7 ms·
Tesla (and Comma.ai ?) seem to be the only groups that find Lidar unnecessary for L5 self-driving. There is another group who find Lidar unnecessary for drivin
by onion2k 6y ago
Tesla (and Comma.ai ?) seem to be the only groups that find Lidar unnecessary for L5 self-driving.
There is another group who find Lidar unnecessary for driving: humans.
- Zhenya 6y agoautonomous driving needs to be safer than humans to be adopted (some say way safer). So this argument of humans don't need it, is sort of, well yes. We're trying to do better than humans.
- dilap 6y agoHumans drive great when they're not distracted, with just two 2d cameras. But we get distracted a lot. So an advanced enough AI could easily be better at humans than driving w/o LIDAR, simply by driving as well as humans when not distracted, and not getting distracted.
- kaba0 6y agoAn advanced enough AI, and we are really far from it. Current AIs are really good at isolated problem solving, but driving is not in that domain.
- YZF 6y agoAgreed. Once we're past "the singularity" a computer with human hardware can drive as good as humans. But our eyes are not cameras and the singularity hasn't happened last time I checked. Btw, we also need some additional hardware to do a good job, headlights, defogger, wiper blades...
- sixQuarks 6y agowow, getting downvoted for stating this fact. I can't believe the anti-Musk bias here on HN. You would think HN would actually appreciate one of the more innovative entrepreneurs on the planet.
- xirbeosbwo1234 6y agoIf Musk is an innovator, then I guess Ian M. Banks was literally John Galt.
- jfoster 6y agoI think it might be due to a form of jealousy. Sad to see HN so unenthusiastic about a guy who is probably the most amazing entrepreneur of our time. It's not that he hasn't invited a little bit of controversy (eg. comments about the diver, covid, etc.) but I think those are inconsequential when it comes to HN's attitude toward him.
- alextheparrot 6y agoWhat are you all talking about, the original top-level post has a reasonably presented assessment of Tesla’s strategy and the ramifications of such a strategy. Sure, that assessment might be flawed, but I don’t see how any of that is “anti-Musk bias” or “jealousy”. The comment regarding humans not needing LiDAR misses the point, which is that either every other self-driving car group is incorrectly deploying an expensive hardware solution or there are valid technical reasons why that might be a better path to L5. Tesla trying to overcome the limitations other companies are deploying LiDAR for is a technical risk they are taking on, and, as the top level post noted, might not be the worst bet given the downside which is that they’ll just have to license the tech.
- jfoster 6y ago> every other self-driving car group is incorrectly deploying an expensive hardware solution Given that Tesla are the only company to have put something approaching autonomy into the hands of customers, that does seem to be the case. What happens next in this thread? Someone points out that Tesla's solution is still far from perfect and claims that they're being highly irresponsible by having shipped it. Then I'd ask how many accidents have happened as a result of FSD beta having shipped to some customers already and they assert that Tesla isn't transparent with the data so we don't know. I'd point out that if we don't know of a single Tesla FSD accident so far, it can't be that irresponsible to have put it in the hands of a few customers and that it's been rapidly improving in the weeks since. It's the same thread, repeated over and over on here, Reddit, etc. Tesla was the company that made electric vehicles successful. To your point about groups taking different approaches, I think Musk & Tesla do take approaches that will differ from others, and succeed partly because of their approach being better in some important way. I think they had a bit of an unfair advantage when it came to autonomy. They needed hardware that could be shipped on vehicles as soon as possible, which meant that it needed to be cost efficient & energy efficient. LIDAR's energy efficiency doesn't get talked about much, and I don't mean just the LIDAR itself but also the subsequent processing power required to make sense of the resulting point cloud. Each Waymo vehicle is an expensive tech demo. It's a very good tech demo, but over the years it's become evident that they're very hesitant to scale it up right now. I'm sure they'll eventually get there, but they are not a company focused on getting this done ASAP. Tesla are getting autonomy done with urgency.
- YZF 6y agoI own a model 3. Often it says "the camera is blinded" (e.g. when the sun is shining at it), often there's rain on the lens... So I'd argue Tesla's cameras are much worse for driving under all conditions vs. the human eye. I'm not sure there's enough redundancy for depth perception either. I tend to agree with the argument that humans prove it's possible to drive with just visual sensors but Tesla's sensors, for better or worse, are not equivalent. Then there's the part about the brain being what gets the sensor data and interprets it. I doubt Tesla's computer has all the capabilities of a human brain. Driving behind a boat trailer, it's clear to me it's a boat, but Tesla's computer is confused. Bicycles can be a little iffy as well. In general the way the car perceives the world seems subpar to how a human does, by far. I'm a fan of the car, I'm mostly a fan of Tesla/Elon, but I don't think these cars with this technology are ever gonna be truly autonomous so they can drive with no human intervention... IMHO.
- Erlich_Bachman 6y ago> Often it says "the camera is blinded" It could very well theoretically be a software issuse.
- javagram 6y agoThe human driver has a number of options to adjust their “cameras” (eyes) that the Tesla doesn’t. E.g., flipping down the sun shield, adjusting head position to reduce glare, or even putting on a pair of polarized sunglasses. While there are technical solutions possible with more hardware, it’s not clear that Tesla could correct for some of these things in just software...
- Erlich_Bachman 6y agoIf these are uncommon enough, a Tesla could just stop on the side of the road and alert the driver. There will be a driver inside anyway. And there is a huge difference between "no L5, you have to keep an eye of the road every 10 seconds all the time", and "drives itself, but once every 2 hours, or in extreme weather conditions safely warns you and makes you take over and force you to drive yourself". This could very well be a great usecase, even if it won't drive itself for those 1% of cases.
- ashtonkem 6y agoSure. But that’s a moot point if it turns out you need to basically replicate large parts of the human brain in order to drive with cameras alone.
- pas 6y agoBut those are needed for L5 anyway, Lidar or noLidar. The big issue is the world model, what's the prior when uncertainty rises, how to spot really out of ordinary stuff, and how to spot really ordinary-looking but dangerous stuff.
- sudosysgen 6y agoI'm fairly certain L5 automation won't need anywhere near the processing capacity of the visual cortex.
- pas 6y agoThe visual cortex part is what current deep learning stuff handles pretty well. What's not so obvious is how to generate a model that can manage the "usual problems" (unclear road signs, debris on road, pedestrians/cyclists, harm minimization in accidents) without "freaking out" (phantom braking). For example if the car in front did not even slow down while going through some obstacle (the proverbial plastic bag), then it's very likely that we don't have to either. But if something looks like a drunk guy trying to cross the highway, then it's caution time. That said, without the actual data/studies/access on what Tesla does and how, it's hard to say how good they are at this. (Sure, in theory everything needed to drive a car well can be inferred from regular human vision visuals. But that means their model has to be able to do counterfactual reasoning - I see this thing as a wall, but other cars go through it, so it's not a wall, so maybe it's a marking. Otherwise it'll very hard to just figure out things frame-by-frame and simply by raw object detection.)
- sudosysgen 6y agoThe visual cortex does a lot more. It can predict future states, it does things like attention, it can classify states that are abnormal and direct attention to it, it can make analogies between different objects and use those to decide what is normal and what isn't. It's a lot more than just object recognition - it can also generate object classes on-the-fly and use them in the future. We are very, very far from that. And having such a developed visual processing system is a prerequisite for feeding an executive processing system that can do counterfactual reasoning.
- jayd16 6y agoYep. It's clear that lidar isn't literally needed. That said, I think the argument just shifts to "the cameras on a Tesla aren't good enough either." That's also just a shot from the hip, weasely sort of argument but it's harder to dismiss outright.
- jfoster 6y agoIt's fairly easy to dismiss that argument; the cameras would not be difficult to replace if they are the limiting factor. Cameras are cheap.
- sudosysgen 6y agoThe human eye has capabilites that you'd need a camera more expensive than a LiDAR by a good bit too replicate. No rolling shutter, very low delay, continuous signal processing instead of per-frame, maximum resolution around ~70MP eq. in the center, f/2.8 aperture at a full-frame image size, servo driven active rangefinding, continuous cleaning, etc... A camera with those features would be around 3000-15000$, and you'd need two. Also, the human brain can use focus information and stereo phase to deduce 3D structure, which is another ace up its sleeve. It might very well be the case that reproducing this is more expensive in training, processing and material than LiDAR. In fact, I'd say it's very likely - we're talking loops that need to update hundreds of times a second and data rates in the gigabits, which our brains deal with by doing a lot of processing inside the eye and along the way.
- jfoster 6y agoYou don't need human eye equivalents. Human eyes are probably not the ideal cameras for driving.
- sudosysgen 6y agoThis may be the case, you might not have to reproduce everything. But just to give an example that is very relevant to self-driving, getting a camera with similar low-light video performance as the human eye costs over 2000$.
- apengwin 6y agoYes, but in fact, imagine how much better at driving we'd be if we had LIDAR lasers shooting out of our heads.