6 ms·
As someone who worked in this space, you are absolutely right, but also kind of wrong - at least in my opinion. The cold hard truth is that LIDARs are a crutch
by chippiewill 1y ago
As someone who worked in this space, you are absolutely right, but also kind of wrong - at least in my opinion.
The cold hard truth is that LIDARs are a crutch, they're not strictly necessary. We know this because humans can drive without a LIDAR, however they are a super useful crutch. They give you super high positional accuracy (something that's not always easy to estimate in a vision-only system). Radars are also a super useful crutch because they give really good radial velocity. (Little anecdote, when we finally got the Radars working properly at work it made a massive difference to the ability for our car to follow other cars, ACC, in a comfortable way).
Yes machine learning vision systems hallucinate, but so do humans. The trick for Tesla would be to get it good enough to where it hallucinates less than humans do (they're nowhere near yet - human's don't hallucinate very often).
It's also worth adding that last I checked the state of the art for object detection is early fusion where you chuck the LIDAR and Radar point clouds into a neural net with the camera input so it's not like you'd necessarily have the classical methods guardrails with the Lidar anyway.
Anyway, I don't think Tesla were wrong to not use LIDAR - they had good reasons to not go down that route. They were excessively expensive and the old style spinning LIDARs were not robust. You could not have sold them on a production car in 2018. Vision systems were improving a lot back then so the idea you could have a FSD on vision alone was plausible.
- goalieca 1y ago> The cold hard truth is that LIDARs are a crutch, they're not strictly necessary. We know this because humans can drive without a LIDAR, however they are a super useful crutch. Crutch for what? AI does not have human intelligence yet and let’s stop pretending it does. There is no shame in that as the word crutch implies.
- lazide 1y agoI think they meant crutch for the AI so they could pretend for investors that AGI is right around the corner haha
- spot5010 1y agoI've never understood the argument against lidars (except cost, but even that you can argue can come down). If a sensor provides additional data, why not use it? Sure, humans can drive withot lidars, but why limit the AI to using human-like sensors? Why even call it a crutch? IMO It's an advantage over human sensors.
- IgorPartola 1y agoI don’t work in this field so take the grain of salt first. Quality of additional data matters. How often does a particular sensor give you false positives and false negatives? What do you do when sensor A contradicts sensor B? “3.6 roentgen, not great, not terrible.”
- giveita 1y agoYou can say that about human hearing and balance. What if they conflict with visual? We are good at figuring it out.
- ben_w 1y agoWe throw up, an evolved response because that conflict is a symptom of poisonous plants messing with us.
- IgorPartola 1y agoHumans can be confused in a number of ways. So can AI. The difference is that we know pretty well how humans get confused. AI gets confused in novel and interesting ways.
- giveita 1y agoDoes removing a sense help in that regard (for car driving?). Probably comes down to lidar (and Ai) failure modes.
- IgorPartola 1y ago
- jfim 1y agoLIDARs have the advantage that they allow detecting solid objects that have not been detected by a vision-only system. For example, some time ago, a Tesla crashed into an overturned truck, likely because it didn't detect it as an obstacle. A system that's only based on cameras is only as good as its ability to recognize all road hazards, with no fall back if that fails. With LIDAR, the vehicle might not know what's the solid object in front of the vehicle using cameras, but it knows that it's there and should avoid running into it.
- sandworm101 1y agoSolid objects that arent too dark or too shiny. Lidar is very bad at detecing mirrored surfaces or non-reflecting structures that absorb the paticular frequency in use. The back ends of trucks hauling liquid are paticularly bad. Block out the bumper/wheels, say by a slight hill, and that polished cone is invisible to lidar.
- bayindirh 1y agoAdd one or a couple of RADAR(s), too. European cars use this one weird trick to enable tons of features without harming people or cars.
- UltraSane 1y agoLIDAR works be measuring the time it takes for light to return so I don't understand how a object can be too reflective. Objects that absorb the specific wavelength the LIDAR uses is an obvious problem.
- sandworm101 1y agoToo reflective, like a flat mirror, will send the light off in a random direction rather than back as the detector. Worse yet, things like double reflections can result in timing errors as some of the signal follows a longer path. You want a target that is nicely reflective but not so shiny that you get any double reflections. The ideal is a matte surface painted the same color as the laser.
- lazide 1y agoThe big promise of autonomous self-driving was that it would be done safer than humans. The assumption was that with similar sensors (or practically worse - digital cameras score worse than eyeballs in many concrete metrics), ‘AI’ could be dramatically better than humans. At least with Tesla’s experience (and with some fudging based on things like actual fatal accident data) it isn’t clear that is actually what is possible. In fact, the systems seem to be prone to similar types of issues that human drivers are in many situations - and are incredibly, repeatedly, dumb in some situations many humans aren’t. Waymo has gone full LiDAR/RADAR/Visual, and has had a much better track record. But their systems cost so much (or at least used to), that it isn’t clear the ‘replace every driver’ vision would ever make sense. And that is before the downward pressure on the labor market started to happen post-COVID, which hurts the economics even more. The current niche of Taxis kinda makes sense - centrally maintained and capitalized Taxis with outsourced labor has been a viable model for a long time, it lets them control/restrict the operating environment (important to avoid those bad edge cases!), and lets them continue to gather more and more data to identify and address the statistical outliers. They are still targeting areas with good climates and relatively sane driving environments because even with all their models and sensors, heavy snow/rain, icy roads, etc. are still a real problem.
- tialaramex 1y agoThis whole "But Waymo can't work in bad climates" thing is very dubious. At some point it is too dangerous to be driving an automobile. "But Waymo should also be dangerous" is the wrong lesson. When the argument was Phoenix is too pleasant I could buy that. Most places aren't Phoenix. But SF and LA are both much more like a reasonable place other humans live. It rains, but not always, it's misty, but not always. Snow I do accept as a thing, lots of places humans live have some snow, these cities don't really have snow. However for ice when I watch one of those "ha, most drivers can make this turn in the ice" videos I'm not thinking "I bet Waymo wouldn't be able to do this" I'm thinking "That's a terrible idea, nobody should be attempting it". There's a big difference between "Can it drive on a road with some laying snow?" and "Can it drive on ice?".
- lazide 1y ago
- raincole 1y ago> The cold hard truth is that LIDARs are a crutch The hard truth is there is no reason to limit machines to only the tools humans are biologically born with. Cars always have crutches that humans don't possess. For example, wheels.
- profunctor 1y agoThe reason is cost, LIDAR is expensive.
- uoaei 1y agoThat's ok, they're supposed to be. That's no excuse to rush a bad job.
- revnode 1y agoThe point of engineering is to make something that’s economically viable, not to slap together something that works. Making something that works is easy, making something that works and can be sold at scale is hard.
- waldarbeiter 1y agoIf it would be easy there would already be a car costing a few million that few can afford but that has solved AD. But there isn't.
- revnode 1y agoThere is no market for such a thing. At that price point, you get a personal chauffeur. That’s what rich people do and he can do stuff that a self driving system never can.
- tialaramex 1y agoAnd the rich people who don't want a chauffeur like driving the car. They will buy a $10M car no problem, but they want driving that car to be fun because that's what they were paying for. They don't want you to make the driving more automatic and less interesting.
- fluidcruft 1y agoMusk's argument "Humans don't have LIDAR, therefore LIDAR is useless" has always seemed pretty dumb to me. It ignores the possibility that LIDAR might be superhuman with superhuman performance. And we also know you can get superhuman performance on certain tasks with insect-scale brains. Musk's just spewing stoner marketing crap that stoners think is deep, not actual engineering savvy. (and that's not even addressing that human vision is fundamentally a weird sensory mess full of strange evolutionary baggage that doesn't even make sense except for genetic legacy)
- mixedbit 1y agoMusk's argument also ignores intelligence of humans. The worst case upper bound for reaching human level driving performance without LIDAR is for AI to reach human level intelligence. Perhaps it is not required, but until we see self-driving Teslas performing as well as humans, we won't know this. Worst case scenario is that Tesla unsupervised self-driving is as far away as AGI.
- hudon 1y ago> they're not strictly necessary. We know this because humans can drive without a LIDAR and propellers on a plane are not strictly necessary because birds can fly without them? The history of machines show that while nature can sometimes inspire the _what_ of the machine, it is a very bad source of inspiration for the _how_.
- ethbr1 1y agoTurns out intelligent design is quicker than evolutionary algorithms. ;)
- DonHopkins 1y agoI'd rather cars have crutches than the people they run over.
- inciampati 1y agoI wish I had radar eyes
- UltraSane 1y agoI want to see gamma rays, I want to hear X-rays, and I want to smell dark matter.
- RaftPeople 1y ago"I've seen things you people wouldn't believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhäuser Gate".
- davidhs 1y ago> Yes machine learning vision systems hallucinate, but so do humans. When was the last time you had full attention on the road and a reflection of light made you super confused and suddenly drive crazy? When was the last time you experienced objects behaving erratically around you, jumping in and out of place, and perhaps morphing?
- hodgesrm 1y agoWell there is strong anecdotal evidence of exactly this happening. We were somewhere around Barstow on the edge of the desert when the drugs began to take hold. I remember saying something like, “I feel a bit lightheaded; maybe you should drive . . .”And suddenly there was a terrible roar all around us and the sky was full of what looked like huge bats, all swooping and screeching and diving around the car, which was going about 100 miles an hour with the top down to Las Vegas. And a voice was screaming: “Holy Jesus! What are these goddamn animals?” [0] [0] Thompson, Hunter S., „Fear and Loathing in Las Vegas“
- ben_w 1y agoMany accidents are caused by low-angle light dazzle. It's part if why high beams aren't meant to be used off a dual carriageway. When was the last time you saw a paper bag blown across the street and mistook it for a cat or a fox? (Did you even notice your mistake, or do you still think it was an animal?) Do you naturally drive faster on wide streets, slower on narrow streets, because the distance to the side of the road changes your subconcious feeling of how fast you're going? Do you even know, or are you limited to your memories rather than a dashcam whose footage can be reviewed later? etc. Now don't get me wrong, AI today is, I think, worse than humans at safe driving; but I'm not sure how much of that is that AI is more hallucinate-y than us vs. how much of it is that human vision system failures are a thing we compensate for (or even actively make use of) in the design of our roads, and the AI just makes different mistakes.
- uoaei 1y agoThis impulse to limit robots to the capacities, and especially the form factors, of humans has severely limited our path to progress and a more convenient life. Robots are supposed to make up for our limitations by doing things we can't do, not do the things we can already do, but differently. The latter only serves to replace humans, not augment them.
- lukeschlather 1y agoI had taken for granted that the cameras in the Tesla might be equivalent to human vision, but now I'm realizing that's probably laughable. I'm reading it's 8 cameras at 30fps and it sounds like the car's bus can only process about 36fps (so a total of 36fps, not 8x30 = 240fps theoretically available from the cameras, if they had a better memory bus.) It also seems plausible you would need at least 10,000 FPS to fully match human vision (especially taking into account that humans turn their heads which in a CV situation could be analogous to the CV algorithm having 32x30 = 960 FPS, but typically only processing 140 frames this second from cameras pointing in a specific direction. So maybe LIDAR isn't necessary but also if Tesla were actually investing in cameras with a memory bus that could approximate the speed of human vision I doubt it would be cheaper than LIDAR to get the same result.
- tialaramex 1y agoMostly human vision is just violently different from a camera, but you could interpret that as a mix of better and worse. One of the ways it's better is that humans can sense individual photons. Not 100% reliably, but pretty well, which is why humans can see faint stars on a dark night without any special tools even though the star is thousands of light years away. On the other hand, our resolution for most of our field of vision is pretty bad - this is compensated for by changing what we're looking it when we care about details we can just look directly at it and the resolution is better right in the centre of the picture.
- asats 1y agoAlso the human vision is backed by the general intelligence, which those cameras are very much not.
- marcos100 1y agoI want my self-driving car to be a better driver than any human. Sure we can drive without LIDAR, but just look up the amount of accidents caused by humans.
- paulryanrogers 1y agoHumans cause one fatal accident per million miles. (They have no backup driver they can disengage to.) Now just look up how many disengagements per million miles Tesla has.
- Eisenstein 1y agoCan you make your point without the stat, or provide the stat for us please?
- phinnaeus 1y agoHumans have the most sophisticated processing unit in the known universe to handle the data from the eyes. Is the brain a crutch?
- bayindirh 1y agoAt least for one marine creature, which I forgot its name, the answer is yes. Said creature dissolves its brain the moment it can find a place to attach and call home.
- chronogamous 1y agoCan't think of the name atm either, but I'm pretty sure it only does so, as it would be pointless to make any further decisions after attaching itself - it simply has no means to act on anything after that... the attaching is the only thing it 'does' in it's life... after that, it's only job, and only ability, is to be. Chose the wrong spot to attach and call home? Brains wouldn't make a bit of difference (unless regretting it's one life-choice is somehow usefull during this stage of just being, being stuck on the spot).
- shagie 1y agoSea squirt. One of the simplest members of Chordata. https://en.wikipedia.org/wiki/Ascidiacea https://en.wikipedia.org/wiki/Ascidiacea
- ModernMech 1y ago> Vision systems were improving a lot back then so the idea you could have a FSD on vision alone was plausible. This was only plausible to people who had no experience in robotics, autonomy, and vision systems. Everyone knew LIDAR was the enabling technology thanks to the 2007 DARPA Urban challenge. But the ignoramus Elon Musk decided he knew better and spent the last decade+ trashing the robotics industry. He set us back as far as safety protocols in research and development, caused the first death due to robotic cars, deployed them on public roads without the consent of the public by hoisting around his massive wealth, lied consistently for a DECADE about the capabilities of these machines, defrauded customers and shareholders while becoming richer and richer, all to finally admit defeat while he still maintains the growth story of for Tesla's future remains in robotics. The nerve of this fucking guy.
- DennisP 1y agoYou might not have the classical guardrails, but you are providing the neural net with a lot more information. Even humans are starting to find it useful to get inputs from other sensor types in their cars. I agree that Tesla may have made the right hardware decision when they started with this. It was probably a bad idea to lock themselves into that path by over-promising.
- maxerickson 1y agoYou could write a rant like this about 4 vs 3 wheels.
- ethbr1 1y ago> Anyway, I don't think Tesla were wrong to not use LIDAR - they had good reasons to not go down that route. They were excessively expensive and the old style spinning LIDARs were not robust. You could not have sold them on a production car in 2018. The correct move for Tesla would have been to split the difference and add LIDAR to some subset of their fleet, ideally targeted in the most difficult to debug environments. Somewhat like Google/Waymo are doing with their Jaguars. Don't LIDAR 100% of Teslas, but add it to >0%.
- ACCount37 1y agoTesla did, in fact, use "ground truth vehicles" - vehicles that were owned and operated by Tesla itself, and had high performance LIDARs installed. They were used to collect the data to train the "vision-only" system and verify its performance. Reportedly, they no longer use this widely - but they still have some LIDAR-equipped "scout vehicles" they send into certain environments to collect extra data.
- ethbr1 1y agoIt seems like an own goal not to sell these to some interested and targeted customers then.
- ACCount37 1y agoWho would buy those and why? They don't use LIDARs for better self-driving somehow. They're just data harvesting units with wheels. And I don't think there's a large and underserved market for LIDARs on wheels.
- ethbr1 1y ago> Who would buy those and why? [...] They're just data harvesting units with wheels. Tesla would subsidize them and offer them at the same price as non-LIDAR models, to select customers in target areas. And yes, you answered the second part of your own question.