Level 4 autonomous driving is the minimum requirement for that, and that is years away.
(Check anon shashua most recent utube lecture.)
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
- Status
Level 4 autonomous driving is the minimum requirement for that, and that is years away.
(Check anon shashua most recent utube lecture.)
Anstandswauwau
Member
Every major technical breakthrough is always "years away" according to naysayers and critics until someone pulls it off anyway
I'm nowhere near as knowledgeable about autonomous driving outside of Tesla's as many on TMC are.
That said, I don't see how claiming autonomous driving is years away squares with how close Waymo seems to be getting.
Waymo makes history testing on public roads with no one at the wheel
"Waymo, the Alphabet self-driving car company, now has cars driving on public roads in the Phoenix metropolitan area with no one in the driver's seat. Waymo CEO John Krafcik plans to announce the news today in a speech at the Web Summit in Lisbon, Portugal."
I'm sure there are caveats and restrictions that apply to what Waymo is doing in Phoenix. However anything this technologically difficult has to proceed in stages. If Waymo shows this works well in one city, they'll continue refining their approach and extending it to other cities with fewer compromises. That seems closer than years away.
I don't know the progress of other self-driving startups. I believe Waymo is on the wrong track and Google will cancel the project within 3 years.
We should finally get some answers about the Tesla Tax today. The parties have agreed on a budget for 2018, and are presenting it in 1.5 hours.
DurandalAI
Member
I think people are making expectations based upon what AP2 cars are doing now, in the thought that it is representative of where EAP is in customer cars. I’ve never been of the opinion that EAP is representative of where things are for FSD internally. I base this on my opinion that after the initial fallout with Mobileye after the one AP1 accident hit the news that they’ve been much more cautious on what they push out to customers. I think that they’re probably only 3-6 months behind where they originally had spoken to.
Tesla has barely reached the level of AP1, which at best is level 2 autonomy.
Level 4 is a ways off .
Tesla today reigns supreme in electric cars, not autonomous software.
Level 4 is a ways off .
Tesla today reigns supreme in electric cars, not autonomous software.
Last edited:
2virgule5
Member
AP is level 2, yes, but it also is the single best system you can buy and use today. We shall see if other manufacturers can catch-up and surpass them.
I speculate that they will show off a tweak to autopilot that amounts to "platoon" functionality. Coast-to-coast no human touch the controls could be a car that is following a car with a human driver.
Anstandswauwau
Member
AP1 / Mobileye's EyeQ3 undoubtedly had a lot of smart tricks up its sleeve (which, apparently was enough to get to semi-level 2 autonomy), but it relied heavily on mere oldschool computer vision. And quite a bit of smoke and mirrors.
Around '14/'15 I think both Tesla and Mobileye figured out this wouldn't make for a great foundation for level 3+, hence their shift to deep neural nets with later versions. This came with the drawback of virtually having to start from scratch again, which was painful but ultimately necessary.
Now, if you compare neural nets (AP2) against traditional algos (AP1), there's a pattern that can be seen throughout the industry right now:
The neural nets usually don't hold a candle to traditional algos until a certain tipping point, but boy, once it reaches it, it runs circles around traditional algos.
I can pretty much guarantee you that Tesla's AP will hit this point within the next 6-24 month, which, by the way, perfectly aligns with Mobileye claims of reaching level 5 autonomy by 2020.
And one last thing: FSD is NOT some sort the holy grail of software engineering. (General AI is.). It's a software problem, like many others before, that is first deemed as being impossible to achieve, then it will eventually be solved and in less than 5 years later you'll be able to grab your open source FSD models from Github. That's how those things work. No magic sauce required.
And the Tesla tax is dead, for now. Maybe next year.
Interesting perspective.
We have AP1 and AP2. Both were on 2017.42 (AP2 just got updated to 0.44 last week). Previous to 0.42 AP2 kinda sucked. Now it’s working great. Some minor things here and there.
AP1, interestingly, got more glitchy with 0.42. I’m thinking the NN move is with 0.42, and some of the glitches I’m seeing with AP1 is because of the transition. I’ve emailed Tesla with my observations regarding the changes.
I think it fits with your idea.
BTW the glitchiness doesn’t impact drivability.
sub
Active Member
sundaymorning
Active Member
A buddy of mine was at his service center and saw a model 3, the branch is in Southern California. He was told that the branch he was at is delivering about 3 model 3s per day. FYI.
Reciprocity
Active Member
Maybe, maybe not. Their issue is that they are going very heavy on the hardware side. The software is much simpler when you have $70k lidar and a truck full of super computers. This is the biggest barrier to those types of solutions. All that hardware has to be made much smaller and cheaper, which means mass produced, to become cheap enough to make it less expensive then a driver much less as cheap as Tesla's solution. The difference is much easier when there are no sacrifices to hardware.
For Tesla every problem seems to look like a nail and Tesla has a really big hammer. Meaning every problem seems like a software issue that can be solved without out massive advances in hardware. They seem to forecast software advances and assume it will run on today's hardware. My guess is that any hardware they thought they might have to upgraded, is upgradable.
Elon classically called the problem solved, it's just about the data and software. Recently they really started consuming massive amounts of data and recently we started seeing reports of new neural networks in the firmware updates. If you all recall, I said that AP or Auto steer was not FSD and that it was a separate development path. One was a bandaid and the other is the real solution. I think we are finally seeing the FSD parts. This is going to be a slow slog but it will be very S curvy. The example Elon gives is alpha go beating a single go champion 3-2 or something like that. It took maybe a year or two to get there. Not 2 months later, alpha go beat 20 champions simultaneously and could not be beaten. So in around 20 months or something like that, they went from impossible to beat anyone to impossible to be defeated by anyone with massive advances in the last 2 months. This was not solved by hardware. Open ai did something similar with Dota but uniquely, the system learned on ours own by playing itself and learned very quickly. This accelerated how fast it went from nothing to beating champions.
I don't know enough to know what it will really take, but I'm confident that Tesla is on the right path and Waymo is not as it relates to making a real product. I have no doubt Waymo will have fully autonomous taxi cabs with no drivers first, it will just require $100,000 in hardware and 5 more years to get the cost, size and power requirements down.
Name one other company that doesn't have a porcupine car with massive computing power in the trunk? Maybe Intel/Mobileye.
Some weeks back a reliable member (Bonnie) reported her MX AP2 was working much better for her after downloading a new software update. The jury is still out on whose approach is going to work best and when. Software often times exhibits dramatic jumps in performance/effectiveness that are not easily predicted during a period when incremental fixes have been the rule.
I love the way Tesla has integrated its sensors and processing into its products. Waymo autonomous cars look weird and kludgy.
My money is still on Tesla to eventually catch and pass Waymo.
Anstandswauwau
Member
No disrespect to anyone, but what he showed in this presentation – which might not even be representative of Mobileye's status quo or potential future – is the complete opposite to how Tesla, comma.ai, Uber and most likely Waymo approach FSD.
- Mobileye's approach, according to this presentation:
Let's divide driving into x sub-tasks / activities / goals and try to come up with a separate mathematical equation or logic for each and everyone of them.
- Everyone else (simplified):
- Let the neural net figure out what driving is and how it works based on provided training data.
- Sandbox the AI and set hard limits ("Under no circumstances do this or that ever." etc.)
- Optional: Have an observer AI judge the driving AI's decision.
- Let it self-improve in simulations and/or shadow driving until it's x-times safer than a human driver in any given situation. Adjust and improve training data, if necessary (go back to step #1, rinse and repeat)
Mobileye's goal is to build the best-possible robot-like driver agent, everyone else tries to recreate a flawless human-like driver using AI.
Both approaches and goals come with their respective set of pros and cons.
My bet is on deep neural networks.
- Status
Similar threads
- Sticky
- Locked
