Hacker News Reader

I think that’s a prime example for the shortcomings of machine learning. It can only “learn” from what is known, anything unknown can be potentially a catastrophe. I believe with the current AI in its infancy, a combination of rule-based and machine learning would be a more dependable approach.

View on HN

The problem with your statement is that AI drivers do rely on traditional logic already, but that cannot be applied to the perception part of the stack. It doesn’t matter if your rule is “if perceives truck, stop” if perception is telling you there is no truck stopped there. The rule also can’t be “if white pixel, stop” or “if Lidar bounces off steam and returns, stop”

if you're doing vision based obstacle avoidance and you're training your model on specific classes of objects and not just object you're doing it wrong

Reply on news.ycombinator.com