AlanSubie4Life
Efficiency Obsessed Member
Good. Tesla still can take the crown, later this year, when FSD is done and in wide release. Remember, AI Day is coming.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autonomous Car Progress
- Thread starter stopcrazypp
- Start date
-
- Tags
- Autonomous Vehicles
diplomat33
Average guy who loves autonomous vehicles
Suuurrre... Hope springs eternal as they say. I guess you believe FSD will be solved this year.
diplomat33
Average guy who loves autonomous vehicles
We don't know. Mobileye has not provided any more details.
I do hope that Mobileye releases some videos from the trip. I don't think they will release the entire unedited trip since that would be 40 hours long. But if they released some unedited videos that show some of the most interesting cases, I think that would still be informative. Mobileye has released unedited videos before of drives so I think there is a good chance we will get some videos from the cross country demo.
diplomat33
Average guy who loves autonomous vehicles
Watch a timelapse of Mobileye mapping large parts of Europe and US in just 12 months. It looks like they've mapped virtually every road in EU and US now.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Shashua touts the recent cross country demo and teases that Mobileye will be bringing SuperVision to more vehicles soon:
IMO, the fact that they have mapped virtually all roads in the US and EU bodes well for Mobileye expanding SuperVision to the US and EU soon.
IMO, the fact that they have mapped virtually all roads in the US and EU bodes well for Mobileye expanding SuperVision to the US and EU soon.
Yet some still believe maps isn’t scalable. Apparently ~1.5 million cars contribute to the mapping. As more cars enter the mobileye network, the coverage becomes even better. Thanks for sharing.
diplomat33
Average guy who loves autonomous vehicles
I wish Tesla would do this. Tesla could build reliable maps and update them quickly with the large number of Teslas on roads. I think accurate HD maps would really help FSD beta. I've had issues with FSD Beta that I think would be fixed if it had better maps. For example, taking sharp turns too fast, not recognizing a 15 mph speed limit, not moving over into a turn only lane when making a turn or moving over too late.
Also, I asked Mobileye how they store the AV maps, whether in the cloud or in the car offline? CTO Shai Shalev-Shwartz himself replied:
Last edited:
It may well be that the enhanced map data is only needed for certain areas that are hard for FSD to perceive correctly without more hints as determined dynamically based on actual driving and network feedback to Tesla. Much of the time FSD can perceive and drive along just fine without needing significant new map data. The added data can probably be sparse and space efficient.
Something like that should be easily doable with Tesla’s network of connected cars, assuming they hire the right expertise.
I figured caches of at least the destination route and surrounding area would be onboard the vehicle. I would rather they cached at least the entire state maps.
Given that Tesla's aim is for cars to be self-driving without the need to map out every pebble in the road, I find it curious that they would even consider making their own maps. This would require a lot of effort that they can licence from someone like google, who specializes in it. I recall reading once that google had 5000 people working map updates. Why duplicate that effort? I doubt that they would ever save money at it.
They would be better off having the cars report map issues (automatically, or by user input) to feed back to their mapping and/or routing supplier with a more streamlined mechanism to get continuous map updates pushed out to the cars.
bradtem
Robocar consultant
You could say, "Looks like the car figures out the map of this location on the fly without additional help, no need to store and remember that map."
It would save you a bit of storage, perhaps a bit of bandwidth.
But that's hardly a priority. Those are not expensive things and they keep getting cheaper. What's expensive is mistakes. You can't be sure the car will always figure out the map on the fly. You can keep comparing the map it makes to the reality to see how it's doing -- but that means you are storing the reality so you are not saving that storage and bandwidth.
And of course the car should be comparing its map to the prior map to see if there are any differences, and then have systems to handle what those differences are, which depending on what they are may be to just accept the new interpretation, rely on the old one, or in rare cases pull over or request manual assist in this section.
The car with no map is not even aware it's interpreting it differently from what the last car did (or what it, itself did last time it was there.) It just takes what it sees now.
That might work most of the time. But the stored map of course has so many advantages. It was created by multiple passes by multiple cars, each one reinforcing the quality of the map. Every object was seen both from a distance but also up close, and from multiple angles. Each image was processed in non-real time, and many were sent to the cloud where they could be processed at leisure by giant supercomputers, rather than within 100ms by HW3. And, for a subset of cases, the results also may have been checked by a human being if flagged as troublesome, up to whatever budget of human checking you care to spend, from very low to, at least in the early days, high.
The decision to just forget everything learned about any piece of road doesn't make much sense. Why forget it? To save a tiny amount of money?
Many say, "but yes, now my car can drive a brand new road it's never seen." Maybe it can. But that's not that exciting. What road in the USA didn't have a Tesla drive it today, at least if the road was not out in the charging deserts where no EVs go? What urban street or common rural road didn't have 100 Teslas drive it?
This is the dawn of self-driving. The time to make it work as well as you can, not the time to cut corners to save some storage and bandwidth.
From the AI day presentation, instead of making hard coded maps based on GPS coordinates, Tesla is saving the features of the road as weights for the NN, so that when it sees any road that looks similar, it can predict the layout of the road. Theoretically that allows it to recognize and predict road structures on roads that were never mapped before. If there are enough weights, it can theoretically predict every single road in the world.
As for actual maps, Tesla is just buying data from a map service (mapbox or tomtom). They dabbled in creating their own, but even though Tesla's fleet is growing, it probably can't match the coverage of the fleets of the map services.
I would note however, the map being discussed is for hands free L2, not for their L4 solution. So far the L4s (not just Mobileye, but other companies too) seem to still require some time actually driving at a given location with the exact end vehicles (for even better equipped mapping vehicles), not just general data from regular cars with cameras.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Mobileye CTO Shai Shalev-Shwartz congratulates Tesla on FSD but shares some key differences between SuperVision and FSD:
Last edited:
You are slightly misunderstanding what they were talking about. I'm assuming you are referring to the part where they showed multiple cars contributing to what appears to be a map of an intersection. Those are not weights but auto-labeled data of the road generated from multiple clips and varying viewpoints on the same intersection, projected in vector space. The labeled data is used to train their NN, ultimately a trained NN model is just weights used at runtime for inference, this applies to just about every major player working in this field currently.
I'm aware Tesla sources maps from 3rd party vendors. I'm just saying it is something they can do themselves with the right expertise and of course it would take time to reach the same level as Mobileye REM Map. There is something they also said at AI day that leads me to believe it is something they might be working on because those vector space auto-labeled road data are pretty much maps. How detailed the label is, is unknown.
Mobileye REM maps are used for both L2 and L4. That's one thing they've been very keen on using as a selling point. It makes sense to use Lidar for mapping as it is very accurate recreation of the world in 3D that is why majority use it as part of their sensor suit and dataset.
diplomat33
Average guy who loves autonomous vehicles
This is incorrect. Mobileye builds their AV maps using cameras only, crowdsourcing data from their large L2 fleet. And the same maps are used for both L2 and L4. Mobileye's L4 is built upon the L2 system. The L4 system uses the same maps, the same vision stack and the same RSS as L2. It just has the extra computing power and the added radar-lidar stack as you can see from the graphic below. Remember that Mobileye defines L4 as just a more reliable L2 that does not need human supervision.
Last edited:
Let's not go crazy here. Accurate extremely-low-definition maps would help FSD Beta.
Maybe I'm wording it poorly, but my point is more that "Supervision" L2 solution can function solely on the maps they are making right now and they can plop a car with that basically anywhere and have that function fine.This is incorrect. Mobileye builds their AV maps using cameras only, crowdsourcing data from their large L2 fleet. And the same maps are used for both L2 and L4. Mobileye's L4 is built upon the L2 system. The L4 system uses the same maps, the same vision stack and the same RSS as L2. It just has the extra computing power and the added radar-lidar stack as you can see from the graphic below. Remember that Mobileye defines L4 as just a more reliable L2 that does not need human supervision.
I don't believe this is the case for their L4 solution, which is still restricted to very heavily geofenced areas and needs a certain amount of time of driving in that area before they work well (at least that has been the case for the cars they demoed so far).
So even though both of them will utilize these maps (just like how they also use data from basic navigation maps also), the L4 solution doesn't work as reliably just on that map.
I guess we'll see. If the L4 solution really needs no additional data beyond the same maps the L2 one is using, the Mobileye solution should be able to launch an L4 solution that has essentially no geofencing as long as it is in that map (which according to Mobileye seems like it should practically cover every public road in the US (beyond any legal barriers like individual state laws). That has not been the case so far for any of the L4 solutions out there. They test in very heavily geofenced areas.
Nope, not talking about that part talking about that, talking specifically about this part (from an earlier post):You are slightly misunderstanding what they were talking about. I'm assuming you are referring to the part where they showed multiple cars contributing to what appears to be a map of an intersection. Those are not weights but auto-labeled data of the road generated from multiple clips and varying viewpoints on the same intersection, projected in vector space. The labeled data is used to train their NN, ultimately a trained NN model is just weights used at runtime for inference, this applies to just about every major player working in this field currently.
At around 2:45:50 Andrej mentions (bold emphasis mine):
"actually in the limit you can imagine the neural net has enough parameters to potentially remember earth so in the limit it could actually give you the correct answer and it's kind of like an hd map back baked into the weights of the neural net".
He touched on a similar thing earlier in the presentation when talking about RNNs at around 1:09:15:
"but you can imagine there could be multiple trips through here and basically number of cars a number of clips could be collaborating to build this map basically and effectively an hd map except it's not in a space of explicit items it's in a space of features of a recurrent neural network".
FSD Beta Videos (and questions for FSD Beta drivers)
It's not a "map" based on GPS coordinates as we traditionally know maps, but rather it's based on the weights of the NN.
The labeled data as per the presentation is only used for training and is gathered only from a small subset of video data the fleet may see (basically only clips that they trigger to capture as interesting). I see nothing in the presentation that suggests they are going back to making their own maps (meaning clips are captured from all roads to do the auto labeling). If that was what they were doing, from the screen shot you posted, they shouldn't be sending a clip to the offline NN. Instead, there should be an online NN that does the autolabeling, and they send the finished data back to the mothership (not the clip itself). Otherwise the data bandwidth required is too much.
See my other post on this subject. L2 and L4 both use REM maps (as well as dumb navigation maps), but that does not necessarily mean L4 can operate without additional data/training beyond that. Put another way, if there was no difference, the supervision demo that was done should have been able to done as L4 with zero interventions if the REM maps was all that was required.
Last edited:
Similar threads
- Article
