Hacker News Reader

The idea of a Tesla self-driving system getting target fixation and driving towards the thing it should avoid is hilarious to me. And it also makes me wonder how likely that is to actually happen. Is the model weight of "shiny, bright reflective thing in IR spectrum" for "probably a road reflector, avoid" higher than "probably some water in the road, so that must be where the road is"? (Obviously totally made up examples).

View on HN

The risk is the image classifier deciding there isn't anything to collide with and not bothering to steer around.

Is this just speculation or do reflective surfaces like traffic vests actually dazzle cameras/LIDAR?

Uber ran over a woman [1] walking with a bicycle because she "didn't exist" as far as their software was concerned. This is a real danger when using the public as beta testers for robots. The Uber vehicle has more sophisticated sensors than a Tesla which is winging it with just cameras.

[1] https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg

Reply on news.ycombinator.com