Hacker News Reader

Comment by udkl | Hacker News Reader

Why would they not use a general purpose classifier [1] instead ?

Sure, the tagging of objects in the field of view in this model may be unnecessary but you leverage an existing model that should allow the drone to 'think' beyond the current limited "obstruction here". It could at-least have been used as a base model to build upon.

[1] https://www.youtube.com/watch?v=_wXHR-lad-Q

Personally, I'm also looking forward to neural networks modeled after real brains [2] .... but the tech to accurately scan the complex interconnections in larger brains seems far away.

[2] http://www.smithsonianmag.com/smart-news/weve-put-worms-mind...

View on HN

Why would an object classifier matter?

If my plane is about to crash, do I really care whether it's a mountain or a building?

Would you care if it's a cloud or a reflection off glass?

Well you could decide the action based on the type of the object.

eg: If it's a human "Bob" : listen for command. If no relevant command, detour & continue to goal.

If it's a vehicle in the middle of the road, wait for it to move and then go no with your path if there are no other moving vehicles on a collision trajectory with you.

It's definitely a step ahead of what the OP are doing.... but isn't it a more practical approach ?

The point I'm trying to make is that training an NN to efficiently detect objects has been a solved problem for quite some time now. We should give more attention to experiments that take it a step ahead.

You would have really cool behaviour.

+ Not going through glass (although humans do not always exceed in that themselves)

+ Going through a fly door curtain

+ Going through smoke

+ Not going through a mosquito net.

+ Not going through a fountain.

+ Pushing through soft objects.

+ Pushing a half open door.

The method used would also end up with many of these behaviors. If it didn't crash, that behavior is enforced. So flying through smoke would not be avoided if it didn't crash

Reply on news.ycombinator.com