In the piece — titled “Can You Fool a Self Driving Car?” — Rober found that a Tesla car on Autopilot was fooled by a Wile E. Coyote-style wall painted to look like the road ahead of it, with the electric vehicle plowing right through it instead of stopping.
The footage was damning enough, with slow-motion clips showing the car not only crashing through the styrofoam wall but also a mannequin of a child. The Tesla was also fooled by simulated rain and fog.
I mean it also plowed through a kid because it was foggy, then rainy. The wall was just one of the tests the tesla failed.
Right, those were the failures that really matter, and Rober included the looney tunes wall to get people sharing and talking about it. A scene painted on wall is a contrived edge case, but pedestrians/obstacles in weather involving precipitation is common.
It’s no longer an edge case if faulty self driving becomes the norm.
Want to kill someone in a Tesla? Find a convenient spot and paint a wall there.
Doesn’t even have to be an artificial wall, for example take a bend on a mountain road and paint the rock.
A better trick would be to paint the road going straight when there’s a cliff. Much easier to hide the evidence that way.
Next test I would love is what is the minimum amount of false road to fool it.
Have you ever seen examples of how the features that ai picks out to identify objects isn’t really the same as what we pick out? So you can generate images that look unrecognizeable to people but have clearly identifiable features to ai. It would be interesting to see someone play around with that concept for interesting ways to fool tesla’s ai. Like could you make a banner that looks like a barricade to people, but the cars think looks like open road?
This isn’t a great example for this concept, but it is a great video. https://youtu.be/FMRi6pNAoag?t=5m58s
I was thinking something that the AI would think the road turns left and humans see it turns right