Using Neural Networks as autopilot for a model airplane

I'd be interested in the answer to this one myself.

Hmmm. You could take a lot of sensor data from real flights and feed it to your network as training data.

Or you could just put it in the plane and have a button for reward when it makes a good move. It could use an impact sensor instead of a "punishment" button. :)

I think that it depends. If you just want it to take off, fly a bit and land in a totally clear area, it could be fairly small.

Once you start adding wind, other planes, not running into people, etc... you wind up with a much more complex network.

It does sound interesting, though.

-- D. Jay Newman

You could try and train it using a software simulation of some kind. Maybe hack away at the source for flight gear or one of the other open source flight sims?

Actually i think it would have to be a fairly complex system. With a R/C car for example, you can reduce it to two functions, speed and heading.

When you get into flight, you don't have a direct way of measuring speed, because you aren't in direct relational contact with a fixed reference. In addition to heading, you will need some way of measuring roll, pitch, and yaw, and be able to compensate for changes in real time (a gust of wind for example).

Collision avoidance will be a hard one, since your stall speed will probably be quite high, say around 15-20 MPH. You'll need to be able to get sensor information for a huge distance around (and below) the aircraft.

This is going to add up to a hefty processor package weight wise, so the aircraft will have to be pretty big to be able to carry it and still be able to maintain a low stall speed, probably ending up with a quarter scale.

And don't forget to have a fail-safe override so you can take *immediate* control of it when it buggers up and heads for a person. And make sure your MAAC dues are paid up :)

Arawn

You could begin with references to developing expert systems and decision trees for this task (decision tree ref: Quinlan, ID3).

Kirt Undercoffer

Excellent idea.

Why not practice on a truck backer-upper (Widrow and Nguyen)

You might also check out decision trees (Lim)

Hope this helps.

Greg

Another possibility could be to have the neural net aboard the plane while you fly it manually, monitoring whatever sensors you have aboard. Let it decide what it would do, then have it see what you in fact do, and adjust its weights accordingly to act more like you. This could be happening more or less continually while you fly it.

- Bob

Which type of neural network you think can be used for this application?

On Mon, 13 Sep 2004, it was written:

I'd say that's a secondary question. The primary question is what kind of stimuli you'll feed to the network - what sensors are you planning to use, and what do you want the network to do? Fly the plane in a straight line or do artistic loops or land it or take off or what?

// Michael

Seems like there is a lot of good stuff on google.

You have to dig through some of thse links to see how they did it.

I think this is the most interesting one I saw:

Mark Browne

Hmm I would think an expert system would be good for this, but NASA's X-36 looks like it uses neural networks.

I was thinking of trying the same thing only with RC cars. Maybe we can share notes :) ?

Im at snipped-for-privacy@hotmail.com .

GL! Charles

Have you seen:

Not a neural net but the neural net could be added on top. The system is not just for helicopters, but fix wing planes as well.