Automatic driving on city streets by EOY - Feasibility

Automatic driving on city streets by EOY - Feasibility

Well..In almost 2 weeks, it will be nearing the end of the year. The stop sign and traffic light recognition have been released (although it seems like it's incomplete), technically. Would the "Automatic driving on city streets" be also rolled out with a surprise before 2020? As much as I am excited about it, I'm very curious to know how our vehicles are able to turn left and right at intersections as well as at any corners. Watching the latest video of FSD posted by Tesla, it looks like the capability is there. With the "Automatic driving on city streets," rollout, how much capability should we expect?

Some Early Access drivers should be getting some of the FSD features by EOY to test out according to Elon, but would 40.2.1 be the last major update of the year..?

What are your thoughts?!!

BadgerErickson | 15. joulukuu 2019

IIRC, the essential aspects of FSD are not in Tesla's court, yet? The camera is dependent on seeing paint lines, correct? If most municipalities can't even afford basic pothole maintenance, I think it's inevitable that EM will why've to develop the COLOR Co., providing the paint to cities.... | 15. joulukuu 2019

No FSD this year.

I'm expecting FSD at the end of 2020, and it will have limitations. Without regulatory approval, it will require nags like today. It may be functional internal to Tesla at the end of 2019, but still, a massive amount of simulation and testing is required, along with fixes as issues show up.

Of course, the fleet has to be upgraded to HW3 too, no small undertaking.

Frank99 | 15. joulukuu 2019

The roll out of FSD on city streets will be...interesting.

I live in Phoenix, where they do a pretty good job of keeping the streets well-striped. As a result, there are half a dozen companies doing autonomy out here; it's common to run into the various Lidar equipped vehicles when driving around. We'd be an excellent (hint, hint) leading-edge rollout for FSD. When you get to some of the NE states, where road maintenance is neither as easy nor as well funded, it'll be a lot harder to get FSD working well most of the time. It wouldn't surprise me if Tesla geo-fences FSD by area - "FSD on city streets is now available in Arizona and California, and will be available in Boston and Chicago as soon as their streets meet modern standards..."

justinyoon2011 | 15. joulukuu 2019

Frank99 - Geo-fencing seems to be a practical move for Tesla to take when it comes to FSD roll-out. Hope they release a beta version to the public sooner than later just like how NOA has been available as a beta version and it's been improving over time.

billtphotoman | 16. joulukuu 2019

FSD won't be happening on city streets with the current side camera sensor range. Tesla's own data shows the side camera can only "see" 60 meters so crossing or turning onto a street with a 40 MPH or higher speed limit (required for me to get out of my own neighborhood) wouldn't be feasible because best case the car could only "see" a motorcycle or car approaching about 3 seconds away (40 MPH = ~20 meters/second). And it gets worse at higher speeds because the car needs to be able to see them further away. Basically for true FSD on city streets the car needs to be able to "see" nearly as far to the side and it does straight ahead. And Tesla themselves uses 3 cameras for forward detection. Given the brainpower Tesla has working for them I am sure they are aware of this and at some point will roll out a new side camera system. Once I see that I will become convinced true L4/L5 on city streets is close. Finally, I believe in efficient markets and with all the private capital looking for good returns these days once the smart money becomes convinced RoboTaxis / Tesla Network will soon be viable the bidding wars to buy Tesla cars will start. If there was really even a reasonable chance of earning $100K per year on a $50K investment the smart money would be all over this. All that being said I really _want_ true L4/L5. I just think it is farther away than many think. I do think Tesla will get there first and am pretty confident when I buy my next Tesla in 2028 or so it will have been out there and proven for a while. | 16. joulukuu 2019

@bill - It may be that those rare areas as you describe will be geofenced. That represents a tiny slice of FSD driving, although if it's a route you take, it means no FSD for that one intersection.

I also doubt that a 60-meter distance is a knife-edge threshold. It may be at 60 meters the side camera can't determine if the moving vehicle is a car or a truck, which is not important but can identify it's a vehicle moving in your direction. Perhaps moving sensitivity is 200 meters. Also perhaps the camera was not the issue, but HW2.x processing power. Perhaps HW3 can process more detail at farther distances. We really don't know the true limitations of the system, especially with the newest hardware. The camera distance specifications were created 3 years ago, and the new HW3 didn't even exist back then.

billtphotoman | 16. joulukuu 2019 - I don't think I would characterize an unprotected turn as "rare" because it will happen any time one needs to turn right at a red light or turn at a stop sign when cross traffic doesn't stop. Or enter a road from a parking lot. As for the cameras, if better processing is all that is required why does Tesla need 3 cameras for the front? As an avid bird photographer it is all about "pixels on the bird" or in the case of Tesla pixels on the object. There are 3 ways to achieve that: 1) throw megapixels at it. IE, keep the focal length the same and just increase resolution. 2) Throw magnification at it 3) combination of 1 and 2. Tesla uses 1 and 2 (3 cameras with different focal lengths) for straight ahead detection. I do agree the side imaging and detection problem is simpler to solve. It just needs to be able to detect an approaching vehicle and have enough resolution to accurately measure the speed. But, picking out say a motorcycle against a complex background still requires quite a few pixels on the subject. Again, given the brainpower Tesla has working for them I have no doubt they will solve it. I am just unconvinced the current set of cameras is sufficient. But, given how fast camera prices are falling Tesla was probably wise to just deploy what they need for highway driving. I can't imaging retrofitting better cameras would be too expensive. | 16. joulukuu 2019

I was only addressing the issue of an unprotected turn with freeway fast traffic. When the speeds are lower, there is far more time to see and make decisions. But you are right that there may be more situations than I expect.

My understanding is HW2 can only process a camera image of about 320x240 pixels. HW3 is supposed to handle full-frame or about 1280 x 720 pixels, or about 12 times more pixels, quite a big difference. The camera lens also makes a huge difference as to distance it is able to resolve. In the case of the front, Tesla is capturing far distances directly in front with a narrow-angle lens, and more content with medium angle lens, and a wide-angle lens for close up information such as pedestrians. I suspect it also offers a bit of redundancy and validation in case part of the image is obscured - more likely to occur on the front windshield than other cameras. Anyway, that's my take. We'll have to wait and see if unprotected turns are any issue or not. Obviously, it does need to be dealt with for total FSD.

billtphotoman | 16. joulukuu 2019

@Tesla.Tap - Enjoying this discussion first of all. I was talking about speeds of 40 MPH in my example which at least here in Austin aren't freeway speeds except for I-35 :). The other issue with detecting vehicles approaching from the side is if the car gets it wrong and misses, say an approaching motorcycle, there is a) probably not enough time for the driver to react and b) very high likelihood of loss of life. So, driving straight ahead if AP/FSD doesn't see a truck tire in the road I have ample time to react if I am paying attention. On the other hand, if AP/FSD determines all is clear in an intersection and misses an approaching vehicle I basically have very little time to react because I am in the intersection already. As you say, we will see how this plays out. I do really look forward to L4/L5 reducing the terrible carnage on the roads but getting there might be a little more challenging than people think.