Tail of the dragon autopilot challenge

[verygreen]

So the other day, I decided to try AP on the tail of the dragon... And frankly about 3.5 minutes in I decided it's waaay too dangerous and I should stop.

I post most of the footage here. at ~3:25 the car solidly went into the other lane (Across a double yellow) and felt comfortable staying there so I aborted my little experiment. Before there were also a couple attempts to accelerate into a tight turn as well.

Here's the video with AP augmentations, driven on 18.42.2 (thanks to Tesla who make it possible to get high framerate videos on that version now. Also @DamianXVI was impressed with their new solution for color channels representation making even compressed videos have better colors and being ap2.0/2.5 agnostic in process!)

Additionally of note - we changed the meaning of the lane colors. Red is a single line, bold red is double (both the car refuses to cross in normal conditions), green is dashed line (car should be ok to cross), yellow is unknown type (by the car), cyan is what we think curbs/road edges and blue we are not sure what it represents at this time.

Would be interesting to compare to what AP1 does after @wk057 post here.

BTW this is a rare sample where I was able to get the backup camera to work for like 1.5 hours before quitting so I was able to get some good footage from it.

 

[strangecosmos]

Be safe out there VG!

 

[wk057]

Be safe out there VG!

I don't think I've been out that way in ages... well before I had an AP car. I believe AP1 would do ok there, at the low speeds. I've found AP1 to be able to do seemingly impossible turns, as long as I have the speed set appropriately for the turn. As an example, there is a hairpin turn on...... 226 I think?.... not too far from me (about an hour). At 35 MPH (which is about what I can personally take that curve at, with some lateral G's, in the S) it doesn't have a chance. But at the posted suggested speed of 20 MPH, AP1 makes it through without too much issue.

In my experience, AP1 will take just about any clearly marked curve if you're going slowly enough... which is probably a good idea in these mountain areas anyway.

I notice quality is greatly improved from your other videos... nice work.

 

[verygreen]

For a good measure, here's some "driving on ap in construction" video.

No big surprises other than in the second segment (from 3:26 onward, though there were similar things on the left side earlier in the video) it totally thinks it's A-OK to drive between those traffic cones onto the closed-off lane, wtf? Smart move on the city's part to block off the exit with the construction trucks

 

[pkodali]

BTW working on a 360 video for this, but positioning of the rear camera makes blending very ugly, and it is taking around 7 hours to rneder at 30 second clip because of all the extra perspective shifts needed for each camera view

 

[verygreen]

BTW another nicety, here's some backup camera footage in the rain during night. Very pretty with all the droplets! backup2-2 - Streamable

 

[skitown]

BTW another nicety, here's some backup camera footage in the rain during night. Very pretty with all the droplets! backup2-2 - Streamable

One of the dudes on the left is giving you the stink eye.

 

[J1mbo]

Wow, big improvements to the picture quality, kudos!

From what we see, path prediction looks a lot more stable than the Paris video (probably not a fair comparison) and seems to be more robust and accurate than any of the boundaries.

I notice the right pillar classification seems to be very poor compared to the rest - drivable area (9:11) and pedestrian detection (7:56) stood out as examples.

Great road to drive, can see why you went manual

 

[lunitiks]

BTW another nicety, here's some backup camera footage in the rain during night. Very pretty with all the droplets! backup2-2 - Streamable

awesomeness
[aw-suh m-ness]

noun
1. @verygreen causing or inducing awe; inspiring an overwhelming feeling of reverence, admiration, or fear:
an awesome sight.
2. exhibiting awe; showing reverence, admiration, or fear.
3. Slang . very impressive:

 

[electronblue]

Nice vid this and the others! Seeing the rear and side vehicle identification really brings home how inaccurate they are though. The front with three cameras and radar is better of course. On version 9 dashboard the cars are very nervous especially to the sides and the rear and these videos show exactly the same nervousness from Teslas car identification so this seems to be source of the nervousness.

Distance estimation is especially off towards the sides and the rear too when looking at the distance numbers as well. Distance numbers are all over the place. It is clear why the dashboard shows the side and rear cars moving back and forth and almost hitting your car.

 

[verygreen]

Wow, so there's a 360 video from @pkodali that's really cool, I think. Time to get some of those VR headsets to enjoy it the most. If you don't have one, be sure to pan around with the mouse.

 

[lunitiks]

That is incredibly cool. If you watch it on the youtube app for iPad, the iPad’s IMU let’s you move the view around by physically moving the pad, pretty sweet.

At timestamp 1:20 the reverse camera freezes