Packed snow and loose snow look very different to RADAR, and probably to ultrasonics as well.
Edit: I missed the context where the original poster was talking about the visual neural net. Yeah, that probably won't work.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Coast to coast drive happening this year for all FSD Teslas!
- Thread starter Daniel in SD
- Start date
Packed snow and loose snow look very different to RADAR, and probably to ultrasonics as well.
Edit: I missed the context where the original poster was talking about the visual neural net. Yeah, that probably won't work.
While you are right that my Cindy Crawford has a mole on her face.... that doesn’t mean your riding with Beyoncé
EVNow
Well-Known Member
No - it needs to be unit tested and generally working.
If they are doing some kind of agile development, the user story should have acceptance criteria in the form of test cases and the code should pass those cases.
Then the code gets released and they find all the edge cases it doesn't work under and fix those ("march of 9s").
EVNow
Well-Known Member
Did the car request the safety driver to take over or did they take over because the car was taking too long ? If the former then, it is L3, otherwise Waymo is L4. BTW, Lex Fridman refers to Waymo as L4.
BTW, the level definitions are conceptual. It has no concept of error rates, so we can never prove that a particular AV is at L3, L4 or L5 etc.
This is not a showstopper. I can drive on snow quite well. And I only have my eyes to guide me on where to drive or not. What I see combined with what I know after years of winter driving. This is comparable to Teslas driving with cameras and a knowledgeable autopilot system.
I know. I just mean the feature list wasn't that long and where I'm from, the GPS speed database isn't very good and hasn't changed in the 2+ years I have my AP2 now. Also in my area we have a bunch of variable speed highways where a fixed database entry doesn't work. So it is a pretty relevant feature to me, and it's still missing. Hence I speak up whenever someone claims they achieved feature parity, because they didn't.
Yep, and that's why there are differing definitions of "feature complete". And as you point out it depends on development methodology as well. So in the end we really don't know what it means to be feature complete for FSD.
EVNow
Well-Known Member
But we can make a guess from what EM has said in various interviews (esp. the one with ARK). He talks about it in the way I've described it (i.e. it will generally work but will not take care of lots of edge conditions).
Ofcourse the big question is will they really hit FC described this way by end of the year or declare FC whatever the quality status is by end of the year.
While FC is an important internal milestone, to me the big leap is when they release NOA on city streets to the fleet. We can look at FC as
- NOA on freeways (done)
- NOA on city streets (WIP)
And then they work on various "edge" cases from most common cases like snow covered roads, construction zones to least common cases like driving in the extremely narrow roads on the Himalayas in their march of 9s.
dhanson865
Well-Known Member
I don't agree. If a collision occurs and there is debris in the road you might have to drive a mile through the traffic jam and then let the computer take back over. It could be the car is faux sentient and perfectly capable but there is a law that says you have to take over in X situations.
If the car can drive 99.99% of the time in sunny weather I'm ready to call it FSD. Imagine this conversation between car and driver
* driver - I'm going to work, you know the way. I feel like a nap wake me up if anything unusual happens. Oh and use Chill mode.
* car - very well sir, traffic data suggests you might have to take over for a half mile near a collision that has occured on our normal route. There isn't a faster route that bypasses the site of the incident. I'll wake you up when we get within a mile or if traffic drops below normal highway speeds.
* driver goes to sleep
* car wakes up driver after about 15 minutes have passed and one of the trigger conditions is met
* driver - disable FSD, go to TACC only ( and proceeds to drive through the unusual traffic pattern)
* car - sir I see we have passed the site of the incident and speeds are back to normal. Would you like me to resume driving to your destination
* driver - yes, enable FSD, I'll read a book until we get there. Keep Chill on.
* car - sir, you have arrived at your destination. Should I find a charger or just park in the shade?
and so on.
If the car can drive on it's own and even follow simple commands and there is basically no risk I'd call it fully self driving even if a law requires me to be in the drivers seat at that point.
I imagine the law requiring that so you don't have a traffic jam full of cars that had to stop to let the driver get out of the back seat. Just so they could get up front and drive the 1 mile before going back to sleep.
Totally fine if you sleep, read, type, whatever but they don't want a ton of cars pulling over disrupting traffic further to get a human in the drivers seat for the .001% cases. The human is there as a backup driver, not as the primary.
Last edited:
You are right though... a mole is a mole, but it matters especially when locked to a lower speed
It will happen, but it will take time.... mainly to get the facial features right on the exmachina Gal Gadot.... as for the snow, yeah no one can work miraclesIn the future I might be dating Gal Gadot
diplomat33
Average guy who loves autonomous vehicles
I think some folks may be overthinking what "feature complete" means. The problem is that some folks are coming up with their own list of features that they think FSD needs in order to be feature complete. But we don't need to do that and we don't need to play guessing games either. Tesla has given us a set of features on the website. I think we can reasonably assume that those features listed on the website are what Tesla means by "feature complete".
First, we have the list of FSD features on the order page:
Full Self-Driving Capability
- Navigate on Autopilot: automatic driving from highway on-ramp to off-ramp including interchanges and overtaking slower cars.
- Auto Lane Change: automatic lane changes while driving on the highway.
- Autopark: both parallel and perpendicular spaces.
- Summon: your parked car will come find you anywhere in a parking lot. Really.
- Recognize and respond to traffic lights and stop signs.
- Automatic driving on city streets.
So Tesla considers the following FSD features to still be PENDING:
- Recognize and respond to traffic lights and stop signs.
- Automatic driving on city streets.
Second, on the Autopilot page, Tesla gives a bit more detail on FSD:
"All you will need to do is get in and tell your car where to go. If you don’t say anything, the car will look at your calendar and take you there as the assumed destination or just home if nothing is on the calendar. Your Tesla will figure out the optimal route, navigate urban streets (even without lane markings), manage complex intersections with traffic lights, stop signs and roundabouts, and handle densely packed freeways with cars moving at high speed. When you arrive at your destination, simply step out at the entrance and your car will enter park seek mode, automatically search for a spot and park itself. A tap on your phone summons it back to you."
From this description, we can infer a bit more information on what "Automatic driving on city streets" will involve:
- Navigate urban streets (even without lane markings)
- Manage complex intersections with traffic lights, stop signs and roundabouts
- Smart Summon: your parked car will come find you anywhere in a parking lot. Really.
- Recognize and respond to traffic lights and stop signs.
- Navigate urban streets (even without lane markings)
- Manage complex intersections with traffic lights, stop signs and roundabouts
diplomat33
Average guy who loves autonomous vehicles
CarlK
Active Member
Yes lane marking is not necessary as had been demostrated already. Machine can even learn to predict direction of roads it could not see. We all can do those things we just never stop and think how we were able to.
I actually think machines can deal with pedestrains, dogs, cars, etc. better than human can. We can only focus on one or two things at a time. It could be hard to track the pedestrian on the left and dog on the right side of the street while also watch for the stop sign ahead. Machine seems to have ability to track many objects and process them simultaneously.
OK, I'm just different then.
I would call all of the features published macro features. Let's use an analogy since I'm looking at the fridge here in the morning. And just for the temperature function of the fridge...
Macro features:
- Keep the food cold
- Keep the freezer section colder
Rest of the features:
- Provide a digital temperature control for each section with appropriate ranges
- Allow fast freeze setting to increase compressor performance for 4 hours
- Detect door cycles to adjust compressor performance
- Measure ambient temperature, when warm provide door chime in 15 seconds, otherwise in 30 seconds
- Detect compressor failure and provide warning
- Detect temperature variations in each section and adjust fan speed to maintain constant temperature
- Run defrost cycle at timer cycle or as needed (detect ice formation on evaporator coil)
- Adjust rate of ice cube production based on consumption
- etc, etc
All of the stuff in the second section would have to be delivered for me to say the fridge is feature complete.
So, a feature like "Navigate on Autopilot: automatic driving from highway on-ramp to off-ramp including interchanges and overtaking slower cars" is so broad for me that it can be interpreted many ways. It is a macro feature. A word like "automatic" is too easy to overload.
Tesla could say this is done now since at least in some cases this works. For me and my interstate road travels this does not work. The car cannot merge into traffic at the on-ramp I use most. It picks lanes for exits that do not make sense (like the left lane for a right turn at the end of the ramp) and then informs me that it is ending in xx feet leaving me to make the next lane change for my right turn. There are a wealth of cases that need to be flushed out for this to be called feature complete. They are not corner cases. The only part that works well is the feature I had two years ago to stay in lane at speed. TACC continues to be useless in traffic. I get jerked around from the constant attempt to maintain following distance. TACC on the 911 is so much better, it allows for a range of following distance so acceleration (+/-) is minimized. There's just so much for me that is not done yet that I cannot call this feature complete.
My conclusion? People in and out of Tesla want so much for EM to be right (and deliver autonomous driving and RoboTaxis) they are willing to compromise what it means to do so. At some point Tesla will declare that their cars are full self driving and then be surprised when the rest of the world concludes they are not.
Huh? That's an accident every 10,000 miles. Waymo for example is way past that and still not ready to pull the safety drivers.
diplomat33
Average guy who loves autonomous vehicles
OK, I'm just different then.
I would call all of the features published macro features. Let's use an analogy since I'm looking at the fridge here in the morning. And just for the temperature function of the fridge...
Macro features:
- Keep the food cold
- Keep the freezer section colder
Rest of the features:
- Provide a digital temperature control for each section with appropriate ranges
- Allow fast freeze setting to increase compressor performance for 4 hours
- Detect door cycles to adjust compressor performance
- Measure ambient temperature, when warm provide door chime in 15 seconds, otherwise in 30 seconds
- Detect compressor failure and provide warning
- Detect temperature variations in each section and adjust fan speed to maintain constant temperature
- Run defrost cycle at timer cycle or as needed (detect ice formation on evaporator coil)
- Adjust rate of ice cube production based on consumption
- etc, etc
All of the stuff in the second section would have to be delivered for me to say the fridge is feature complete.
So, a feature like "Navigate on Autopilot: automatic driving from highway on-ramp to off-ramp including interchanges and overtaking slower cars" is so broad for me that it can be interpreted many ways. It is a macro feature. A word like "automatic" is too easy to overload.
Tesla could say this is done now since at least in some cases this works. For me and my interstate road travels this does not work. The car cannot merge into traffic at the on-ramp I use most. It picks lanes for exits that do not make sense (like the left lane for a right turn at the end of the ramp) and then informs me that it is ending in xx feet leaving me to make the next lane change for my right turn. There are a wealth of cases that need to be flushed out for this to be called feature complete. They are not corner cases. The only part that works well is the feature I had two years ago to stay in lane at speed. TACC continues to be useless in traffic. I get jerked around from the constant attempt to maintain following distance. TACC on the 911 is so much better, it allows for a range of following distance so acceleration (+/-) is minimized. There's just so much for me that is not done yet that I cannot call this feature complete.
My conclusion? People in and out of Tesla want so much for EM to be right (and deliver autonomous driving and RoboTaxis) they are willing to compromise what it means to do so. At some point Tesla will declare that their cars are full self driving and then be surprised when the rest of the world concludes they are not.
Yes, Tesla is defining "feature complete" at the macro level. The features listed on the order page are macro features. I would assume Tesla does this because they need a concise way of telling customers what they are getting. There isn't room on the order page to list every single feature that Autopilot will be able to do. All the things you mention are certainly important though. Tesla's approach is to deliver macro features, like NOA, to customers when the features are functional and then mostly from fleet learning, gather the data needed to perfect the feature. So Tesla delivers a functional macro feature and then adds the micro features, like you mention, later.
Your post does explain the disconnect and why so many on this forum are so skeptical when Tesla and Elon says things like "feature complete by the end of this year". For Tesla, this is a realistic goal because they are aiming at delivering a couple functional macro features (the last 2 on the website) by the end of the year. But many on this forum, have a list of like 100 micro features that they consider requirements for feature complete, so they see the goal of end of the year as completely unrealistic. And they would be correct except for the fact that Tesla is not aiming to finish those 100 micro features by the end of this year, they are merely aiming to deliver 2 functional macro features by the end of this year and then work on those 100 micro features later.
I also think that aiming for L5 autonomy presents Tesla with a huge challenge because there are so many different driving scenarios that a L5 has to be able to handle. This means that a feature like NOA may work really well in say Indiana but maybe not work so well in New York. This means that people may view feature complete differently depending on how well it works for them. If NOA works well for me, I am more likely to view it as feature complete but if it does not work well for you, then you will probably be less likely to view it as feature. Tesla may well consider a feature to be complete because it works reliably well by their standards but does not work well enough in your specific cases. But this is why Tesla has fleet learning. Tesla is constantly gathering data when there is a failure or weakness detected so that they can improve the feature for all those cases when it does not work. This takes time as you know but the fact is that features like NOA are getting better over time. It might just take some time before you see improvement in the driving scenarios or driving environment that you experience.
In conclusion, Tesla will deliver features like traffic light detection and automatic city driving at some point. No, those features will not be perfect right out of the gate. There will certainly be edge cases and micro features that are missing. So in your eyes, it won't be feature complete. But Tesla will work to solve those problems. So we can be sure that Autopilot and FSD will get better over time.
Yes, exactly! We have differing things we each would include at the micro level. I can see the macro features being very useful for marketing. Can we have a data sheet of what actually is in the delivery though? Seems like every release there are improvements but we don't know what they are and folks have to decipher them through reverse engineering.
We know these descriptions of improvements should be available. Otherwise how does management know that their developers are actually doing something? I sure would appreciate granular detail in some document that I could access as an owner describing what they did. Something like "AP will now identify the primary car in white and secondary cars in grey with noticed cars in shadow relief, secondary cars are defined to be ones that could become primary" or "AP now detects the primary car leaving the current travelling lane to avoid phantom braking". There are so many.
This kind of granularity would also be helpful to Tesla. Instead of saying something like "we improved the autowipers", and getting responses like "no you didn't". If they told use "we changed autowipers to detect mist and in that situation wipe once per 5 seconds" they would get feedback like 5 seconds is too long or too short. Folks who don't want the granularity can skip it all. But me thinks that this group wants it. Just look at the disappointment voiced each time release notes are too macro to be useful.
Similar threads
