Please post. You didn't say you're a domain expert.
Sorry my bad - seriously, please contribute. I mean it.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Sorry my bad - seriously, please contribute. I mean it.Also sorry I didn't see the last paragraph where you put your opinion upon first read. I only saw the definition of L5.
Yes I was hoping some experts would take those factors into consideration. It sounds like you're saying processing power and code is the limiting factor - am I right? If so, how does adding even more sensors lesson that load? Is the argument that radar signatures and lidar data is easier to interpret than visual data, therefore requiring less computing power/simpler neural networks and is thus easier to achieve in the short term?
It's called a screen/sort. Yes - when looking for info in a field one doesn't understand, one uses imperfect signaling devices (such as degrees/employment field) to make a quick judgement on who to listen to. It's imperfect - invariably you throw out good sources of information, but you also likely/hopefully throw out more bad sources than good ones when time is limited. Sorry if it upsets you. I have no agenda - I just want to hear what actual domain experts think because the forum is thick with uninformed opinions (including mine).
thegruf
Active Member
... many years working with machine vision systems.
my 2c
wrong question!
It's not what the systems "see".
It's what they miss.
Human drivers process visual field only, no LIDAR, no RADAR.
So in theory this is possible to replicate with machine vision alone, this I believe is primarily Tesla thinking.
LIDAR and RADAR both suffer from noise, and LIDAR in particular is compromised in anything but excellent visibility, though both can measure over considerable distances more accurately than a camera typically is able to.
However though machine vision can be more accuarate and reliable in some conditions, having sufficient AI to apply context as a human would to interpolate a situation is some way off.
The problem however is this:
Human drivers kill 1000's of other humans every year. This has become accepted as a (somewhat unpalatable) norm.
Automation kills one single human (even if said human was not acting according to the rules provided) but quietly reduces accident rate by 40% thereby statistically saving more lives than the one sadly lost. Result - Hysteria.
Likely outcome: Tesla will continue to make phenomenal strides in real world application of autopilot. Most probably L5 FSD will be restricted to Highways only with curent technology and good visibility due to the reduced scenarios for accidents, with lower levels applicable elsewhere. Will still meet Musk's claim though.
my 2c
wrong question!
It's not what the systems "see".
It's what they miss.
Human drivers process visual field only, no LIDAR, no RADAR.
So in theory this is possible to replicate with machine vision alone, this I believe is primarily Tesla thinking.
LIDAR and RADAR both suffer from noise, and LIDAR in particular is compromised in anything but excellent visibility, though both can measure over considerable distances more accurately than a camera typically is able to.
However though machine vision can be more accuarate and reliable in some conditions, having sufficient AI to apply context as a human would to interpolate a situation is some way off.
The problem however is this:
Human drivers kill 1000's of other humans every year. This has become accepted as a (somewhat unpalatable) norm.
Automation kills one single human (even if said human was not acting according to the rules provided) but quietly reduces accident rate by 40% thereby statistically saving more lives than the one sadly lost. Result - Hysteria.
Likely outcome: Tesla will continue to make phenomenal strides in real world application of autopilot. Most probably L5 FSD will be restricted to Highways only with curent technology and good visibility due to the reduced scenarios for accidents, with lower levels applicable elsewhere. Will still meet Musk's claim though.
Wow thank you so much - most detailed commentary on the problem/situation I've yet read. I now have even more questions - will ask you some tonight after work here on this thread.
I have a BS in Computer Science and currently work in the implementation(inference/forward pass) side with deep neural networks for computer vision.
That out of the way, yes, of course. Even vision alone is demonstrably sufficient.
If the bar is being able to safely handle driving in all environments that a human can, then you have been using an L5 suite since you drove your first car that relies almost entirely on just two cameras(with the ability to swivel to something like 135 degrees in each direction).
The trick here isn't the sensors. Tesla's AP2 sensor suite has better coverage and tracking than you do, by far. The hard part is being able to train a network to use those inputs to perform at least as well as a human can, while still being able to process those inputs fast enough.
That out of the way, yes, of course. Even vision alone is demonstrably sufficient.
If the bar is being able to safely handle driving in all environments that a human can, then you have been using an L5 suite since you drove your first car that relies almost entirely on just two cameras(with the ability to swivel to something like 135 degrees in each direction).
The trick here isn't the sensors. Tesla's AP2 sensor suite has better coverage and tracking than you do, by far. The hard part is being able to train a network to use those inputs to perform at least as well as a human can, while still being able to process those inputs fast enough.
To start with, I insist on the use of SAE definitions from the PDF posted earlier. Any other basis will not work for purposes of discussion, because I refuse to discuss on the basis of arbitrarily defined parameters of what L5 may be.
On that basis, look through the PDF I posted earlier. There are 4 SAE vectors:
1. Execution of Steering and Acceleration/ Deceleration
2. Monitoring of Driving Environment
3. Fallback Performance of Dynamic Driving Task
4. System Capability (Driving Modes)
L2 through L5 progressively adds each of these to System role, from Human Driver. Item #1 is essentially Autosteer in Tesla language (which is why AP1 is a L2 system) . Item 2 is significantly covered by 8 cameras on AP2. *But* 'monitoring' and system fallback are hugely different roles, which means item #3 is a non-trivial jump from item #2, i.e. from L3 to L4 is hard. There *will* be regulatory and liability insight into a system driven fallback approach. Item #4 (from some driving modes to all) is the hardest, which implies L4 to L5 will be a very hard move.
In my opinion, Elon's definition of FSD is essentially something between L3 and L4, skirting the point where he can technologically accomplish substantial autonomous capability, but not complete 'lie in the back and sleep' autonomous capability that regulatory reform will not permit anytime soon.
Do you agree with @Sir Guacamolaf and @thegruf in their characterizations that training the neural network to successfully process the input from the cameras is years off. If I've misinterpreted your positions @thegruf and @Sir Guacamolaf please correct me.
And to all three of you - and any other experts here - does the addition of lidar and/or more radar sensors reduce the computation required?
Last edited:
No comment in that regard. It's impossible to even guess, unless you're part of Tesla's AP management or trying to read the tea leaves of what they say or do, what kind of time table they're on.
That said, I highly doubt the networks we have are anything close to what Tesla is developing towards FSD. They're clearly bringing in some of the vision networks(evidence of this in that 8.1 appears to use parallax between the two cameras since it can identify stopped cars), but since they aren't using the side or back cameras at all, the networks that take them as input clearly aren't there(or aren't being used). More qualitatively, it *feels* like steering and TACC are rules based currently, and I'd imagine those would need to move to an ML approach for FSD.
If every bit of technology they have towards FSD can be said to already be activated in our current cars, then yes, they're a long ways away. But that's almost certainly not the case, and it's tough to comment on the status of software we've never seen.
Missed the question about decreasing computation required. Unfortunately, the best answer to that is: *shrug*. It all depends on what the network can learn. If the addition of lidar/radar allows a smaller network to achieve the same performance of a larger network without it, then yes, it'll likely shave off processing time. If not, then it won't.
Since the whole *thing* in machine learning is training a computer to do a task you're unable to fully define, the only real way to tell which of those is the case is empirically; by running a set of experiments and seeing what happens.
Since the whole *thing* in machine learning is training a computer to do a task you're unable to fully define, the only real way to tell which of those is the case is empirically; by running a set of experiments and seeing what happens.
thegruf
Active Member
many safety systems work on a 2oo3 (two out of three) principle or similar.
The theory being that two systems have to concur to result in a response.
It couid be argued that camera's alone are just one input. Many cameras could be considered many inputs (if they overlap) but these are consolidated to a single input through MVP.
Adding RADAR and/or LIDAR in theory can provide confirmation inputs to the system: camera thinks it detects something/radar agrees - system responds.
To some extent Tesla's now massive autopilot database can also be considered a 3rd input, at least for location information. Fundamentally knowing exactly where you are is critical to any decisions to be made, threats to be assessed, scenarios to be interpreted.
I think it is important not to exclude the latter when considering FSD, this is outside my specialist field but clearly many in the industry see that Tesla have a big, potentially unassailable lead in this area.
Even for highway use, I still have a conern that with only forward facing radar whether overtaking vehicles in adjacent lanes can be sufficiently characterised for speed by cameras alone, especially relevant during lane change of course. Had three radars been fitted, with two rear facing (perhaps from the mirror housings) I could forsee an easier path to FSD.
The theory being that two systems have to concur to result in a response.
It couid be argued that camera's alone are just one input. Many cameras could be considered many inputs (if they overlap) but these are consolidated to a single input through MVP.
Adding RADAR and/or LIDAR in theory can provide confirmation inputs to the system: camera thinks it detects something/radar agrees - system responds.
To some extent Tesla's now massive autopilot database can also be considered a 3rd input, at least for location information. Fundamentally knowing exactly where you are is critical to any decisions to be made, threats to be assessed, scenarios to be interpreted.
I think it is important not to exclude the latter when considering FSD, this is outside my specialist field but clearly many in the industry see that Tesla have a big, potentially unassailable lead in this area.
Even for highway use, I still have a conern that with only forward facing radar whether overtaking vehicles in adjacent lanes can be sufficiently characterised for speed by cameras alone, especially relevant during lane change of course. Had three radars been fitted, with two rear facing (perhaps from the mirror housings) I could forsee an easier path to FSD.
thegruf
Active Member
I thought the improved detection of stationary/partially obstructing objects in path was down to the new active reconstruction mode of the radar signal with whitelisting that Tesla have developed?
Ah, possible. I hadn't seen any explanations of it and just suddenly saw that my AP2 car could pick up stationary objects it never saw move before. And, given that they activated a second forward facing camera for the same release, parallax seems the most obvious explanation for that.
Gizmotoy
Active Member
I meet one of your expert criteria but do not feel I am qualified to give an expert opinion. Your criteria are too easy.
If FSD is vision based, what happens if the vision is obscured with glare, snow, rain, mud, etc?
As a human, we squint our eyes or turn our head. The cameras are in fixed locations without any ability to clean or shade themselves.
I'd be curious what the car will do if it is temporarily blinded.
If FSD is vision based, what happens if the vision is obscured with glare, snow, rain, mud, etc?
As a human, we squint our eyes or turn our head. The cameras are in fixed locations without any ability to clean or shade themselves.
I'd be curious what the car will do if it is temporarily blinded.
Similar threads
