning the destructive potential of a car vs. a 16-wheeler.
I would agree with your point if majority of the roads were occupied by 16-wheelers, but it isn't the case (at least in the metro area that I commute to work in).
Source for numbers used: https://trucksmart.udot.utah.gov/motorist-home/stopping-dist...
Note: I agree that it would be safer if everyone drove the exact speed limit, as opposed to everyone going 10mph above the speed limit. However, in a situation where everyone is driving 10mph above the speed limit, you are creating a more dangerous situation by driving 10mph slower instead of driving 10mph above like everyone else.
Anyway, perhaps I'm in a minority here, but I feel as though my driving has gotten _significantly safer_ since getting a Tesla, particularly on longer road trips.
Instead of burning energy making sure my car stays in the lane I can spend nearly all my time observing drivers around me and paying closer attention farther down the road. My preventative and defensive driving has gone up a level.
> I just don't understand how you can have a "self driving car but you must but be ready to put your hands back on the steering wheel and your foot on the pedal(s)".
I've not hit animals and dodged random things rolling/blowing into the road at a moment's notice. This isn't letting autopilot drive, it's like a hybrid act where it does the rote driving and I constantly take over to quickly pass a semi on a windy day, not pass it on a curve, or get over some lanes to avoid tire remnants in the road up ahead. I'm able to watch the traffic in front and behind and find pockets on the highway with nobody around me and no clumping bound to occur (<3 those).
To your suspicion, it is a different mode of driving. Recently I did a roadtrip (about half the height of the USA) in a non-Tesla, and I found myself way more exhausted and less alert towards the end of it. Could be I'm out of habit but egh.
Anyway, so far I've been super lucky. I don't think it's possible to avoid all car crashes no matter how well you drive. But I _for sure_ have avoided avoidable ones and taken myself out of situations where they later occurred thanks to the extra mental cycles afforded to me by auto-pilot. My safety record in the Tesla is currently perfect and I'll try and keep it that way.
I don't think autopilot is perfect either but I do think it's a good tool and I'm a better driver for it. Autopilot has definitely helped me spend better focus on driving.
I wouldn't want my, strangely enough upvoted a lot, comment, to be mistaken for Tesla hate. I like what they're doing. I just think the auto-pilot shouldn't give a false sense of security.
> I've not hit animals and dodged random things rolling/blowing into the road at a moment's notice.
> I don't think it's possible to avoid all car crashes no matter how well you drive.
Same here... And animals are my worst nightmare: there are videos on YouTube just terrifying.
For I do regularly watch crash videos to remind me of some of the dangers on the road.
You're talking about Autopilot which is just driver assistance technologies; lane keep assistance, adaptive cruise control, blind spot monitoring, etc. It's not to replace driver attention, it's just monitor sections of the road the the driver can't pay attention to full time. The driver is still remaining in control and attentive to the road.
The person you're responding to seems to be talking talking about the Full Self Driving feature who's initial marketing implied that the driver need not be mentally engaged at all or too impaired to drive normally. Which was later back pedal led to say that you need to pay attention.
Also, your subjective impressions may be what they are simply because you have not yet encountered the unlucky set of conditions which would radically change your view, as was surely the case for all the drivers involved in these sorts of incidents.
Obviously that's not exactly the same thing as taking over for a real car when the driver assistance features give up, but seems similarly challenging to take over the controls at the most precarious moment of travel, without being sort of "warmed up" as a driver.
This is a situation that you simply shouldn’t put yourself in. There is no reason to ever drive right next to a large vehicle, on either side, except for very short periods when overtaking them on a straight road.
Perhaps this is for the best.
[0] https://news.ycombinator.com/item?id=27996321
[1] https://news.ycombinator.com/item?id=27863941
[2] https://news.ycombinator.com/item?id=28053883
"The involved subject vehicles were all confirmed to have been engaged in either Autopilot or Traffic Aware Cruise Control during the approach to the crashes," NHTSA said in a document opening the investigation.
TACC is very different from Autopilot.
Safety Consideration When Using Adaptive Cruise Control
• The system can only brake so much. Your complete attention is always required while driving.
• Adaptive Cruise Control does not steer your vehicle. You must always be in control of vehicle steering.
• The system may not react to parked, stopped or slow-moving vehicles. You should always be ready to take action and apply the brakes.
[1]https://my.gmc.com/how-to-support/driving-performance/drivin...
Doesn't autopilot require you to put your hands on the wheel fairly regularly? Are these incidents just a matter of people using this feature outside of its intended use case?
Newer versions of Autopilot watch to make sure you keep your eyes on the road, probably to prevent this very scenario[1].
[1]: https://www.theverge.com/2021/5/27/22457430/tesla-in-car-cam...
Here’s more detail: https://www.tesla.com/support/autopilot
Oh wait NM that's tradional ICE cars.
FUD is dangerous
I think the best situation would be to have 'automated' stretches of highway specially designed to 'help' self driving systems.
Only self driving vehicles would be allowed on such special highways, and everything would be built around such systems.
This is kind of a position I've held for a long time but a different aspect of the problem. I think a system similar to IFF in aircraft would solve all of these issue. If every car knew where every other car was at all times, you could easily devise a system that would be nearly flawless. The issue is, there is no incremental path to this solution. You would essentially have to start over with the existing transportation network.
The question should be - how many lives were saved by this system vs how many would die if driven "normally"?
For example: it's not morally equivalent to die while drunk driving at 150 km/h vs dying as a pedestrian because someone ran over you.
I would prefer 10 drunk drivers die instead of just one innocent person.
It is also necessary to project this into the future, i.e. looking at the integral of expected lives lost 'rushing' self driving cars vs. 'waiting-and-seeing' (as Americans die at a rate of 40,000 per annum).
If twice as many people die for a fixed number of years to create a self driving system that results in half the fatality rate of the status quo, that becomes worth it very, very quickly.
I have no idea how self-driving fits into this. I don't have a feel how self-driving responds to emergencies. I'd have to experience an emergency in one. For that reason, I don't see myself ever trusting self-driving.
1. always pay full attention to where you are because there might be a truck or a family of 5 coming from the opposite direction, 2. never lift, 3. always look in the direction you want to travel in, not in the direction you currently travel
The rationale being that swerving most likely puts more people at risk more of the time. Especially true here where leaving the road often means either colliding with the granite cliff wall or ending up in the fjord or lake.
Does anyone know if the FSD Beta has this ability?
Recently, Tesla switched from radar-based to pure optical obstacle recognition. This should vastly improve this kind of behavior. Ironically that the investigation starts at a moment when they basically got rid of the old system.
Look on youtube for videos of the FSD beta. It is amazingly good at recognizing the surroundings of a car, including parked vehicles at the road side.
Neither can Volvo or VW.
Actually my 2015 Ap1.5 Model S does detect stopped vehicles, unfortunately not reliably.