Anyone that posts an anti-Tesla post in any way and is new here is an instant troll. That's the TMC group think. There is one mod here that literally engages in censorship at levels that are a joke.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Will Tesla be able to complete the FSD?
- Thread starter void walker
- Start date
if your watching the 737 max debacle,
runaway trim + lack of training = BITE!
so what's old is new again!
Knightshade
Well-Known Member
Suggest the HW can eventually do it? Absolutely.
Promise the actual named feature sold for $, with a specific (shorter) list of abilities, will do it? Not so much.
Though as I say the pre-3/19 buyers have a lot better case to make about them having paid for something nearer Elons original statements than the post 3/19 buyers do. Neither was ever promised literal SAE L5 for their money though.
CertLive
Member
They were told their hardware is level 5 ready? Seems legitimate to me. Unless there is a counter disclaimer to suggest this is not the case you would think the CEO is correct and using it as a selling point for the cars to this day.
Knightshade
Well-Known Member
My human body hardware is "ready" to paint a masterpiece painting.
I am utterly lacking in the software to do so however, so expecting me to paint one isn't going to get you far.
The fact it has all the HW? Sure that's a selling point.
The SW isn't there though. And indeed the SW package they ARE selling only brings a limited feature set- even more limited since they narrowed the specific promises in March 2019.
HW capability != SW capability.
And "expected" != promised while we're at it.
See again how he "expected" to drive coast to coast with no driver interventions by 2017- but never actually promised that specific thing to anybody buying FSD software.
At this point, people have bought "Full Self Driving", not just the current set of features. As long as Tesla continues to develop FSD on the Model 3, I think it reasonable to expect them to provide that development to those who (pre)bought that functionality. Can they come up with a new feature... "Really Full Self Driving, Level 5"?... I don't think that would fly.
Knightshade
Well-Known Member
Post 3/19 buyers have bought, specifically, the current set of features....PLUS... a promise of 2 other specific features.
- Recognize and respond to traffic lights and stop signs
- Automatic driving on city streets
That's it.
(and with relatively little detail on the second, other than the admission it'll be level 2 at least out of the gate)
Pre 3/19 buyers on the other hand bought.... zero current features- there aren't any above the EAP they had to buy before buying FSD.... but a promise for a lot broader set of future ones.
My most optimistic read of them changing the promised future FSD features is to minimize the # of folks they will owe refunds to if they can't get to what FSD originally promised.
As long as they can get stoplight/stopsign action and city driving at at least level 2- the new FSD folks have had what they bought delivered, technically.
If they can give em more- great. I agree they likely will. But at least it'll only be the pre-3/19 buyers they owe money if they can't get much further than that.
I don't think they made the change so they can charge the "new FSD" folks more later for "extra FSD" or something.
That said- I DO think it's possible they'll find they simply can't do L4 or L5 on the current sensor suite. Then it's decision time for Tesla and they have a number of options.... Free HW upgrades (beyond the computer) for pre-3/19 buyers and nothing for post 3/19? or free for both? or upgrades for nobody because it's too expensive, but only the pre-3/19 folks get refunds? Hopefully not a bridge they have to figure out how to cross.
smatthew
Active Member
I hope this makes sense. From what i understand ap is going to need a hefty update first to achieve the 'FSD' goal.
AP is optimised for HW2 and pumps out 99fps uses 90% of HW2 compute and the cameras update the neural net at different rates. Updates are every frame for main camera and every 4th frame for all other cameras. It requires neural net to stitch and post process info from both main and other cameras to achieve predictions and accuracy is low (cameras aren't in sync due to HW2 compute budget).
It also hinders accurate updates at junctions and traffic lights due to neural net predicting differences in frame information from main vs other cameras.
An example of 36fps sync'd across all cameras will use 15% of HW3 compute budget so there's a lot left in HW3 before HW4 is required. Doing this will create a fisheye 'all round ap' camera and fix the above issues. It'll also sort windy / tight roads as camera's are 'sync'd' and 'fov' (and more importantly camera update info) is accurate. Neural net no longer has to stitch and post process and can make predictions based on all cameras at the same time. But there's a catch...
The compute budget is also constrained by the design of the neural net which is designed to require these separate (unsync'd) outputs from main and 'other cameras' to be stitched together which affects the path prediction etc.
After AP gets rewritten into this 'all sync'd camera ap beast' the neural net will need a rewrite to take advantage and most likely drop the compute requirement of the neural net as it no longer has to stitch / post process inbound data etc. Much simpler, no more stitching for predictions and the improvements will allow an increase in training data harvesting for Tesla.
This probably makes some sense of why city street driving, traffic light and road sign info is coming next...
Long way to FSD or whatever Tesla's interpretation of FSD is though...
edit: found part of it - it'll be a bit heavy if this isn't your thing but anyway Multi-Task Learning in the Wilderness
AP is optimised for HW2 and pumps out 99fps uses 90% of HW2 compute and the cameras update the neural net at different rates. Updates are every frame for main camera and every 4th frame for all other cameras. It requires neural net to stitch and post process info from both main and other cameras to achieve predictions and accuracy is low (cameras aren't in sync due to HW2 compute budget).
It also hinders accurate updates at junctions and traffic lights due to neural net predicting differences in frame information from main vs other cameras.
An example of 36fps sync'd across all cameras will use 15% of HW3 compute budget so there's a lot left in HW3 before HW4 is required. Doing this will create a fisheye 'all round ap' camera and fix the above issues. It'll also sort windy / tight roads as camera's are 'sync'd' and 'fov' (and more importantly camera update info) is accurate. Neural net no longer has to stitch and post process and can make predictions based on all cameras at the same time. But there's a catch...
The compute budget is also constrained by the design of the neural net which is designed to require these separate (unsync'd) outputs from main and 'other cameras' to be stitched together which affects the path prediction etc.
After AP gets rewritten into this 'all sync'd camera ap beast' the neural net will need a rewrite to take advantage and most likely drop the compute requirement of the neural net as it no longer has to stitch / post process inbound data etc. Much simpler, no more stitching for predictions and the improvements will allow an increase in training data harvesting for Tesla.
This probably makes some sense of why city street driving, traffic light and road sign info is coming next...
Long way to FSD or whatever Tesla's interpretation of FSD is though...
edit: found part of it - it'll be a bit heavy if this isn't your thing but anyway Multi-Task Learning in the Wilderness
Last edited:
OK, I do agree with the sentiment, but I'm pretty sure our lack of update over the last 2 months means that 'FSD' will be dropped from a helicopter shortly. It won't actually be FSD but it will be the seed, so I guess it just depends how you think of it. Then the HW3 rush and entitlement comes. Then HW4 and HW5 comes...then FSD comes
vjason
Member
Short of a miracle, I don’t see how you’ll ever have an empty driver seat with any Tesla currently on the road. I’m talking mainly about robotaxis here, as I bought FSD expecting that I’ll always have my hand the wheel. I mainly just wanted to make sure I got the computer upgrade (which is scheduled next Monday) and an AP that can turn for me.
The current sensor suite simply can’t handle enough variations in the weather, big examples being snow and direct sunlight. Raleigh has its yearly snowstorm last week, and AP was freaking out every 3 minutes (cameras and other sensors) during my short drive home. Add to that the all those times when the sun is shining directly into a camera, which it also does not like.
The current sensor suite simply can’t handle enough variations in the weather, big examples being snow and direct sunlight. Raleigh has its yearly snowstorm last week, and AP was freaking out every 3 minutes (cameras and other sensors) during my short drive home. Add to that the all those times when the sun is shining directly into a camera, which it also does not like.
I wish the car could just take selfies from all the cameras, that would be useful. Perhaps make a music video after each drive and send it to my phone. In the meantime I'll hold onto my money and the steering wheel which will be needed regardless. The best single feature Tesla could offer is a legal disclaimer on the screen that waives all liability and takes the nudges off AP. That would sell very well because what good is auto lane changing if I need to keep my hand on the wheel!
Been reading this thread and mostly agree that full FSD with the current sensor suite is a pretty heavy lift. But then, I consider that I have even fewer sensors than the Model 3 and I have been driving safely for many years and in all sorts of conditions. Just says that two eyes, coupled with the right processing can be adequate, although that processing is likely beyond the current state of computer vision technology.
Similar threads
