3 ms·
You don't need the sensor data, it's obvious. The lane markers widened, and 'straight' is into the (not seen yet) barrier. If the barrier end had a mockup of th
by Applejinx 9y ago
You don't need the sensor data, it's obvious. The lane markers widened, and 'straight' is into the (not seen yet) barrier. If the barrier end had a mockup of the ass-end of a car stuck on it (about three times as wide) the Tesla would freak out, panic brake, and still think it was in the center of a single widening lane. It's not mysterious what the car was thinking.
If the overall shape of the road wasn't veering slightly to the right, the car would probably not have chosen to swerve to the left into a barrier, but the car sees the lane widening equally in both directions. Simple as that.
- Y_Y 9y agoObvious? Not to the car! My point is that gathering failing cases like this could be a a simple way to gain confidence in a self-driving system, either from the perspective of internal development at Tesla, or even for regulation. Imagine if before you were allowed to push an update to road-going cars you had to show that it doesn't crash in any of the simulations of difficult conditions that the Road Safety Authority has prepared. I'd be in favour of such a thing at least. It's not proof, but it doesn't give any false positives.