Model 3 - Beta FSD 12.3.1 - Drive around Santa Clarita - 2023.44.30.30 - 30 March 2024
HTML-код
- Опубликовано: 5 сен 2024
- Crazy drive around Santa Clarita. Some serious rain on this drive and pushed FSD Beta to its limits for heavy rain.
Long video but some interesting experiences. Some long waits at lights too. lol
Completely unedited video for you to make your own determination on how FSD Beta v12 is doing.
Difficult 44 minutes drive, and no safety disengagements. One disengagement due to poor visibility and no lane detected on an empty road - fsd would correct itself soon. Other one because driver did not see that fsd tried to avoid potential accident. Other incidents just to report the isse (using blinker would be enough). Some minor issues with signalling.
11:40 - heavy rain, car passing on the left drifted to the right, so FSD made a room for it to avoid collision, probably unnecessary disengagement
12:51 - no turn signal while changing into a starting lane right after the fork
19:50 - 0.4 mil to the turn - IMPATIENCE - should it be changing lane already? Driver unhappy. Low speed in the rain, not many cars ahead.
20:06 - 0.3 mil - intervention, forcing lane change
20:18 - 0.2 mil - fsd would probably change the lane around here, no cars around
20:23 - 0.1 mil - so 17 seconds window for FSD to make one lane change if not disengaged, we will never know if it would
21:50 - no turn signal while exiting a roundabout - maybe because of no cars behind? Still it should to it.
23:05 - fsd a bit too slow, would probably miss the light, driver accelerates
23:15 - 0.4 mil to the turn - IMPATIENCE - remark that car needs to change lane right after the turn (two lane changes total)
23:25 - 0.3 mil - fsd makes the the first lane change
23:45 - 0.2 mil - last chance to make that hypotetilcal first lane change (done already)
24:10 - 0.1 mil - fsd too slow to change to the actual turning lane?
24:26 - 0.1 mil - fsd makes the lane change in time before the turn
24:44 - wrong lane detection after the turn, you cannot see lines in the camera, a lot of reflections on the wet road, disangagement [poor weather warning on]
32:40 - 0.7 mil to turn - driver planning lane change ahead of time, as three lanes of cars would want to make into one lane
33:33 - 0.4 mil - IMPATIENCE
33:48 - 0.2 mil - fsd makes lane change, without any issues
34:01 - 0.1 mil - last chance to change into this lane, traffic is not heavy, a lot of posibilities
In general, seems like planning lane changes ahead of time could be improved. Driver's mood is consistent - impatent around 0.4 mil before the turn if the fsd is not starting to change the lane. One issue noticed when the visibility was impaired by the wet road condition (and sunlight ahead). Otherwise awesome performance in the rain and heavy rain. Autowipers could be improved of course, but they were actualy decent. Dynamic driver used accelerator taps quite a lot for faster travel. FSD could be improved in speeding up as well, but ony in the assertive mode, as it could be a bit jerky for passengers.
24:47 An observation here: The visualization (mainly provided by the old v11 stack, right?) shows the lines correctly. That proves that the cameras and a NN are able to detect them correctly and no hardware limitation, simply the new v12 NNs seem to not yet have reached that precision.
Edit: Or are they...? It's really difficult to make that out on the screen. I can definitely see some gray lines at some points.
I agree about it not moving into the needed lane for upcoming turns soon enough and often when it has a good opportunity... my observation as well. Also drifting into the other lane during two-lane turns (as I saw once in your video as well) is problematic for me. Also, sometime hesitant to pull into traffic when it has a good chance. Otherwise, yes, pretty amazing stuff.
it needs medium distance planning for more than 1/4 mile out, lane selection etc. looking good though.
I think, you tested the software really hard, in rain time driving. Night and heavy rain and heavy fog testing will be even a harder, but needed test too. I am impressed with the performance of the software. Radar sensors which used to be in it could be useful in those situations, but minimalistic and too much efficiency based logic are dumping on that functionality which could be corrected as radar is another data input too. Night driving require IR sensors and RGY sensors but that's up to the manufacturer and customers. Wipers are required inside both of the optical rear view mirrors placed outside, for many cars, which are currently absent for checking sides, during turning in rainy weather, hope the manufacturers, take a note on that, as that will improve road safety in general.
An easy rainy day drive for a human, is quite difficult for the beta FSD software:
0:40 hesitation, making a Left turn
8:58 "FSD may be degraded because of weather", in broad daylight
and very light rain. ??
11:38 Human must intervene: vehicle is drifting out of lane
12:58 Human must intervene: lane change, without a turn signal
20:00 Human must intervene: vehicle not changing lanes
23:03 FSD does not react to stale green light ahead
Human intervenes: accelerate to make the green light
24:42 Human must intervene: FSD is on the shoulder, wrong spot.
15:00 mins pouring rain. Are you insane 😱. Still in beta.