Waymo CEO Explains Why Camera-Only Self-Driving Falls Short (electrek.co) 215
Longtime Slashdot reader AmiMoJo shares a report from Electrek: Waymo co-CEO Dmitri Dolgov laid out the clearest technical case yet for why cameras alone can't take a self-driving system to full autonomy, arguing that "weak sensing" hits a safety ceiling long before it reaches superhuman performance. [...] Dolgov made the comments in a talk at Y Combinator's Startup School, walking through the lessons Waymo has learned building its driver over close to two decades. He put the sensor question on the table plainly: "there's been a long-standing debate about what kind of sensors do you actually need for autonomous driving."
His answer draws the line that camera-only advocates tend to skip right past. "Humans of course can drive with just eyes, so there's that proof of existence," he said. "If the goal were to just approximately match human performance or to build an assist product, that's a very reasonable way to go." Then the catch. If you're targeting full autonomy and strongly superhuman performance, he said, "you find that weak sensing just leads to a safety curve that flattens out way too early." Dolgov said that cameras, lidar, and radar are complementary rather than redundant: "These different sensing modalities, they're not backups to each other," and combining them produces a view "vastly superior to what you get with any one sensor." He said multiple sensor types also protect against physical failures, such as a leaf or branch blocking a camera, while helping Waymo climb the "exponential ladder of nines" required for fully driverless safety.
Dolgov warned that camera-only systems may improve quickly before plateauing "way before the performance that is required by your product." He also pushed back on the cost argument against lidar, calling it "a number that has a fairly short shelf life" as hardware costs continue to fall.
His answer draws the line that camera-only advocates tend to skip right past. "Humans of course can drive with just eyes, so there's that proof of existence," he said. "If the goal were to just approximately match human performance or to build an assist product, that's a very reasonable way to go." Then the catch. If you're targeting full autonomy and strongly superhuman performance, he said, "you find that weak sensing just leads to a safety curve that flattens out way too early." Dolgov said that cameras, lidar, and radar are complementary rather than redundant: "These different sensing modalities, they're not backups to each other," and combining them produces a view "vastly superior to what you get with any one sensor." He said multiple sensor types also protect against physical failures, such as a leaf or branch blocking a camera, while helping Waymo climb the "exponential ladder of nines" required for fully driverless safety.
Dolgov warned that camera-only systems may improve quickly before plateauing "way before the performance that is required by your product." He also pushed back on the cost argument against lidar, calling it "a number that has a fairly short shelf life" as hardware costs continue to fall.
Gosh Elon... (Score:2)
Lidar falls short too (Score:4, Interesting)
Snow, rain, fog, changes in roads, construction, flooding. There are hundreds of novel situations Waymo fails at. It is easy to trash your competitors but far harder to deliver true driverless cars.
And before the Waymo gooners show up to defend them, lets just get this out of the way: Waymo has millions of short trips, but all of their miles combined are just rounding error on the number of trips that happen in a single day in the United States. The plain fact is that they are largely unproven from a statistical perspective, especially on highway and novel situation driving
Re:Lidar falls short too (Score:5, Insightful)
Snow, rain, fog, changes in roads, construction, flooding. There are hundreds of novel situations Waymo fails at. It is easy to trash your competitors but far harder to deliver true driverless cars.
And before the Waymo gooners show up to defend them, lets just get this out of the way: Waymo has millions of short trips, but all of their miles combined are just rounding error on the number of trips that happen in a single day in the United States. The plain fact is that they are largely unproven from a statistical perspective, especially on highway and novel situation driving
You're concerned about if it's actually capable of delivering meatsacks safely to a destination.
You should be FAR more concerned with the finger-pointing corruption that will ensure anyone killed by a murderous autonomous car, will NEVER be able to face the actual entity responsible. Because 37 legal entities exist between the gas pedal and the steering wheel.
By Fuck-You design.
Re: (Score:2, Insightful)
Snow, rain, fog, changes in roads, construction, flooding. There are hundreds of novel situations Waymo fails at.
You're addressing a claim that wasn't made just so you can ignore the claim that was made.
Re: (Score:2)
And it isn't just passenger safety - those things are a menace during civil emergencies. There needs to be both a lot of work put in to contextually proper response to exceptional circumstances - from what should be easy stuff like "don't block fire trucks or drive over fire hoses [sfstandard.com]" on through figuring out how to safely get out of the way when power goes out [sfstandard.com]. For reference, right now we're at the point where they still confuse themselves [ktvu.com].
Of course if you
Took them a long time to figure that out (Score:2)
The DARPA demonstrations made too many people too sure that multiple cameras and AI would make self-drivers cars simple to pull off. Then they followed up with tests where it's warm and dry and the weather doesn't get weird. They wanted to get investment in before things started getting interesting. What they've got now is a technology that has slowed down and needs to learn more about what to do when there's a road crew digging holes in front of them. It's going to be much more of an adventure then the opt
Re: (Score:2)
Except Tesla FSD works better than humans already. I haven't had to do a critical intervention ever. There are hundreds of thousands of users daily .. no fatal accidents.
Re: Took them a long time to figure that out (Score:2)
Re: (Score:2)
So? Lidar plus camera fails a lot less than camera alone. I have no idea why you event try to do FUD here.
Re: (Score:2)
Snow, rain, fog, changes in roads, construction, flooding. There are hundreds of novel situations Waymo fails at. It is easy to trash your competitors but far harder to deliver true driverless cars.
Holy poster missing the point batman! Capabilities of a company's current product have zero to do with discussion of the capabilities of the underlying platform. What Waymo currently does or doesn't do it irrelevant. The fact of the matter is LiDAR + Camera is better than Camera alone. We're talking about a regulatory control system not a USA presidential election. The quality and quantity of information matters. And it is very much possible to trash your competitors for choosing a fundamentally less capabl
Re: (Score:2)
Those are not limitations of lidar. For example, snow and heavy rain are just extra noise, and we are pretty good at filtering noise.
Where they have issues is not knowing how to deal with or recognize certain obstacles, but even there lidar and radar have proven how valuable they are. As TFA notes, lidar can see solid objects with certainty, and radar can measure their speed. Tesla has a well known problem where it will drive straight into stationary objects at full speed (disengaging a few milliseconds bef
Re: (Score:2)
Two problems with your statements
First problem: Waymos have been driving on highways for a long time.
Second, is you are looking at the data totally wrong. To be better than human drivers, all any *individual* Waymo driver has to have is more miles than *the average individual driver* - and that number is already passed by tens-of-thousands.
The "number of trips that happen in a single day" is not relevant whatsoever as a safety metric.
Re: Lidar falls short too (Score:2)
The same problem as AI (Score:4, Interesting)
AI is proving to be a useful augment to skilled humans, but not a replacement. We were promised a robot and instead have invented a robotic arm. I think LiDAR based driving is headed in the same direction.
The systems they have invented are a great augment to human judgment, but not a replacement for it. I would love to buy a car equipped with Waymoâs lane and collision detection, but I would certainly not trust a Waymo in most cases and especially in serious weather.
Re:The same problem as AI (Score:4)
but I would certainly not trust a Waymo in most cases and especially in serious weather.
Give me a system that stops driving in bad conditions over a human judgement that thinks the increased risk is acceptable any day. If getting to your destination despite everything is your concern, Waymo's system isn't for you. If your own safety is your concern, you'd take Waymo's system even in serious weather. Your risk of dying in bad weather increases by around 35% depending on which study you look at. Waymo's record is currently 0 deaths.
Re: The same problem as AI (Score:2)
Re: (Score:2)
And that's the thing --- if it's just augmenting a human driver, it means that now, instead of driving, you're supervising. And that's a much more difficult thing to devote continuous, full, unwavering attention to (still required) than it is to devote that attention to something you're actually doing yourself.
I have yet to feel safer in a car that argues with me than I do in a car that just lets me drive. And until it can take over *fully* from me (which it may never), I would rather it not express opini
Humans do more then see (Score:4, Interesting)
Re: (Score:3)
Humans have already lost. Both Waymo and Tesla FSD have better "accident per mile" numbers than humans. By a lot, it's not margin of error it's not driving locales. They have mimicked human reasoning for nearly all relevant situations aside from highly contrived scenarios you may dream up but never see in real life. It even understands construction worker gestures and signs.
https://www.youtube.com/shorts... [youtube.com]
Re: (Score:2)
First, humans use other senses like hearing.
Hearing is a sense that is rarely used while driving. It's also the most commonly obscured sense (with cars putting serious effort into obscuring external noise, and drivers almost universally playing music or conversations while on the road).
There are places where this doesn't apply. Drive in India some time and you'll find there hearing plays an important role, but it's also a place where a horn is used more often than an indicator precisely because people there use hearing when driving. That doesn't appl
Re: Humans do more then see (Score:2)
Lidar is nice, not required (Score:3)
I think the Rivian approach is the best. They're basically copying the Tesla model of having their fleet send the sensor recordings back and then using them as training data. While they're far behind right now, they'll catch up quickly, as they can copy a lot of the basic technology that's now established instead of having to invent everything from scratch and make wrong decisions along the way. The difference is Rivian is including lidar, so they just include the matched video and lidar recordings as a larger training data set. I'm convinced that if the price of lidar was more like it is today back when Tesla started selling FSD in 2016, then that's exactly what Tesla would have done, but Tesla is now stuck with their solution with ten million cars already on the road.
But in clear weather conditions, I don't think lidar really helps that much. Most human accidents would be avoided if drivers had better situational awareness and didn't have gaps in paying attention, and the cameras solve that.
Re: Lidar is nice, not required (Score:2)
I worked on self driving software (Score:2)
...years ago
I strongly agreed that lidar and radar would be necessary in addition to cameras. The projects I worked on used them.
My concerns with cameras were dirty lenses and limited dynamic range.
It's interesting that Tesla seems to be making a camera only approach work well.
I continue to follow developments, but will admit that I was wrong is the evidence shows it.
Laughable (Score:3)
You can look it up on google. It's amazing how different their tune gets when they are under oath to Congress...
This isn't the future this is just more outsourcing and you're going to be on the road with remote control cars pilot it from another country so that they can pay taxi cab drivers even less than they pay them now. And over time it will become increasingly unsafe because they will start using layers of contractors and they won't follow driving laws the way the current batch does because they won't be sitting on limitless amounts of investor cash hoping to reduce wages even more than what Uber pays which is already sub minimum wage
In the dystopian cyberpunk books they never really get this deep into all the nasty little pieces of dystopia that affect daily life it's usually the big cool stuff of course. I'm not enjoying learning about the banal aspects of a techno-feudal corporatocracy
Re: (Score:2)
Again he already admitted to Congress that when they need the vehicles to do anything more complicated than basic Lane assist driving it's a call center in the Philippines doing the actual driving.
And he repeats his lie again. No they didn't say this to congress. They specifically mentioned what the Philippine's centre does and doesn't do, but you are paid to ignore that.
You must be paid.
You've been corrected on this post so many times that even the dumbest amoeba at this point would have seen the error they made. The only remaining conclusion is that someone is paying you to spread bullshit. Despicable human you are.
Re: Laughable (Score:2)
Re: (Score:2)
Are you talking about Waymo? That's not true at all. Their cars do drive themselves most of the time. They have the ability to call a human when they get into a situation they don't know how to handle, but even then the human doesn't do the driving, it only tells the computer if its plan is acceptable or if it should do something else.
In particular, all the safety stuff is done on the car.
Multiple approaches (Score:2)
I think debating which is better is moot.
Just build your car your way, Tesla will build theirs their way. And the results will show in the end. Clearly Tesla's option will be cheaper that could turn out to be an advantage, and Waymo's can "see" in more wavelengths which in theory could be an advantage, but the proof is in the putting. Maybe sensing isn't the defining factor in safety, maybe it is.
We'll see. Good that several companies have several approaches.
Re: Multiple approaches (Score:2)
Performance Bonuses and Equity, anyone? (Score:2)
And trust him because he knows (Score:2)
His guided cars have hit more things and screwed up traffic more than any other. What we need is a set of national Car2Car standards so these megalomaniac douche bags CEOs don't get to decide what is good enough. If done right, the only thing getting hit is the Darwin winners who J-walk.
He must be new here... (Score:2)
...project ARGO [unipr.it] in 1998 was about a self driving car equipped with just two camera and a 200 MHz MMX PC. Too bad that the project was left unfunded after its remarkable achievments.
I'm a human and I want radar (Score:2)
I have a very old car, but I recently drove a new rental. Radar has two huge benefits for me. First, I'm bad at checking my blind spots. On a self driving car it could fill in for camera glitches. Second, radar can detect when the car I'm following slows down faster than I can see that the car's getting closer. Especially if it's doppler radar. A self driving system deserves the same info.
people don't want perfection (Score:2)
What causes the majority of accidents? (Score:2)
This is cope... (Score:3)
Re:Meaningless statements? (Score:5, Insightful)
He's saying "do you want robot cars to be as good as humans, or safer?"
If the goal is just "as good as a human" then (in theory) we have "proof of concept" that that can be achieved (humans drive cars all the time) ...
Double standard for bots (Score:5, Interesting)
I believe the goal has to be "better than humans" because humans seem to have a double standard for bots per screwups. Bots F-ing up spread like wildfire on social media.
Maybe we are collectively bigoted against automation, but being bots probably won't lobby to correct that or fight for their (alleged) rights, that's probably the way it will stay (at least until true sentience). Thus, I agree with the premise the driving system must be better than a human driver to be accepted.
Re: Double standard for bots (Score:2)
Re: (Score:2)
I can't calculate the square root on 58643 that I should accept if my computer can't do it?
I don't know, maybe you should work on your mental math. I can calculate the square root of 58 squared without blinking an eye.
Re: (Score:3)
I can't calculate the square root on 58643 that I should accept if my computer can't do it?
I don't know, maybe you should work on your mental math. I can calculate the square root of 58 squared without blinking an eye.
Reminds me of Jim from Taxi!
Elaine: "Rio de Janeiro to here! Do you know how far that is?"
Jim: "6,842 miles."
Elaine: "Well look, do you know what kind of mail system they must have?"
Jim: "The Brazilian mail system is government run with an annual budget of 27 million, 485,000 pesetas."
Elaine: "How do you know these things?"
Jim: "You mean I'm right?"
Jim: "I wasn't even sure if pesetas was the right money."
Re:Double standard for bots (Score:5, Insightful)
I mean, humans have a double standard for mistakes they make while they're in control, and mistakes someone else makes while in control of their lives. Which is pretty reasonable.
And we also have a double standard for mistakes that a human might make (which we might anticipate and be able to prepare ourselves for), versus mistakes a computer vision system might make (which are more likely to take us completely by surprise when they're in a category that a human would never make). Driving in an automated car is knowing more than ever that at any minute your car may kill you, and you have no power to stop it or even know when to be anxious. And that *is* a worse experience than being a passenger in a car with a human driver, and certainly worse than driving yourself.
Re: (Score:2)
I believe the goal has to be "better than humans" because humans seem to have a double standard for bots per screwups.
It's not so much we have double standards for bots, we have double standards for humans driving cars specifically.
Re:Double standard for bots (Score:5, Insightful)
Not so much a double standard, it's liability. If a human crashes their car, they are responsible. If a robotaxi crashes, the operator is responsible.
The operator will have a fleet of vehicles, so as well as the potential for a large number of very large claims being made against them, if the problem is systemic it might get their whole fleet "grounded" until they can put it right.
Re:Meaningless statements? (Score:5, Informative)
If you want "safer than human", as Google evidently does, you need LIDAR.
No you don't. Humans get distracted, drunk, and drowsy. Automation doesn't. That's where automation is safer than humans.
Lidar and radar can add safety, but they do not guarantee it.
Re: (Score:2)
You mistake what "safety" means. It just means accident rated below an acceptable threshold. It does not mean perfection at all.
Re: (Score:2)
Where did I say anything about perfection? It's true that automation has a perfect record of not getting drunk, but I doubt that's what you meant.
Don't make up things to blame me for. There's plenty of real things for that.
Re: Meaningless statements? (Score:2)
Re: (Score:3)
That depends on what statistical measure you use. It's statistically significant enough that there's a special law against it, so it's not trivial.
Re: (Score:3)
It's true that automation has a perfect record of not getting drunk
Not so sure about that, my Tesla went through a phase were it would randomly drift over and hug the centerline, then eventually cross over it, wobble around a bit, and lurch suddenly toward the edge -- with the on-screen visualization showing exactly what it was doing. Seems to have been silently fixed a few software updates ago, right around the same time it stopped spamming errors about the left headlight E-Fuse and front radar in the service log. Whatever it was on, I don't want any, but could probably
Re: (Score:2)
If you want "safer than human", as Google evidently does, you need LIDAR.
No you don't. Humans get distracted, drunk, and drowsy. Automation doesn't. That's where automation is safer than humans.
Lidar and radar can add safety, but they do not guarantee it.
I think you settled with "safer". Are you happy with 30000 deaths per year from the intended and correct use of a product? That would make that product far safer than the current human use, but is that enough for you? Do you know any other product (that isn't protected by the second amendment) with that kind of a headcount which hasn't been banned from sale?
Re: (Score:2)
Are you happy with 30000 deaths per year from the intended and correct use of a product?
I kind of feel that SDCs are the wrong technology to make a paltry 30% improvement, and also might not succeed that well[*]. The problem is the Great North American Stroad, a result of poorly designed road building rules and poorly designed zoning rules. It mixes high peed and local traffic in the worst possible way making a unique combination of poor safety, bad traffic, low efficiency and astoundingly high cost.
[*] wel
Re: (Score:3)
You left out the limits of vision, as in fog, snow, and deer standing on the side of the road at dusk.
There is where something other than conventional cameras would be helpful.
Re:Meaningless statements? (Score:5, Interesting)
How about we get as good as humans first, and then we can talk about being better than humans. Sorry, even Waymo is not there yet. The claim is that safety performance is due to a lack of being able to see and more sensors is the solution. If we take that statement at face value (ie lacking data), then additional cameras could just as easily overcome the limitation as additional sensing modalities. But most accidents caused by humans are due to judgement errors, not an inability to see, so one could just as easily argue that super-human performance should be achievable without additional sensors. Incidentally, all approaches to self-driving (including Tesla’s) use way more sensors than humans possess, so if more sensors is all that is needed, we should have achieved super-human performance a long time ago.
Look, let’s be straight. When Waymo started their experiment in self-driving more than 20 years ago, cameras were not as good as they are today, embedded computers were not as powerful, there was a limit to how much custom hardware design they could do, power consumption was an issue, and computer vision was in its infancy. The only feasible approach at that time was to use a rich sensor array coupled with large data heuristics—maps that tell the car which lane to be in when going through an intersection, which light corresponds to the lane, which intersections have stop signs or yield signs, which intersections have no turn on red, what is the speed limit, which intersections should be avoided entirely because they are too complex, etc—to lower the burden of perception and decision-making that had to be made directly by the car in real time. In other words, the current Waymo driver and all of the engineering that was put into it is a sunk cost that they are not going to back out of now. If they were starting fresh today, I would be willing to bet that they would not take the same approach.
Re:Meaningless statements? (Score:4, Interesting)
I agree that he was making empirical claims as though they were first-principle claims, and ignoring the fact that cameras are better optical sensors than the human eye, that reaction speeds can be faster than humans can manage, etc. There's no reason a camera system alone can't deliver better-than-human performance. Whether it can deliver a performance sufficiently better than humans to be good enough, depends on what bar is set for sufficiency.
All that said, I don't see an inherent reason to stick with a single sensor modality beyond the complexity of integrating multiple streams of different types of information, and that doesn't feel like a good enough reason to me. Most humans themselves don't rely on a single information stream when driving -- we use our ears as well as our eyes, which I think is an important clue; drones routinely use fused multi-sensor information including many other types of sensor information eg thermal, GPS, barometer, magnetometer; etc. And clearly hardware costs for these other sensors have plummeted and will fall much further too.
Re:Meaningless statements? (Score:4, Insightful)
A camera only system can only be capable of superhuman performance if object recognition is at least as good as humans are capable of. Currently, computer vision systems are nowhere near as good as humans at that and probably won't be for a very long time. Besides, for all "better optical sensors than the human eye" inferring is still performed with low resolution images, much lower resolution than a human eye is capable of, because it is computationally very expensive. This is why all those recaptcha images that are used to train computer vision systems are so small.
Hence his claim is spot on.
Re: (Score:2)
You're overclaiming.
Object recognition is an important capability, but what matters always is system performance, not any single capability, whether object recognition or anything else (eg number of sensors, low light performance, fatigue impacts, etc etc).
internal network resolution may be lower than the sensor resolution, but that’s not the same as saying AVs only use “low resolution images”. They often crop regions of interest at high resolution and use multi-scale processing.
Captcha im
Re: (Score:2)
A camera only system can only be capable of superhuman performance if object recognition is at least as good as humans are capable of.
If performance is limited by object detection, then yes, but if performance is limited by decision-making and contextual awareness, then not necessarily.
Re: (Score:3)
I agree that he was making empirical claims as though they were first-principle claims, and ignoring the fact that cameras are better optical sensors than the human eye,
Not necessarily. Pretty sure the older HW3 sensors from Tesla had considerably less instantaneous dynamic range than the human eye — probably not enough to be a problem most of the time, but still...
But more than that, cameras are notably worse than eyes, because they're in the wrong spot. A pair of eyes three feet back from the window sees the scene and focuses through whatever is hitting the glass, so unless it's a deluge, the contents of the glass are largely out of focus. With small sensors mou
Re: (Score:2)
All that said, I don't see an inherent reason to stick with a single sensor modality beyond the complexity of integrating multiple streams of different types of information
I don’t disagree, but I will go back to the point I was trying to make previously. Is the limitation one due to perception or decision-making? If the problem is you can’t see enough of the right things, that supports the argument that you need more sensors. But if the problem is that you aren’t making the right decisions with the information you have, more information is not necessarily better.
Re: (Score:2)
Sheesh, how to over-interpret. Notice I didn't say that camera-only was good enough. I said it depends where the bar is set. I think this is just a technical engineering question, capable of being solved in multiple different ways. And I don't think Tesla is close to having an unequivocal lead in this field; many Chinese players are doing extremely well, some with camera-only eg Xpeng, some with mixed-sensor systems eg Li Auto, Nio, Xiaomi, BYD...
Re: (Score:2)
How about we get as good as humans first, and then we can talk about being better than humans. Sorry, even Waymo is not there yet.
Two things:
a) You can't strive for the best when you intentionally limit yourself from the onset. You don't perfect cameras and then add LiDAR to the mix, the change in engineering platform becomes a huge setback and undoes a whole world of R&D that came before it. You start from the onset with an idea hardware platform and work to improve with that.
b) You're wrong, Waymo objectively has far better safety record than humans, even when you dice statistics down to make sure you're comparing the same kind
Re: Meaningless statements? (Score:3)
Re: (Score:2)
The person you are responding to makes the unsubstantiated claim that Google are only retaining lidar due to it being a sunk cost, but it seems much more likely that Tesla is on
Re: (Score:2)
The person you are responding to makes the unsubstantiated claim that Google are only retaining lidar due to it being a sunk cost,
That is not what I said. I said Waymo’s approach to self-driving is a sunk cost. The difference between the Waymo and Tesla approaches is not just sensor count. Waymo chose their approach a long time ago and are locked in to it. Tesla is also locked in to their approach, but they started later and framed the problem differently. It’s more complex than just lidar vs no lidar.
Re: (Score:3)
a) You can't strive for the best when you intentionally limit yourself from the onset.
Perhaps, but again, if perception (ie ability to see) is the primary factor limiting safety performance, there is more than one way to overcome that.
b) You're wrong, Waymo objectively has far better safety record than humans,
Objectively, the Waymo driver operates in a much smaller service area than most humans drive, with the majority of miles accumulated on a small set of curated routes. If “safety is solved”, as you claim, why doesn’t Waymo operate everywhere that humans do?
It doesn't matter how good cameras are. A primary method of perceiving depth is always better than an inferred method.
You are bringing up engineering trade-offs. LiDAR brings some advantages, yes, but it als
Re: (Score:2)
Re: (Score:3)
And even that's wrong, because people also use their ears when driving. And even a bit of "sensing vibration". So people don't JUST use their eyes when driving, even though that's the primary mode.
You don't need superhuman! (Score:3)
I don't buy that argument.
You don't need superhuman, if you get p99 human level you're already better off.
Max Verstappen level of car control with cameras would be a significant upgrade over the average driver.
Re: Meaningless statements? (Score:4, Interesting)
All of those words have meanings and make sense as he used them. Your lack of comprehension isn't his fault.
What's crazy is imagining that a computer can definitely do what a brain does when we don't even know how the brain does it. MAYBE it can, but we have no proof of that.
Re: (Score:3)
It doesn't need to do everything a brain can do. Autonomous driving doesn't need to get a boner when it sees noods for example. Maybe you're thinking of Turing completeness which is another story.
Re: (Score:2)
This is true, indeed as you allude there is much that a brain does that you don't want your software to do. But you may not be able to separate all of those things away from the things you do want without going badly awry, either. Also, some of those things may be artifacts of the attached peripherals.
Meaningless statements (Score:2)
All of those words have meanings and make sense as he used them. Your lack of comprehension isn't his fault.
What's crazy is imagining that a computer can definitely do what a brain does when we don't even know how the brain does it. MAYBE it can, but we have no proof of that.
What he (Dolgov) is saying is a belief that stems from rationalization. It *seems* like it's a proof because that's how he's phrasing it, but in the final analysis he's a biased observer expressing an opinion and bolstering it with a rationalization.
It's what we get every time we ask an expert for advice about something they don't directly know. Ask a physicist about dark energy, or a doctor about your unusual symptom, or a nutritionist about what you eat, or an economist about whether the Fed should lower
Re: (Score:2)
Your post reminds me of a good Heinlein quote: "man is not a rational animal; he is a rationalizing animal".
I agree with it. The only thing I'm not sure of, is whether FSD is sufficiently trained yet to be safe on roads around the whole world (not that we expect humans to reach this standard, either, but it matters for an OEM).
I did see an impressive demo of a Chinese car doing FSD in a busy Chinese city. It was an interesting comparison of Tesla vs Xpeng.
https://www.youtube.com/watch?... [youtube.com]
Re: Meaningless statements (Score:2)
Re: (Score:2)
He isn't even an expert. He's a hype man, a salesman.
Frankly, he's engaging in a zero sum fallacy. Because it's not perfect, it can't be done. But it quite clearly has been done, successfully, with a drastic reduction in fatalities and injury. Is it perfect? No. But it's better than anything Waymo has been able to achieve with fancy intrusive useless sensors (there's nothing quite as infuriating as getting in a Waymo and having it just sit there confused about the homeless guy standing on the curb), but it'
Re: (Score:2)
The most important difference is that Waymo depends on having an extremely detailed 3D map of the area in which it is driving, whereas Tesla tries to YOLO it wherever it goes.
Re: (Score:2)
There are tons of things computers can do that humans could do first. But what we basically always find is that there is a way to do it without insight, often by brute-forcing things and mixing that with some algorithms (that have no clue what they are doing).
Hence basically most cases where computers replicate human skill, we are pretty sure they do it differently.
Re: (Score:3)
My favorite way to pick on Tesla owners is to tell them that AutoPilot, Full Self Driving, and Full Self-Driving Supervised to not allow for ANY truly hands / attention free operation because you need to be able to take over driving in an instant. You are ALWAYS liable if the car gets into an accident.
On the flip side Mercedes Drive Pilot, under specific circumstances, actually all
Re: (Score:2, Insightful)
He's lying.
"Better than humans" is possible with cameras only, because most of the errors humans make have nothing to with their eyes. Humans get distracted, drunk, and tired. Automation doesn't.
I'm sure lidar and radar can do better than cameras alone. But lidar and radar do not guarantee better than humans.
He's just trying to justify his expensive system.
Re: (Score:3)
Humans get distracted, drunk, and tired. Automation doesn't.
It doesn't get tired. It might get into an altered state. It can definitely get distracted, and get confused about what something even is so long that it's a problem. It doesn't have infinite resources.
Using other types of sensors takes guesswork out of certain decisions. If you can measure density and depth like you can with a mix of sensors which don't even include cameras then you can reliably do AEB which overrides your beloved optical-only system and keeps it from screwing the pooch.
Re: (Score:3)
You didn't read what I wrote. The claim is that camera-only automated cars can never be safer than human cars. That's wrong, because the safety comes from automation itself being a better driver than a human, even if it only has cameras. Other sensors may or may not add safety. The do not guarantee additional safety.
An automated car can also be worse than a human driver, even with a dozen sensors. But it won't be because of the sensors.
Re: (Score:2)
The claim is that camera-only automated cars can never be safer than human cars
No it isn't. That's literally not what "you find that weak sensing just leads to a safety curve that flattens out way too early" means.
Re: (Score:2)
More importantly, cameras can be better because they can have real binocular vision, no blind spots, and plenty of redundancy. No, current systems don't have that, and I am happy my Tesla still has/uses ultrasonic sensors for better distance estimation close-in. Ultimately by the point cameras are completely useless, so are humans. Get off the road and wait for better conditions!
Re: Meaningless statements? (Score:2)
Re: (Score:2)
Observable evidence from about 40 years of research into this say he is exactly right and that the additional sensors make a huge difference. Your entirely invalid claim is that computer vision can be as good as human vision. No idea where you take that from, but it cannot be actual research, because actual reality is that nobody knows.
Re: (Score:2)
No. That is not what he said. He said cameras-only driverless cars cannot be better than humans.
That is a lie, and that's what I called him out for. Research be damned; cameras already have better resolution than the human eye, but that's not where the safety comes from. It comes from being automated, and not getting distracted, drunk, or drowsy. If you won't believe that without research papers, then you're a damned fool.
Re: Meaningless statements? (Score:2)
Re: Meaningless statements? (Score:2)
A human can still excel at predicting what's coming based on incomplete information.
AIs in cars always needs full confirmation before a decision is made.
It's possible to make an AI that can perform some predictions, but it can cause a lot of trouble too.
Re: (Score:2)
That isn't half of it.
How many people in the audience own a Tesla?
He's talking to people who, if they are not personally intimate with fully self-driving capabilities in a Tesla, are at least aware of the capability.
It's like Steve Jobs going to Xerox to give a talk about why mice are superior input devices for navigation over just a keyboard.
Re: (Score:2)
This is obvious stuff and its demonstrable. Mark Rober did a series of tests on YouTube proving this exact point pitting a Tesla against a Lida
Re: (Score:2)
Automation is safer because it doesn't get distracted, drunk, or drowsy, not because it uses lidar. Adding lidar can add safety, but it doesn't guarantee it.
Re: Meaningless statements? (Score:2)
Re: (Score:2)
Humans also drive using feedback from their ass (but really accelerometers in their ears and also the rest of kinesthetic feedback from the whole of their body).
This is obvious, yes?
So do the self driving cars. The computer is constantly evaluating input from many sensors simultaneously
Re: (Score:2)
Re: (Score:2)
Thanks for you input, Elon.
Re: (Score:2)
Re: (Score:2)
Re: Trust me, I'm a CEO (Score:2)