EVNow
Well-Known Member
I don't think it has anything to do with HW3. This is the best they currently have.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Enhanced Summon, where are you?
- Thread starter dckiwi
- Start date
I was referring to it from a resource allocation perspective, and timing of it.
Why would you spend a ton of time optimizing performance and capabilities on an obsolete platform?
Especially when the Neural Network capabilities will drastically change with the new platform? and, I imagine a huge chunk of EAP owners who didn't already have FSD took advantage of the discount (which isn't really a discount) to upgrade to FSD, but there are probably a lot of hold outs who never felt like Tesla delivered on EAP.
Tesla has to release something because if they don't they'll be really screwed in owing EAP owners it, but not wanting to give them the HW3 computer for free. Cause then they'd have to refund all the EAP->FSD people.
I simply don't see any reason to have high expectations for it.
Not to mention HW3 is an entirely different computer that has redundancy designed to handle this kind of thing.
Plus it's an entirely separate computer from the Infotainment computer.
EVNow
Well-Known Member
First of all, summon doesn't depend much on NN. I think it is mostly software 1.0 - rule based code. This is why it "sucks"
Second, Musk has said the HW3 & HW2 NNs won't diverge until later in Q4. It seems to me, V10 will deliver the same NN to both HW2 and HW3. Verygreen has confirmed that HW3 has the same V9 NN now.
Third, Karpathy in his talk made a joke about summon - "its great …. when it works". I doubt he would have said it if summon used NN much and he was partly responsible for its functioning.
Ofcourse, summon would use NN for basic object recognition (including drivable area, curb etc). Some of them are new tasks too - and are needed for both Feature Complete and for Summon. But they are all in V9.
First of all, summon doesn't depend much on NN. I think it is mostly software 1.0 - rule based code. This is why it "sucks"
Second, Musk has said the HW3 & HW2 NNs won't diverge until later in Q4. It seems to me, V10 will deliver the same NN to both HW2 and HW3. Verygreen has confirmed that HW3 has the same V9 NN now.
Third, Karpathy in his talk made a joke about summon - "its great …. when it works". I doubt he would have said it if summon used NN much and he was partly responsible for its functioning.
Ofcourse, summon would use NN for basic object recognition (including drivable area, curb etc). Some of them are new tasks too - and are needed for both Feature Complete and for Summon. But they are all in V9.
Basic object recognition that is fast, and highly accurate is of monumental importance for anything in the parking lot. Sure V9 has a lot of object recognition capabilities (cars, driveable areas, pedestrians), but HW3 offers a lot more power to detect those at a faster frame rate, higher resolution, and accuracy.
I'm aware that the NN's won't diverge until later in Q4 for public cars. But, I consider that to be near term when talking about development. Resources are so tightly constrained that if something is going to be abandoned then I doubt much effort will go into getting them to work beyond basic functionality.
You seem mostly focused on the NN aspect of HW3, but there is also the FSD driving aspect as well. Those two things go hand in hand.
I think we both agree that the lack of rules based code is what makes summons kinda suck by the average observer. It's not that it's running into anything , but that it simply doesn't seem to take correct path.
It's not just Enhanced Summons, but NoA as well.
I find NoA almost infuriating because of a lack of rules based coding. Like as example I expect the car to get out of the passing lane at a specific time/space interval. But, after lots of testing I can't determine what this interval is. Sometimes its 10 seconds, and sometimes its 20 seconds. It seems pretty random, and there are times where it just doesn't (especially when an HOV lane is the left most lane, and I'm in the passing lane).
It seems to me that Tesla is simply trying to reach "feature complete" with EAP, and they're not too concerned making any of the features work all that consistent.
As to comment Karpathy made I interpreted it as a joke based on a feature plagued by delays. I wouldn't put much meaning into what it means for Summons. Plus Karpathy is the director of AI at Tesla so he's most certainly responsible for Summons working well. NN's are not the only thing under the umbrella of AI.
EVNow
Well-Known Member
Lack of rules ?
No - what I said was, since it is rule based code (rather than done by NN), it sucks i.e. I said the opposite.
Basically NN passes on the perception to software 1.0 / rule based code / heuristics - which implements the planning & execution. This makes summon difficult because in a parking lot the rules are not clear, unlike on the roads.
Generally speaking if you have a rules based system without a lot of work going into the rules then its going to suck. Just like if you did reinforcement learning on a very limited dataset it would suck.
The complexity of parking lots combined with the failure of lots of drivers to follow the rules that do exist make parking lots an extremely challenging environment.
But, that doesn't mean that parking lots can't be mapped, and generic rules generated. I'm not seeing any evidence that any of that was done.
EVNow
Well-Known Member
Yes, not enough rules have been implemented. That takes time.
Also, you could go completely NN and forgo rules based code. Thant has not been attempted either.
As long as the car can reliably detect every item in and around its path, why would you need a NN over rules? The NN’s realm is pattern recognition.
The rules aren’t that complicated (otherwise people wouldn’t be able to negotiate a parking lot). They are basically: “follow the route (humans use a more or less fuzzy memory of the overall parking lot layout, computers use navigation maps), stay on the right side, DON’T CRASH INTO ANYTHING, move slowly and follow the basic yield rules”.
The “don’t crash into anything’’ is the big nut to crack as this requires a total 3D understanding of what’s around the car. And that’s where the NN comes in. And that’s where the problem with Mr. Musk is, IMO. It seems he was impressed by a MobileEye/NVidia machine vision demo a few years ago and thought, “wow, the computer can identify what it sees”. Turns out, it’s not so easy after all.
The rules aren’t that complicated (otherwise people wouldn’t be able to negotiate a parking lot). They are basically: “follow the route (humans use a more or less fuzzy memory of the overall parking lot layout, computers use navigation maps), stay on the right side, DON’T CRASH INTO ANYTHING, move slowly and follow the basic yield rules”.
The “don’t crash into anything’’ is the big nut to crack as this requires a total 3D understanding of what’s around the car. And that’s where the NN comes in. And that’s where the problem with Mr. Musk is, IMO. It seems he was impressed by a MobileEye/NVidia machine vision demo a few years ago and thought, “wow, the computer can identify what it sees”. Turns out, it’s not so easy after all.
diplomat33
Average guy who loves autonomous vehicles
It's easy for us laymen to say "it sucks" because enhanced summon does not meet our expectations. But enhanced summon is difficult. I am sure it required a lot of hard work. I think if we knew what was actually happening under the hood, we might be more impressed with the feature.
EVNow
Well-Known Member
That is changing - there is realization that you need to move more and more features into NN to have a shot at L5.
Actually rules are very complicated. They don't even exist for most part - for eg., it is not uncommon to go on the wrong side of the parking lane. If there are people on the lanes constantly, when can the car drive ? Try to write down the rules (in if -then -else) and see how far you get.
If a feature doesn't meet expections or doesn't work as expected, it sucks by definition, no matter how much work went into it. Further, why offer a feature that sucks? The answer is that people will still buy a product that doesn't work as intended.
Finding a path through known obstacles based on a rule set has been solved by CS decades ago. I don’t have to see how far I will get. This is basic computer science.
The core of the autonomy problem is reliably understanding all objects in and around the path. That’s why Tesla designed their own NN optimized chipset and that’s why others slap LIDAR and other sensor monstrosities on their cars.
Enhanced Summon still almost does not suck because the car is not reliably understanding what’s in its path and an error is EXTREMELY costly. That’s why it hesitates and “drives” like a oddly concerned drunk person with severe eye impairments and no hearing in these youtube videos.
diplomat33
Average guy who loves autonomous vehicles
I don't think you'll ever find complete agreement as to whether a product sucks or not.
EVNow
Well-Known Member
You are talking about the static world. In the AI interview with George Hotz, he explains this well.
- Static world
- Dynamic world, where lots of actors are moving and you have to predict their behavior
- your action is going to change the way the various actors are going to change their behavior and you have to predict the changes depending on your choices
willow_hiller
Well-Known Member
That's starting to sound like the n-body problem from physics! No wonder this is so hard to solve.
boonedocks
MS LR Blk/Blk 19”
diplomat33
Average guy who loves autonomous vehicles
well, I agree that Enhanced Summon sucks less than the current Summon we have.
AlanSubie4Life
Efficiency Obsessed Member
Do you think that's true on an E[Property Value Loss] per use basis? I'm not sure, even though I am only going by videos of Enhanced Summon at this point.
Could we use that as a metric for suckage?
Similar threads
