Kant.Ing
Member
Once a while similar posts like this will show up. The hardest part of Tesla's AP/FSD solution is that it depends on Tesla owners to find out its limitations.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Challenges - Things that make it so hard to do...
- Thread starter aSlowOldGUy
- Start date
Daniel in SD
(supervised)
tljdesign
Member
I think the poster was trying to show another example of why FSD is unlikely to happen any time soon.
Daniel in SD
(supervised)
But FSD is totally different code being worked on by a totally different team, right?
The reason it can't see the jersey barrier is that the neural net isn't running fast enough and HW3 will solve that I do feel that if they can't get the vision system to see barriers like that then there isn't much hope for a camera only approach.
I will not defend Tesla's FSD because we have seen nothing but some hardware specs.
But as a software engineer I thinkany of the things written on this thread are not really that impossible.
Image recognition is something already done by many companies like Google for many of this services. We just need to keep evolving and improving the technology so that we can put most of it in a car. Maybe at the end we will need 100 of the new CPUS from HW3 and a 5TB SSD, but that is just a matter of time. And maybe it can be done like the HW3 update for HW2.5: remove motherboard, install new motherboard board. Done.
Anything will be recognizable by neural networks or by whatever the new technology we come up with is. So plot holes, bumps, a sudden person dress in a particular set of colors doing gestures that are well known etc.
Of course the biggest problem is going to be blind and dirty cameras. This can be fixed by adding more specialized cameras and some kind of "auto cleaning system".
The only obstacles I can think about that can be a challenge are small things very close to the car, where cameras have pore visibility, like a kid coming out of nowhere or a dog. This is going to be harder to handle for Tesla because for example there are no cameras on the nose.
Then, you can enhance this with information coming from GPS and from a database like Waze: traffic, construction work, crash etc. You can know in advance where a bus stop is located by GPS, and with a "hive mind" every car can communicate whatever it sees: a policeman here, a tree blocking the road there, etc.
And then there is the "what to do". I like to think of it as a video game like Forza Motorsports. You can use a hard coded NPC algorithm to specify what to do, or you can train an IA to drive. Those things have been imemented since video games exists. Of course in this case you have to refine it a lot because you you are actually handling human lives, but nothing really new, just bigger. And again maybe we need 4 times more processesing power than right now, but it's just a matter of time to get there. And again maybe it can be done like the HW3 update for HW2.5: remove motherboard, install new motherboard board. Done.
Finally, Tesla's FSD is going to be under development / beta for a very long time, the same way autopilot keeps getting better and better over the years. But I honestly think we just need to keep evving what we have to achieve a sweet spot for FSD with the current hardware installed.
TL;DR: a computer with 8 eyes and constant connection to a humongous database of information that is constantly updated by a huge number of cars will drive better than the avarage driver. It's a matter of proving power and a big enough IA. I don't know when tho.
But as a software engineer I thinkany of the things written on this thread are not really that impossible.
Image recognition is something already done by many companies like Google for many of this services. We just need to keep evolving and improving the technology so that we can put most of it in a car. Maybe at the end we will need 100 of the new CPUS from HW3 and a 5TB SSD, but that is just a matter of time. And maybe it can be done like the HW3 update for HW2.5: remove motherboard, install new motherboard board. Done.
Anything will be recognizable by neural networks or by whatever the new technology we come up with is. So plot holes, bumps, a sudden person dress in a particular set of colors doing gestures that are well known etc.
Of course the biggest problem is going to be blind and dirty cameras. This can be fixed by adding more specialized cameras and some kind of "auto cleaning system".
The only obstacles I can think about that can be a challenge are small things very close to the car, where cameras have pore visibility, like a kid coming out of nowhere or a dog. This is going to be harder to handle for Tesla because for example there are no cameras on the nose.
Then, you can enhance this with information coming from GPS and from a database like Waze: traffic, construction work, crash etc. You can know in advance where a bus stop is located by GPS, and with a "hive mind" every car can communicate whatever it sees: a policeman here, a tree blocking the road there, etc.
And then there is the "what to do". I like to think of it as a video game like Forza Motorsports. You can use a hard coded NPC algorithm to specify what to do, or you can train an IA to drive. Those things have been imemented since video games exists. Of course in this case you have to refine it a lot because you you are actually handling human lives, but nothing really new, just bigger. And again maybe we need 4 times more processesing power than right now, but it's just a matter of time to get there. And again maybe it can be done like the HW3 update for HW2.5: remove motherboard, install new motherboard board. Done.
Finally, Tesla's FSD is going to be under development / beta for a very long time, the same way autopilot keeps getting better and better over the years. But I honestly think we just need to keep evving what we have to achieve a sweet spot for FSD with the current hardware installed.
TL;DR: a computer with 8 eyes and constant connection to a humongous database of information that is constantly updated by a huge number of cars will drive better than the avarage driver. It's a matter of proving power and a big enough IA. I don't know when tho.
Why testing self-driving cars in SF is challenging but necessary has a few.
There are some more on pages 36 to 39 of Waymo's safety report at https://storage.googleapis.com/sdc-prod/v1/safety-report/Safety Report 2018.pdf.
In some places of the world, they don't even follow road rules or the lanes painted on the ground.
From talking to some folks who visited India on business, they mentioned that the way a vehicle's horn is used can mean any number of things. Try checking out a few videos at https://www.youtube.com/results?search_query=india+city+traffic.
This thread should probably be moved to the Autonomous Vehicles as it's not Model 3 specific.
Similar threads
