this post was submitted on 24 Aug 2023
415 points (89.7% liked)
Technology
59204 readers
3157 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Without LIDAR, this is a fool's endeavor.
I wish this was talked about every single time the subject came up.
Responsible, technologically progressive companies have been developing excellent, safe, self-driving car technology for decades now.
Elon Musk is eviscerating the reputation of automated vehicles with his idiocy and arrogance. They don't all suck, but Tesla sure sucks.
Even with LIDAR there are just too many edge cases for me to ever trust a self driving car that uses current-day computing technology. Just a few situations I’ve been in that I think a FSD system would have trouble with:
I pulled up at a red light where a construction crew was working on the side of the road. They had a police detail with them. As I was was watching the red light the cop walked up to my passenger side and yelled “Go!” at me. Since I was looking at the light I didn’t see him trying to wave me through the intersection. How would a car know to drive through a red light if a cop was there telling you to?
I’ve seen cars drive the wrong way down a one way street because the far end was blocked due to construction and backtracking was the only way out. (Residents were told to drive out the wrong way) Would a self driving car just drive down to the construction site and wait for hours for them to finish?
I’ve seen more than one GPS want to route cars improperly. In some cases it thinks a practically impassible dirt track is a paved road. In other cases I’ve seen chains and concrete barriers block intersections that cities/towns have determined traffic shouldn’t be going through.
Temporary detour or road closure signs?
We are having record amounts of rain where I live and we’ve seen roads covered by significant flooding that makes them unsafe to drive on. Often there aren’t any warning signs or barricades for a day or so after the rain stops. Would an FSD car recognize a flooded out road and turn around, or drive into the water at full speed?
In my opinion, FSD isn’t attempting to solve any of those problems. Those will require human intervention for the foreseeable future.
Musk’s vision is (was?) to eventually turn Tesla’s into driverless robo-taxis. At one point he even said he could see regular Tesla owners letting their cars drive around like automated Ubers, making money for them, instead of sitting idle in garages.
Musk is an idiot
Well FSD is supposed to be Level 5 according to the marketing and description when it went on sale. Of course, we know Tesla's lawyers told California that they have nothing more than Level 2, have not timeline to begin building anything beyond Level 2, and that the entire house of cards hinges on courts and regulators continuing to turn a blind eye.
Or there are better other ways to tell a FSD car that the road is closed. We could use QR code or something like that which includes info about blockade, where you can drive around it, and how long it will stay blocked. A FSD should be connected enough to call home and give info to the servers, those then update the other FSD cars, et voila tadaa.
Sure. A QR code. That couldn't possibly get obscured.
Why should it get obscured? Just plug that giant QR code there. You can create QR codes, where you only need to see very little of the square to get all the info in the QR code. I don’t feel like obscuring would be any problem.
Wow, you solved one of the really easy self driving problems (sign recognizion) with a more complicated solution.
Sign recognition and traffic light recognition are available in a lot of cars. Detecting that the dark lump on the road is just a shadow and not slamming the brakes automatically is the hard part.
Or that the white sky is in fact not a white sky but the sideways view of a semi trailer.
(The latter issue is why relying on multiple sensors camera+radar+ultrasound in the case of my car's emergency brake system and drive assistant is always a lot better - each sensor on its own has its failure modes)
Umm, we had it about blocked streets not normal sign detection But if those would be standardized to designs good "readable" for cam/pc (like QR codes), the car would never recognize e.g. a "S" for a "5". Due to high contrast, one could assure that it still works by night. But honestly I would prefer, if manually cars are completely banned from roads and only FSD cars are allowed. The streets would then be made for FSD cars (instead for human("monkeys" drivers) and about all problems are solved.
In huge parts of the world these are standardized and readable as well - where Vienna Convention signs are being used. (Those with have more pictograms).
Detours are also labelled by standard signs in a pre-determined font with a specific reflectivity. Easy for a car to recognize.
And even without text recognition is one of the really easy parts of self driving. I've done development on document recognition for random bank statements, builinng plans and legal documents - all the paperwork for financing a house - before. Those documents come in in various fonts, layouts differently formatted etc.
Mud for one. Trees and bushes for another. Strong wind for a third. All of those things already obscure signs or make them very hard for humans to read, let alone a computer.
Just like that cheaper non-lidar Roomba with room mapping technology, it will get lost.
I don't know why people are so quick to defend the need of LIDAR when it's clear the challenges in self driving are not with data acquisition.
Sure, there are a few corner cases that it would perform better than visual cameras, but a new array of sensors won't solve self driving. Similarly, the lack of LIDAR does not forbid self driving, otherwise we wouldn't be able to drive either.
challenges in self driving are not with data acquisition.
What?!?! Of course it is.
We can already run all this shit through a simulator and it works great, but that's because the computer knows the exact position, orientation, velocity of every object in a scene.
In the real world, the underlying problem is the computer doesn't know what's around it, and what those things around doing or going to do.
It's 100% a data acquisition problem.
Source? I do autonomous vehicle control for a living. In environments much more complicated than a paved road with accepted set rules.
You're confusing data acquisition with interpretation. A LIDAR won't label the data for your AD system and won't add much to an existing array of visible spectrum cameras.
You say the underlying problem is that the computer doesn't know what's around it. But its surroundings are reliably captured by functional sensors. Therefore it's not a matter of acquisition, but processing of the data.
won’t add much to an existing array of visible spectrum cameras.
You do realize LIDAR is just a camera, but has an accurate distance per pixel right?
It absolutely adds everything.
But its surroundings are reliably captured by functional sensors
No it's not. That's the point. LIDAR is the functional sensor required.
You can not rely on stereoscopic camera's.
The resolution of distance is not there.
It's not there for humans.
It's not there for the simple reason of physics.
Unless you spread those camera's out to a width that's impractical, and even then it STILL wouldn't be as accurate as LIDAR.
You are more then welcome to try it yourself.
You can be even as stupid as Elon and dump money and rep into thinking that it's easier or cheaper without LIDAR.
It doesn't work, and it'll never work as good as a LIDAR system.
Stereoscopic Camera's will always be more expensive than LIDAR from a computational standpoint.
AI will do a hell of a lot better recognizing things via a LIDAR Camera than a Stereoscopic Camera.
This assumes depth information is required for self driving, I think this is where we disagree. Tesla is able to reconstruct its surroundings from visual data only. In biology, most animals don't have explicit depth information and are still able to navigate in their environments. Requiring LIDAR is a crutch.
I disagree with you, I don't think visual camera's alone are up to the task. There was an instance of a Tesla in auto pilot mode driving at night with the driver being drunk. This took place in Texas on the high way, the car's camera footage was released and it showed the autopilot not identify the police car in the lane with it's red/blue lights flashing as a stationary obstacle. Instead it didn't realize there was a car in the way around 1 second before the 55 mph impact, and it turned of autopilot that 1 second before.
Having multiple layers of sensors, some being good at actually sensing a stationary obstacle, plus accurate range finding, plus visual analysis to pick out people and animal, thats the way to go.
Visual range only cameras were just reported to have a harder time recognizing people of color and children.
If the obstacle was visible in the footage, the incident could have been avoided with visible spectrum cameras alone. Once again, a problem with the data processing, not acquisition.
If we're talking about the safety of the driver and people around them, why not both types of sensors? LIDAR has things it excels at, and visual spectrum cameras have things they do well too. That way the data processing side has more things to rely on, instead of all the eggs in one basket.
Cost seems to be a pretty good reason. Admittedly, until I looked it up 5 minutes ago I thought it was just 100-200% more expensive than cameras, but it seems to be much more than that.
On top of that, there are the problems of weather and high energy usage. This is more of a problem than just "not working on rain": if the autonomous driving system is designed to rely on data from a sensor that stops working when it rains, this can be worse than not having that sensor in the first place. This is what I refer to by saying that LIDAR is a crutch.
That's a pretty good point, the part about if it's raining or snowing, LIDAR can't be used, which could leave the system in a much worst spot. It's getting to the point where I'm beginning to think that fully self driving cars just won't be 100% possible in all conditions in all locations.
For instance, where I live, we can have some bad winters, snow, ice, slippy conditions. People have a tough time with these conditions, and I'd imagine it'd be even harder for a self driving car, especially given how the sensor suites work. My car has that intelligent cruise control where it'll slow down when it senses a car ahead of me, then match it's speed. That feature stops working if too much snow accumulates on the sensors.
Optical cameras alone have issues as well that can’t be handled though. It’s the combination of the two along with other things like ultrasonic sensors that makes them safe. More sensors in general are better because they reduce the computational burden and provide redundancy - even if that redundancy is to safely stop.
Cost is certainly an issue, but on $40k+ vehicles it’s cheap enough for other EV makes to include it in the cost. Volvo for instance is using Luminars version at a cost of about $500 (https://www.wired.com/story/sleeker-lidar-moves-volvo-closer-selling-self-driving-car/).
Image processing is expensive even with dedicated hardware and LiDAR provides enough extra information to avoid needing to make make certain calculations off of images alone (like deltas between image series to calculate distance). Those calculations are further amplified by conditions where images alone don’t provide enough information - similar to how there are conditions where the LiDAR data alone wouldn’t be sufficient.
and you're suggesting to use LIDAR which is more expensive and power hungry as a replacement for those computations?
I meant the computations are expensive, i.e. slow to perform even with good processors. When you need to do something millions of times, anything to make that faster helps with the overall safety of the system.
Yes, self driving is not computationally solved at all. But the reason people defend LIDAR is that visible light cameras are very bad at depth estimation. Even with paralax, a lot of software has a very hard time accurately calculating distance and motion.