from the bigger-they-are-the-sooner-they-stop dept.
A nameless submitter writes:
Motor Trend is reporting that a Tesla FSD owner/beta tester wondered why his car was stopping well before the end of the off ramp...where there is a stop sign before the surface street. It was so far ahead of the sign that he had to quickly switch to manual before getting rear ended by the next car down the off ramp.
His theory is that these stop signs are larger than the normal ones--and a tape measure confirms it. So the Tesla (by whatever mechanism) seems to be using the apparent size of the stop sign to determine where to stop.
This is why (imo) any sort of self driving needs ranging information. There are options, stereo cameras (which may be hard to keep aligned?), radar for certain types of objects (but stop signs might not be a good target??) or the original choice, lidar.
Also reported here with the following quote:
As they suspected, the stop sign on the exit ramp is, in fact, significantly larger. Many of these signs measure just over 48 inches tall, which is around 60 percent larger than the 30-inch signs in their neighborhood.
"60 percent larger to the camera gets interpreted as 60 percent closer?" asks the Tesla owner on Twitter. "I still have the slow creep problem here [...] and it'll come to a complete stop while sitting in the near lanes even if no cars are around!"
(Score: 2) by krishnoid on Friday September 02 2022, @03:09AM
I mean, it's not like radar, lidar, et al doesn't exist for a good reason [youtu.be].
(Score: 2) by bryan on Friday September 02 2022, @06:58AM
Or maybe be consistent with traffic signs? Or don't freak out if your local edge case isn't handled perfectly by beta software?
(Score: 4, Insightful) by RamiK on Friday September 02 2022, @08:35AM (7 children)
The onboard computer has more than enough data to make the right decision here just like human drivers do.
The case for lidar shouldn't be a shortcut around bad AI. It should be for faster and better than-human decision making. Like, faster highway speeds and tighter gaps. Or centimeter gaps when standing in traffic / parking...
Use lidar as a crutch and you'll just end up crippling the AI modelling around the edge cases where lidar fails but cameras / human eyes don't.
compiling...
(Score: 3, Touché) by JoeMerchant on Friday September 02 2022, @10:45AM (3 children)
>faster highway speeds and tighter gaps
Yes, eventually. Let's start out with big "Student AI Driver" signs on them and let them prove themselves capable of better than human driver safety while following current human driver guidelines for safe following distance and speeds.
I would rather be a passenger in a safely auto-driven vehicle moving at 65mph than a driver on the same route doing 80mph. Sure, the AI could eventually drive 100mph more safely than human drivers, but probably not until human drivers are retrained to coexist with AI drivers on the roads, or humans are simply banned from driving.
How much simpler will CIA containment by termination ops become when everyone uses AI FSD for all of their ground transportation?
Україна досі не є частиною Росії Слава Україні🌻 https://news.stanford.edu/2023/02/17/will-russia-ukraine-war-end
(Score: 2) by Freeman on Friday September 02 2022, @01:31PM
I've seen that movie. They were using vehicles on rails, though. (Minority Report)
Joshua 1:9 "Be strong and of a good courage; be not afraid, neither be thou dismayed: for the Lord thy God is with thee"
(Score: 1, Insightful) by Anonymous Coward on Friday September 02 2022, @02:34PM (1 child)
I don't know about where you live, but in my neck of the woods, if the AI follows these guidelines it will never be able to merge onto or operate safely on the highways. There are no gaps to merge into that actually meet human safe following distances, at least not during a normal weekday.
(Score: 3, Interesting) by JoeMerchant on Friday September 02 2022, @03:27PM
Yeah, I try to just stay off the roads on "normal weekday" rush hour times... it's nutso here too.
But, what's the point of having an AI driver if you can't let it do the crappy duty for you? So, AI follows the guidelines (what is it, like 3 seconds behind at 65mph?) when possible, gets abused by people cutting in front constantly, but still gets to destination with maybe 2-3% longer total travel time. And, when it needs to merge from an on ramp, if there are no "safe gaps" it does what all the human drivers do: puts itself in an "acceptable gap" before the merge lane runs out. One benefit of the AI signage on the car would be: predictability. Any assholes trying to block the AI from merging would know: A) everything they do is being recorded in high def and livestreamed to multiple archives as potential evidence, B) the AI will merge itself into the opening in a predictable fashion - which may well be encoded into the rules of the road: "when a vehicle is on a limited access merge lane, following at a distance of less than 3x the merging vehicle's length makes the tailgating car liable for a hefty fine and points if the merging vehicle is forced to stop in the merge lane." Sort of like "let the bus back in" laws.
Україна досі не є частиною Росії Слава Україні🌻 https://news.stanford.edu/2023/02/17/will-russia-ukraine-war-end
(Score: 2, Interesting) by Anonymous Coward on Friday September 02 2022, @11:36AM (2 children)
And it's not even edge cases where Lidar fails, here's a recent attack on Lidar-camera fusion systems, https://arxiv.org/pdf/2106.07098.pdf [arxiv.org]
"Security Analysis of Camera-LiDAR Fusion Against Black-Box Attacks on Autonomous Vehicles" From the Abstract:
Beware the frustum attack!!
(Score: 0) by Anonymous Coward on Friday September 02 2022, @12:25PM (1 child)
I can handle most attacks fairly well, but that one has always frustrumated me.
(Score: 2, Funny) by Anonymous Coward on Friday September 02 2022, @12:29PM
“Beware the frustum attack, my son!
The jaws that bite, the claws that catch!
(Score: 4, Interesting) by JoeMerchant on Friday September 02 2022, @10:33AM (3 children)
If FSD is reliant on a single camera, that needs to change for one simple reason: the world is messy and goop on the lens is no excuse for running over small animals and children.
Cameras are cheap, even good ones with good lenses. Multiple cameras are simpler and more reliable than complex automatic lens clearing schemes (like rolls of clear film).
For the money, two arrays of mid quality range 12MP cell phone cameras, say 20 each across the top of the windshield and front bumper, would provide simple resilience against dirt, insects, dazzle by sun reflection, etc. Half could be redundant (and varying perspective) wide angle and the other half could be overlapping telephoto, and total cost of both arrays should still be less than Lidar.
Sure, humans who are blind in one eye and only have one functional arm and no legs can still drive, and safely, but would you choose one to be your chauffeur?
Україна досі не є частиною Росії Слава Україні🌻 https://news.stanford.edu/2023/02/17/will-russia-ukraine-war-end
(Score: 2, Informative) by Anonymous Coward on Friday September 02 2022, @11:50AM (2 children)
> ... and total cost of both arrays should still be less than Lidar.
Maybe. Maybe not. From two years ago, here's Velodyne (one of the big names in Lidar) with a $100 unit. https://www.geoweeknews.com/news/velodyne-releases-100-lidar-sensor [geoweeknews.com]
There are Lidars for less than $100 (of varying quality) on the market now.
Maybe defense in depth by multiple cameras AND multiple lidars? I still wonder what they are going to do in a variety of low visibility situations (heavy rain, white out, etc) where people aren't very good...but people usually muddle through if speeds are low.
(Score: 2) by JoeMerchant on Friday September 02 2022, @12:57PM (1 child)
As you say though, there are LIDARs and then there are LIDARs - the cheap ones are potentially better than optical _for some situations_ but they suffer one simple problem: people don't "see" the way LIDAR does, so the environment isn't built to accommodate LIDAR the way it is visual sensors, even if LIDAR can "see better" it's also going to have confounds that nobody ever cared about before.
Any system benefits from multiple modalities of sensing data: visual cameras, IR and even UV, good LIDAR, crappy LIDAR, ultrasonic range finders, radar range finders, short range data links with nearby vehicles, long range data links with weather and traffic info systems, systems to estimate available traction in current conditions (temperature for sure, particulates kicked up by the tires, individual wheel speed sensors of course)
The problem is: when you integrate all this data, there will be unintended side effects. Traction estimation on a vehicle 1km ahead has an undetected faulty sensor and calls out "oil or ice", and the whole roadway slows to a crawl for no good reason, creating a hazard because of the slow and congested traffic...
Україна досі не є частиною Росії Слава Україні🌻 https://news.stanford.edu/2023/02/17/will-russia-ukraine-war-end
(Score: 0) by Anonymous Coward on Friday September 02 2022, @09:08PM
> the cheap ones are potentially better than optical _for some situations_
This made me stop and think for a minute. Maybe a near-term future development will be combined video and lidar, I wonder if they could share the same lens? Sensor fusion built-in right at the sensor. First they appear in high end cell phones integrated with VR sensor fusion, then in cars.
(Score: 2) by VLM on Friday September 02 2022, @12:41PM (1 child)
Its a self driving / self crashing car, so the owner would be blamed regardless of obvious software bugs or obvious design mistakes. Those products suck.
(Score: 2) by stretch611 on Saturday September 03 2022, @10:27AM
Actually, unless you can prove that you were cut off, the driver of the car in the back is generally the at fault party in all rear end collisions.
It does not matter if you stop short in the middle of nowhere... The car in back is usually cited with following too close if they do not have time to react to your actions.
I do not see any reason why this would change even if one or both cars are under AI control.
Now with 5 covid vaccine shots/boosters altering my DNA :P
(Score: 3, Interesting) by inertnet on Friday September 02 2022, @07:58PM (2 children)
Another way to determine distance would be to calculate the rate at which the size of objects changes. I don't know if this has been done before, but I believe it could assist in preventing errors as per TFA.
(Score: 2) by Osamabobama on Friday September 02 2022, @09:49PM (1 child)
Motion parallax would also be useful, except when heading directly toward an object. It would be perfect for a stop sign, though; stop at a location where the sign is moving quickly to the edge of the field of view. ('quickly' to be normalized to vehicle speed)
Appended to the end of comments you post. Max: 120 chars.
(Score: 2) by inertnet on Friday September 02 2022, @11:18PM
Combine these 2 and add stereo vision. Then you won't need lidar anymore, which might get confusing if every vehicle is sending out radar signals. With lidar you need a way to distinguish your own beam's reflections from the others. Even more fun in bad weather.
(Score: 5, Insightful) by istartedi on Friday September 02 2022, @10:02PM (1 child)
Current AI: This sign is a stop sign, but its dimensions fall outside the parameters of known valid stop signs. For all I know, it could be a stop sign on a child's T-shirt, so I'm going to ignore it.
Humans: That stop sign is a bit smaller than usual, but it's obviously a stop sign. I'm stopping.
Appended to the end of comments you post. Max: 120 chars.
(Score: 0) by Anonymous Coward on Saturday September 03 2022, @02:56PM
(Score: 1) by UncleBen on Thursday September 08 2022, @01:30PM
This is a thoughtful, citation-filled video on the effects of Tesla's decision to remove radar. It's not just stop signs that Tesla's getting wrong, and the video author's hypotheses seem highly likely.
https://www.youtube.com/watch?v=yRdzIs4FJJg [youtube.com]