Alan Zavari on Twitter
Car accelerates towards car transporter.
Car accelerates towards car transporter.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autopilot fail
- Thread starter DanCar
- Start date
seattlite2004
Active Member
diplomat33
Average guy who loves autonomous vehicles
The most likely explanation is that the camera vision saw the car that was on top of the carrier and did not "see" the rest of the truck, especially when it was very close to the carrier. As a result, AP thought there was more space in front than there really was. The NN may not be properly trained to correctly see this type of car carrier up close. Basically, another "edge case" that the NN is not properly trained for.
If I am right, it is another example of why camera vision alone is not good enough for safe autonomous driving. Sure, at some point, the NN might be trained well enough to handle all the edge cases but in the mean, you will have accidents, if the driver is not paying attention. Including lidar would easily solve these problems, because the lidar would have detected the edge of the carrier and known not to accelerate. Yeah, maybe lidar is a crutch, but it's a crutch that can make your autonomous car a lot safer as you work towards your perfect camera only system. Tesla's approach is to basically accept a lot of easily avoidable failures on the march towards perfect vision NN that works.
I’ve experienced this with Volvo’s Pilot Assist, expect I’ve never rear ended anything (likely because I pay attention, who knew?). Pilot Assist has done this with certain landscape trailers, car haulers, low-boy trailers. It doesn’t do it all the time but it has happened to me. Funny, someone in the Twitter post mentions Volvo’s system would handle this just fine. It doesn’t. Real talk.
This happened to me twice now over 50k miles and 3 years. First was a garbage truck with very early AP2 in 2017. Second was a matte black model S at night stopped at light. Car in both instances forgot about stopped car in front after we stopped and sat. I hit brakes in both instances well before collision. This driver wasn't paying attention and trusted too much... But this proves we are far from FSD.
I wonder if he had HW2. 5 or HW3. That can make a big difference in object processing
AlanSubie4Life
Efficiency Obsessed Member
Yeah this was a really rapidly evolving situation.
Avoid cognitive dissonance...avoid cognitive dissonance...
AlanSubie4Life
Efficiency Obsessed Member
“Unattended acceleration” is more like it!
I find it a bit puzzling because the semi was moving so it should be registering on the radar.
It was also going so slow that wouldn't it at least show up on the ultrasonics within the last few feet to stop it?
From a purely vision standpoint it really depends on how the system is implemented. I believe Subaru implemented a stereo vision system that will even stop for "blobs" meaning that it uses stereo depth to sense anything in front of it, and get a good idea of it's size. So you don't have to train a Neural network on every possible combination of stuff.
I believe the Tesla system has to be trained to recognize everything. In this case it failed since it likely saw the car further away, and wasn't trained on the car carrier.
The ironic thing is Tesla uses so many car carriers you'd think they've have a highly trained system to recognize them.
It was also going so slow that wouldn't it at least show up on the ultrasonics within the last few feet to stop it?
From a purely vision standpoint it really depends on how the system is implemented. I believe Subaru implemented a stereo vision system that will even stop for "blobs" meaning that it uses stereo depth to sense anything in front of it, and get a good idea of it's size. So you don't have to train a Neural network on every possible combination of stuff.
I believe the Tesla system has to be trained to recognize everything. In this case it failed since it likely saw the car further away, and wasn't trained on the car carrier.
The ironic thing is Tesla uses so many car carriers you'd think they've have a highly trained system to recognize them.
Last edited:
Daniel in SD
(supervised)
The car carrier definitely wasn't moving when the car started moving.
Subaru doesn't have radar as a crutch to rely on.
I imagine that the binocular vision works great at close range like this. I can sympathize with the driver. After 20k miles he probably thought his experience was enough to "know" that Autopilot won't accelerate into stopped vehicles.
AlanSubie4Life
Efficiency Obsessed Member
The More You Know....the worse you are at using AP. Just like Daddy Elon said.
I meant that the car carrier was moving which enabled radar tracking. So when the car carrier came to a stop it's still there. It's not like its going to magically disappear.
AlanSubie4Life
Efficiency Obsessed Member
I’m not sure that AP is quite to the point of object permanence. It is but a helpless infant. Probably about as good as a 1-4 month old at driving when viewed from this perspective.
Object permanence - Wikipedia
This is not the first time an object has magically disappeared. There was a video a long time back with a white truck against white clouds which magically disappeared.
Blending perception with a constructed 3D model of an environment (not clear that this is done at all) and then using that model to cross check your perception seems tricky.
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
@S4WRXTTCS. This is a good example of the continuous updating going on. There is not really a conception of the past. Objects just change size instantaneously (in this particular example).
green on Twitter
green on Twitter
EVNow
Well-Known Member
Barklikeadog
Active Member
None of the cars currently on the road will ever be full 'Robotaxi' style FSD. Elon bet all of his chips, trashed critics (Lidar), but this is not a winning hand.
diplomat33
Average guy who loves autonomous vehicles
Yes. That's why in another thread, I predicted that Tesla will achieve "self-driving" on the current AP3 hardware but only with driver supervision and will struggle to reach the safety level required to remove driver supervision. Basically, they will get the car to navigate a route, stop at lights, turn at intersections, auto park, etc but will encounter cases that they simply cannot solve reliably without more sensors so they will be stuck with keeping driver supervision. The real question is how long will Elon/Tesla go before they admit the truth and add more sensors.
Barklikeadog
Active Member
Similar threads
