Prof. Andrew Ng: So, what you're seeing today is by far the most difficult aerobatic manuevers flown by any computer controlled helicopter.
Adam Coates: The goal is always for us to take really hard problems, and humans are very good at solving. For instance, Garett can pick up any helicopter, even ones he's never seen and go fly amazing aerobatics. So the question for us is always: why can't computer do things like this? Our goal is to build software that can learn to go fly a helicopter the way that a human being can, and that way we don't need so much engineering effort to build control systems like this.
Prof. Andrew Ng: The basic technology used to develop the helicopter controller was a type of AI algorithm (a type of Artificial Intelligence algorithm) called the apprenticeship learning.
Prof. Andrew Ng: What the helicopter does is it watches a human pilot flying a helicopter around for a while, and by watching a human pilot learn to fly, it then learns to fly by itself. And so, all it does is watch a person fly, and then it will try to fly the same stunt maneuvers by itself, and maybe try a few times until it nails the manuever. What you're seeing is the end result of this machine learning process called apprenticeship learning.
Adam Coates: Basically, what the software does is it acquires all of the measurements off of the instruments, and then it puts them into something called the common filter that figures out where the helicopter is, which way it's facing, what its velocity is, and so on. Then that information is fed into a control system that decides based on where the helicopter is and where we want the helicopter to go. The controller actually computes new controls to send out to the helicopter every twentieth of a second.
Pieter Abbeel: What we do is we collect several demonstrations from our expert pilot, and then our algorithm processes demonstrations to figure out what would be the ideal version of, let's say a flip, or a roll, or a loop, etc. and then we use that as a specification for our control problem. And there are a lot of us to find trajectories that are better trajectories than the pilot can demonstrate.
Garett Oku: The helicopter doesn't want to fly; it always wants to just tip over and crash. You always have to be maintaining the orientation of it. For as an airplane, if you let go of the stick, it will more or less fly fairly straight. I've never seen anything like this done. The only autonomous helicopters I've seen -- they're hovering at waypoints and doing quite basic manuevers.
Prof. Andrew Ng: In order to make these applications work well, in order for us to trust helicopters and these sort of permission-critical applications, it is important we have very robust, very reliable helicopter controllers that can fly maybe as well as the very best human pilots in the world can. And the work you saw today I think took a large step towards achieving that.
