Well, I guess in theory that could be of sufficient quality to feed into the NN piecemeal.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
HW2.5 capabilities
- Thread starter verygreen
- Start date
-
- Tags
- Autonomous Vehicles
Well, I guess in theory that could be of sufficient quality to feed into the NN piecemeal.
Sure, that's always been the question - do you need lidar or not? I don't dispute (nobody does) that having lidar would help advance the reliability of self driving cars in the short run. Tesla's gamble was that massive fleet reinforcement learning could overcome the lack of expensive lidar. And on the flip side, lidar's high price meant that nobody was every going to put it into a massive commercial scale customer fleet for large scale reinforcement deep learning.
Oh I thought it was several minutes of the trip. Only 10 seconds eh? But for that short 10 second clip is it least the entire total of all sensory input?
Very good and interesting conversation, guys. Just a short post of appreciation here. 
Yes, more data can be found here: Massive data uploads since 2017.34 2448cfc
Sorry . I think I double posted.
Also, there is a thread around started by @verygreen I believe where he captures the actual snapshot of whats sent to the mothership including all 8 camera feeds and someone else actually procssed the data. but I can't seem to find the thread right now, but it's here in the forums somewhere.
Found it: AP2.0 Cameras: Capabilities and Limitations?
Lot's of great data in this read, probably will want to read the entire thing honestly.
Lot's of great data in this read, probably will want to read the entire thing honestly.
stopcrazypp
Well-Known Member
@calisnow beat me to it already. Having around 100k vehicles on the ground (and poised for a massive increase with Model 3 release) is a fairly big advantage. When the others talk about volume, it's measured in the hundreds, because of the same difference I discussed in the other comment (something made for consumer use, vs commercial fleet use).
AP2 is delayed versus the initial expected schedule, but putting aside the whole false advertising argument, in terms of what they accomplished in this time frame, it's still impressive to me. And this has been discussed on other threads, but it's not safe to assume that the team working on AP2 and FSD is the same, so progress on AP2 may not necessarily be a good indicator of how far along FSD is.
As for FSD development, Tesla has kept quiet about it, and I think that is a good thing. I don't think they want a repeat of what happened with AP2 (both the expected timeline initially published on the website, and also Elon's tweets). We'll see our next update on the cross country trip and also in the disengagement report released next year.
I believe you know I agree with this.
I am not basing my thought on EAP progress alone, not at all.
lunitiks
Cool James & Black Teacher
I am. That's all I have. Everything else seems to me like flying teapots
Yes, Tesla has an advantage on the verification possibilities through their fleet.
The bigger question, though, is how far behind are they of the competition? How much can such advantages offset the things where Tesla is behind.
It will be interesting to watch. I have a horse in the race too - I'd certainly love for my Model X to get a Level 5 update sometimes soon.
The bigger question, though, is how far behind are they of the competition? How much can such advantages offset the things where Tesla is behind.
It will be interesting to watch. I have a horse in the race too - I'd certainly love for my Model X to get a Level 5 update sometimes soon.
Mobileye claims to be doing exactly that and to have signed agreements during the last year with multiple huge auto companies. Well - they claim they are about to start doing it and that the agreements are in place. They must have hashed out the cost/benefit and data sharing issues already. You can look this up - it's all public and Amnon gives some very good talks about his "road experience management" mapping project to get low bandwidth maps built - it does look ingenius. He does not, however, claim to be doing or to be planning to do reinforcement learning using a giant crowd sourced fleet - or if he has I have never seen the claim.
Yes, Tesla has an advantage on the verification possibilities through their fleet. The bigger question, though, is how far behind are they of the competition? How much can such advantages offset the things where Tesla is behind. It will be interesting to watch. I have a horse in the race too - I'd certainly love for my Model X to get a Level 5 update sometimes soon.
Yes and the follow-on question is - can the competition, though it might be ahead in terms of disengagements in the real world at this moment, actually cross the finish line to an acceptable level of reliability without a massive scale fleet reinforcement project? The competition could be 95% of the way to the finish line and Tesla is only 20% - but if it turns out you need a Tesla-scale testing network then if the competition doesn't build one with another OEM manufacturer then they could diddle around, right in front of the finish line for several more years while Tesla catches up and passes them.
I am surprised, actually, that after this many years - no other automaker has teamed up with an AI group to do exactly what Tesla is now doing (though many partnerships are being announced left and right - indicating that maybe everyone is privately reaching the same conclusion but won't admit it yet). Maybe it is a structural problem and a cost/political problem with other automakers that can't be overcome - maybe you need a fast moving automaker with a silicon valley mindset to have produced this kind of test fleet by 2017.
OR - maybe you do not actually need huge fleets - I dunno. Maybe that's why nobody else is doing it this way yet.
Final thought then I have go back to the boring business of running my actual business - is that perhaps Tesla has another large advantage in its customer base. Tesla's sales growth has proven that the customers have been more than willing to let Tesla "move fast and break things" and in fact to be part of a massive beta test. It's possible that Audi/Daimler/GM would have a hard time get their customer base to accept being part of a giant test project even it ultimately advanced the state of the AI better and faster than any other method.
Every time I read some dumb journalist throwing off the sound bite "Only Tesla is irresponsible enough to use its customers as beta testers for self driving" I wanted to say "Read just a tiny tiny bit about how deep learning works and you will realize that perhaps there is no other way to bring self driving to fruition without the data set that only a commercial scale vehicle network can provide."
Every time I read some dumb journalist throwing off the sound bite "Only Tesla is irresponsible enough to use its customers as beta testers for self driving" I wanted to say "Read just a tiny tiny bit about how deep learning works and you will realize that perhaps there is no other way to bring self driving to fruition without the data set that only a commercial scale vehicle network can provide."
@calisnow Given sufficient many years into the future, anyone can win and be right. It is a question what can be seen today, from the available data. It seems quite foggy to me.
That said, I can see a scenario where everyone basically is wrong. Tesla going Lidar, yet winning on fleet or something. Many things are possible, if we consider all alternatives.
Which ones are likely, that is a bigger question.
Frankly, I think Google will win the autonomous driving one - and Audi will probably win Level 3.
That said, I can see a scenario where everyone basically is wrong. Tesla going Lidar, yet winning on fleet or something. Many things are possible, if we consider all alternatives.
Which ones are likely, that is a bigger question.
Frankly, I think Google will win the autonomous driving one - and Audi will probably win Level 3.
I wouldn't be surprised if you're right. Google's Android is so far ahead of Apple's iOS in terms of usability, design and useful AI it isn't even close. If that kind of expertise with learning and big data carries over to self driving - I can see Google getting there first in a commercial solution.
Similar threads
- Article
