-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Video: Tesla is going to win Level 5
- Thread starter Curt Renz
- Start date
diplomat33
Average guy who loves autonomous vehicles
Tesla's machine learning incremental approach is a smart approach to get to L5 because it will slowly add all the features needed one by one until the car eventually gets to L5. And since Tesla uses data from a lot of cars all over the word to train the NN, Tesla's FSD is not dependent on mapping or location. The features can work everywhere given enough reliability. And Tesla just needs to grind through the work of training the NN with data. It's laborious but in time, it will get there. With machine learning, Tesla can train the NN to do anything with enough data. The big question is when. We have no idea how long it will take Tesla to finish training the vision NN needed to get to full autonomy. And even when Tesla does get there, that will still leave a lot of work of driving policy etc before Teslas can reliably get to L5. So, yes, I believe it is the right approach, but it's not a done deal. There is uncertainty about whether Tesla can actually get to L5 in a timely manner.
On the other hand, Waymo has a proven approach to get to L4. In fact, their cars are essentially L4 now. But the system only works in geofenced areas that are carefully premapped first. This makes it hard to get to L5 because you can only deploy your cars after the mapping and testing phase. Waymo cannot just deploy their cars everywhere. Waymo will need to either improve the reliability of their system where it does not need premapping anymore or get HD maps of the entire US if they want their L4 to work in the entire US. Chances are that Waymo won't even bother doing that. Instead, Waymo will just focus on premapping select areas of cities where their robotaxi service will be most profitable.
On the other hand, Waymo has a proven approach to get to L4. In fact, their cars are essentially L4 now. But the system only works in geofenced areas that are carefully premapped first. This makes it hard to get to L5 because you can only deploy your cars after the mapping and testing phase. Waymo cannot just deploy their cars everywhere. Waymo will need to either improve the reliability of their system where it does not need premapping anymore or get HD maps of the entire US if they want their L4 to work in the entire US. Chances are that Waymo won't even bother doing that. Instead, Waymo will just focus on premapping select areas of cities where their robotaxi service will be most profitable.
Silicon Desert
Well-Known Member
I want level 6 where the car goes on a vacation for me and takes photos to bring back home 
Seriously though. I don't care who gets there first as long as my Tesla gets to level 5 before I die
Seriously though. I don't care who gets there first as long as my Tesla gets to level 5 before I die
Last edited:
I’m 63 years old. I lost about 80% of my hearing, half of my my vision, my hand eye coordination is not what it used to be and my reflexes just suck.
My timeline for level 5 is five years MAX!
My timeline for level 5 is five years MAX!
Uncle Paul
Well-Known Member
Jack.
Many of us older guys are in the same situation. Our driving skills are in decline, and a great Autopilot would give us many more years of independent driving.
Many of us older guys are in the same situation. Our driving skills are in decline, and a great Autopilot would give us many more years of independent driving.
I'm 53, in OK health. I have no expectation that real level 5 will be a mass-consumer product in my lifetime. Maybe some kind of limited automated ride share will exist, but jumping into my car and waking up at some rural remote B&B hundreds of miles away the next morning? I'm not holding my breath.
Maybe some kind of limited automated ride share will exist, but jumping into my car and waking up at some rural remote B&B hundreds of miles away the next morning? I'm not holding my breath.I know this is the narrative of both Tesla and the Tesla community but I am not convinced it is true personally.
I think it is very possible Waymo has a more generalized solution today than Tesla does. And a faster roadmap to Level 5 than Tesla does.
The theory goes like this:
1. Tesla’s suite is possibly not sufficient for Level 5 and this may delay Tesla in both development time and in deployment advantage
2. Tesla’s progress has been underwhelming, there is higher than zero chance they just can’t get FSD to work reliably
3. People may be overestimating how hard it would be for Waymo to operate their cars ”anywhere” (remember, Google maps the world all the time with increasing precision)
4. Tesla faces risks as a company that Waymo probably does not that may limit their investment capabilities
It is one theory.
I have been on the sidelines of the recent exchanges because it really is all speculation until we know more. But just throwing it out there.
My personal investment in an AP2 FSD car of course means I hope this theory is wrong.
Last edited:
It amazed how fast we went from resizing JPEGs to uploading every photo of visits to Starbucks. The storage capacity has increased exponentially—and that growth continues.
Each Tesla with suite of cameras will upload and astonishing amount of data. I haven’t delved into the subtleties of Tesla’s NN, but I don’t underestimate what can be achieved with amount of information.
This will happen sooner than many think.
Each Tesla with suite of cameras will upload and astonishing amount of data. I haven’t delved into the subtleties of Tesla’s NN, but I don’t underestimate what can be achieved with amount of information.
This will happen sooner than many think.
diplomat33
Average guy who loves autonomous vehicles
I'm 53, in OK health. I have no expectation that real level 5 will be a mass-consumer product in my lifetime.
Assuming you live for another 30+ years which is likely, L5 will definitely happen in your lifetime.
If it is possible to generalize a L4 geofenced system into a L5 system, then yes, Waymo has a faster roadmap to L5 than Tesla. Put differently, if it is just a matter of "copying and pasting" their L4 system into other cities, with different HD maps and some minor tweaks, then yes, Waymo has a direct path to L5.
Now I know that Waymo is testing in lots of different areas but I looked up where Waymo One is currently operating their taxi service and it only covers about 300 square miles. That's really tiny! The continental US is 3 million square miles. If Waymo has a direct path to a generalized solution why aren't they deploying faster? It is entirely possible that generalizing a L4 system into a L5 system is not so easy. Waymo's L4 may still be missing edge cases and lack the reliability to be L5. The presence of safety drivers seems to support that. So I don't think Waymo's path to L5 is quite so obvious.
There is also the possibility that Waymo is not even interested in L5 at all. Waymo's business model may be to simply replicate a L4 robotaxi service in say the top 10 or top 20 major metro areas of the US. That would be far easier than L5 and would still give Waymo plenty of customers and demand for their service to have a very profitable business.
But I am certainly not implying that it will be easy for Tesla. I agree that Tesla's hardware is a bit of a question mark. Elon is counting on vision being good enough in which case the 8 cameras should be sufficient. In theory that is true. In practice, it may be a lot harder. And like I said, earlier, even if Tesla does achieve a super reliable vision NN for L5, there is still the work of driving policy etc... Dealing with edge cases and weird driving behaviors of other drivers, may be a whole challenge in itself.
I think it is entirely possible that Tesla does achieve a pretty good L3 or even L4 system but struggles to get to L5.
EVNow
Well-Known Member
After reading about how they tackle parking lots, I'm less convinced about Waymo. They literally annotate the parking lots with drivable path, alternate path to take if the first one is blocked (!). How is that scalable ? Definitely not L5, but then, they can be L4 and serve just top 200 US/EU markets.
I think that also makes them a lot more conservative ... not the best for bold innovation.
Personally - I've argued this elsewhere - it is not clear to me what path is faster.
- Tesla from L2 everywhere to L4 everywhere.
- Waymo from L4 in one suburb to L4 in 200 large urban/sub-urban markets
I've wondered how this system knows when to slow for curves. E.g. cruse is set for 60 mph and the upcoming curve is sign rated at 35. How would it know to slow down or how much to slow down? How would it know to return to 60? Hope it's not relying on some type of map. Maybe it's beyond AP. Seems both pretty simple (the task) and incredibly complex (to do).
EVNow
Well-Known Member
I think it currently relies on the map - and possibly the sharpness of the curve too. They probably have some in built logic about the max speed for a given curvature. Returning to 60 mph would require confirmation of the speed limit using the map.
BTW, map is always required - otherwise there is no sure way for the car to determine the speed. Even if it can read the posted signs, the speed postings are sporadic/random.
Why would they deploy faster? It is still a development system. They are developing it in dozens of locations with different characteristics like snow. But they are not done yet.
With Tesla we keep talking about the future march of 9s. I’d wager Waymo is already in the march of 9s. But again, they are not done yet. Once they are ”done”, I would expect wider deployment to begin.
diplomat33
Average guy who loves autonomous vehicles
You are correct but only if Waymo actually has a "general solution", just temporarily geofenced while they improve the reliability. In that case, yes, once Waymo improves the reliability, then they can scale up pretty fast and do a wider deployment. But if Waymo only has a "specialized solution" that relies on special maps to work, then it won't be so easy to generalize or scale up to L5. If that is the case, then even when Waymo finishes the "march of 9's" they will still be stuck at L4.
As @EVNow pointed out in his post, Waymo solved parking lot navigation by annotating the heck out of a highly accurate HD map to tell the car exactly what to do. Yes, it is L4 and it works really great but it will only work in geofenced areas that have the special HD maps that have been annotated properly. But it means Waymo is indeed using a specialized solution for doing FSD, that cannot be widely deployed quickly.
So I am suggesting that Waymo's slow deployment is precisely because they don't have a general solution. They have an excellent L4 system (the best self-driving right now actually) but one that relies on special maps to work. So in order to deploy their fleet, they have to premap the area and do a lot of work annotating the map first. This takes time. Hence, why the deployment is slow. If I am right, then getting to L5 will be very difficult for Waymo because it will be too time consuming to premap the entire US. Not going to happen in a timely manner. Which is why Waymo will be much better off not even trying to do L5, and just deploying to select metro areas that have the most potential for profitability. They won't get to L5 but it won't matter because they will have the best robotaxis operating in most major US cities.
Indeed it comes down to that — or if Waymo’s solution allows faster adding of generalization than the system others like Tesla would use.
That was the alternative theory I was offering for the sake of keeping the discussion diverse.
Also, some level of mapping may not be a hinderance with the likes of Google, because they can and have mapped effectively entire continents with an ever increasing suite of positioning and mapping technology.
Interesting to see over time.
EVNow
Well-Known Member
Well, that kind of mapping is very different from what Waymo needs. With Lidar and highly annotated. How is that going to work in a large continent where things are constantly changing.
I believe Google already maps with Lidar. However my point is more general than that: Whatever Waymo needs for mapping, I can see Google integrating into their processes down the line as things mature — and Google is very good at mapping at scale.
As for how that will work when things are constantly changing? I continue to believe it would be a mistake to assume Waymo requires all that annotation and HD mapping to work, especially as things mature. It may very well be it is or will be just one redundant data point amongst many others.
This is one theory anyway.
EVNow
Well-Known Member
Possible- but looks like they are using the easier path first, leaving out the more difficult path for later.
So, instead of driving using vision they use hd maps for localizing. Yes, this could be a fall back mechanism later, but they have to first develop drivable path without localization first, which puts them behind Tesla.
Frankly I don’t understand Waymo strategy. They have the best AI engineers and platform, most money of all competitors and yet very reluctant to run robotaxis without safety drivers or publish a plan For scaling.
That is all I really wanted to point out anyway.
I fully agree we don’t know where things are headed. Just putting some less talked about theories and points on the table.
