Champion-level drone racing using deep reinforcement learning

[Thenop]

https://www.nature.com/articles/s41586-023-06419-4

And so it goes... not unexpected but still. Not that I understand any of the formulas on the 2nd half of this story, but I get the idea.

[The Picnic Wasp]

I recently decided to learn a bit more about baseball, the stats of which I think might be complex enough for me. Drone racing is definitely a bridge too far.

[Slim]

That's fascinating. So: the drone is able to work out where the gates are in relation to itself from images provided by its onboard camera. It also needs to recognise and avoid other obstacles. presumably it builds a 3D map of its environment from the 2D images provided by its sensors, like a human does (I wonder if the drones have stereoscopic vision?)

It also needs to understand the implications to its own manoeuvrability from an excess of speed in a turn, and I suspect that's the main benefit of the deep reinforcement learning.

[Nick]

For simpletons like myself here's some footage.

https://www.youtube.com/watch?v=pq53uCDZelQ

[Slim]

Seems to me that virtual drones would be the obvious next stage in the evolution of that competition.

[R6GYY]

Very interesting.

I remember some time ago competitions where people would code combat robots and send them off to compete. A bit like Robot Wars but no hardware required.

This was some years ago, and I don't think that AI was involved, just good old-fashioned human ingenuity and coding skills.

[R6GYY]

Then again, I'm getting old and maybe that wasn't quite it. Something like that anyway.