If I remember correctly, it took at least a year for AP2 to catch up and pass AP1 — and arguably it still hasn’t in a few ways like reading speed limit signs and lane merges. While I imagine much of the AP1 to AP2 situation was because Tesla was moving from an outside vendor to internal development, I would think swapping out the current software for a total rewrite will cause similar issues. I don’t recall Musk addressing this question. Your thoughts?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Will the autopilot rewrite be a replay of AP1 to AP2?
- Thread starter azred
- Start date
diplomat33
Average guy who loves autonomous vehicles
I don't think the two situations are quite the same. AP2 was completely different hardware than AP1. In this case, the hardware is the same, it's just the software that is changing. So I do expect things to be a bit easier this time. However, that does not mean that it will completely problem free. The rewrite is significant so it could take time to debug.
n.one.one
Info w/o Drama
I'm not a subject matter expert but as I remember (and in my words) they may be running the new software in the background while running the current software in a portion of the chip. IF...IF that is something close to correct the transition should be easier. I hope others who are more knowledgeable will expand the discussion.
I'll look back at the autonomy day video because that may be where I saw this.
I'll look back at the autonomy day video because that may be where I saw this.
We are finally seeing some real progress on Autopilot for the streets. Hopefully that progress won’t be substantially reversed or delayed. (It has been nearly four years since I purchased my AP2.0 S and I have become a bit cynical.
)
diplomat33
Average guy who loves autonomous vehicles
We are finally seeing some real progress on Autopilot for the streets. Hopefully that progress won’t be substantially reversed or delayed. (It has been nearly four years since I purchased my AP2.0 S and I have become a bit cynical.
I am cynical too. Tesla is a very long way from L5.
powertoold
Active Member
I don't think the rewrite will be like AP1 -> AP2.
Tesla has gone a long way in unit testing and creating their software stack. They'll definitely be able to validate the rewrite before doing a wide release. I have little doubt it'll be better than the current software once released (probably for HW3+ only).
Tesla has gone a long way in unit testing and creating their software stack. They'll definitely be able to validate the rewrite before doing a wide release. I have little doubt it'll be better than the current software once released (probably for HW3+ only).
Especially with the recently introduced Autopilot behaviors related to traffic lights and stop signs, these were implemented in the context knowing a rewrite was in progress, so making sure tests are compatible with the existing main-camera code and birds-eye-view rewrite was probably a requirement.
One potential downside is Autopilot team needed to spend time implementing the traffic light behavior on existing codebase that will become unused a few months later thus looking like wasted engineering resources that could have sped up rewrite development. But then again, if one thinks of the old code as gathering data for the new code as well as allowing Tesla to recognize some deferred revenue, this temporary scaffolding might have accelerated Tesla overall.
powertoold
Active Member
It's all part of the iterative development process. It's good to know that they're bold enough to start from scratch again. Most companies will just keep going with what they have.
I still feel like lidar has an advantage for fully 3D labeling. However, the companies using lidar right now don't have access to the data. It'll be interesting to see who's really ahead next year.
DDotJ
Member
I don't think so, IIRC Tesla's hand was forced when they went from AP1 to AP2 because of a spat with MobilEye. I remember seeing leaked schematics with a dual camera setup on the front, with rumors that Tesla would develop Tesla Vision in the background while running EyeQ chips.
When Tesla and MobilEye broke up their partnership, Tesla was forced to move AP development in house. They couldn't use MobileEye tech anymore so they scrambled to recreate the AP software and it took them a long time to reach feature parity. There's a good TMC thread here about the timeline of that part of Tesla history. There was no AP functionality whatsoever for a few months and even when Autopilot was released it was very primitive and terrible compared to AP1.
This situation is completely different. Tesla isn't being forced to roll out their new AP software stack. They can release it when they feel that they are ready to release without the threat of selling vehicles without AP functionality. Maybe some areas might see some slight regressions that might be corner cases but I expect it to be a lot smoother of a transition.
When Tesla and MobilEye broke up their partnership, Tesla was forced to move AP development in house. They couldn't use MobileEye tech anymore so they scrambled to recreate the AP software and it took them a long time to reach feature parity. There's a good TMC thread here about the timeline of that part of Tesla history. There was no AP functionality whatsoever for a few months and even when Autopilot was released it was very primitive and terrible compared to AP1.
This situation is completely different. Tesla isn't being forced to roll out their new AP software stack. They can release it when they feel that they are ready to release without the threat of selling vehicles without AP functionality. Maybe some areas might see some slight regressions that might be corner cases but I expect it to be a lot smoother of a transition.
Tesla have a massive tool kit and training library which they wouldn't have had back in the AP1>AP2 days.
It was interesting to see Elon's tweet that they were moving from 2D to 3D. Would have hoped they went straight to 4D (3D+time) but I guess that'll be the next re-write...
It was interesting to see Elon's tweet that they were moving from 2D to 3D. Would have hoped they went straight to 4D (3D+time) but I guess that'll be the next re-write...
Tesla have a massive tool kit and training library which they wouldn't have had back in the AP1>AP2 days.
It was interesting to see Elon's tweet that they were moving from 2D to 3D. Would have hoped they went straight to 4D (3D+time) but I guess that'll be the next re-write...
3D uses time also. A group of time separated frames are used to generate the scene. Then the labels are propagated through the time slices to each frame.
I am glad to see optimism in most of the posts in this thread. My skepticism is rooted in owning a HW2 S for 3 1/2+ years and losing count of all of the B.S timelines put out by Tesla about FSD.
Did they underestimate the challenge of new HW2 and the impending EyeQ software divorce? I doubt it. I think it is more likely it was just Musk’s philosophy of throwing out unrealistic timelines (perhaps) to enthuse prospective buyers or (more likely) to keep pressure on his workforce.
Will this rewrite be implemented in 2 to 4 months — or even 6 months? The past says there is almost no chance but his programmers certainly will feel pressure to deliver. I guess that is good but I think that could be accomplished internally without building up customer hopes.
Did they underestimate the challenge of new HW2 and the impending EyeQ software divorce? I doubt it. I think it is more likely it was just Musk’s philosophy of throwing out unrealistic timelines (perhaps) to enthuse prospective buyers or (more likely) to keep pressure on his workforce.
Will this rewrite be implemented in 2 to 4 months — or even 6 months? The past says there is almost no chance but his programmers certainly will feel pressure to deliver. I guess that is good but I think that could be accomplished internally without building up customer hopes.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
I think it is more than unrealistic timelines just to motivate the workforce. In 2015, Elon said that FSD was "like a solved problem". He is absolutely convinced that solving FSD is just a matter of "solving vision" by doing the machine learning with enough big data. Elon also said that "parking lot driving" and "highway driving" would be easy to solve, that only city driving would be somewhat challenging. So he saw solving FSD as a relatively easy and straightforward process. I think this supreme confidence is a big reason for the unrealistic timelines. He fundamentally underestimated what it would take to solve FSD.
That's a big reason why Tesla is doing this rewrite. They realize 3D labeling and pseudo-lidar are critical components of camera vision if you want to do FSD.
Based on past results, I think it is likely that Tesla will miss the 2-4 months deadline. But I am optimistic that the rewrite will produce very positive improvements to FSD when it is ready.
Last edited:
i feel the same man. My ap2 lease that i paid for FSD on goes back in 1 month. They upgraded me to ap3....now im upset that i cant afford the buyout payment to see what the future of autopilot is like...hilarious how i went through all the crap for them to finally get to some real functionality and now my car goes bye bye.
i may just get the cheapest model 3 available to still use AP on long trips...idk.
very impressed with the newest stop light detection it works extremely well.
Similar threads
