Comment by Wouter van Verre on August 24, 2010 at 1:15pm

Wow! That is pretty impressive!

Comment by Tom Yochum on August 24, 2010 at 1:47pm

I could watch these all day...

There is some really useful technology here for the hobby autopilot: automatic control gain tuning. I hope to someday try to get something like this working. So far, the tricky part has been coming up with a "linear-enough" system model.

Comment by Fabien Bruning on August 24, 2010 at 2:16pm

That looks very nice, I wonder if the system actually knows what it's lost, or just has a self-tuning controller that figures something doesn't work, and adjusts the gains.

Comment by Sami Finnila on August 24, 2010 at 4:15pm

Coupled PID would work very similarly, to loosing a wing for example, if properly tuned. No live tuning of the parameters is needed, really, just the initial tuning. The problem with coupled PID is that they're very hard to tune - at least by hand - since 6 separate coefficients need to be optimized and every coeff affects the other. Automatic tuning of PID is very interesting subject and after that's achieved on the amateur level more complex control loops like coupled PIDs can be introduced as well. Maybe this'd be a future project for the ArduPilot development team? ;P

Comment by I Heart Robotics on August 24, 2010 at 7:36pm

They are using something called relay tuning for their controller design, which looks like an automated version of the Ziegler Nichols method.

This paper was written by some of the team that is now at Rockwell.
http://www.aiaa.org/store/cart.cfm?BackTo=..%2F%2Fagenda.cfm%3FluMe...

This has some info about relay tuning.
http://www.chem.mtu.edu/%7Etbco/cm416/Atune.html

Comment by John Arne Birkeland on August 25, 2010 at 12:53am

That is just impressive. Mathematically not that complex, but ohh so hard to make work in practice.

Comment by john bero on August 25, 2010 at 7:28am

Why did they blot out the ground view of the auto landing; is this some secret location?

Comment by Ed on August 26, 2010 at 4:12am

This is something I'm very interested in. Does the controller have a discrete set of possible known faults or does it just re-learn its model constantly? How quickly can it detect a fault? I think this fault tolerant stuff is where UAV research (and a lot of practical control applications in general) is heading. Very interesting and very important.

Comment by Tom Yochum on August 27, 2010 at 8:09am

Ed:
The controller looks at the inputs it is putting into the system and the outputs that result. It generates a mathematical model that maps the inputs to the outputs. When the wing falls off, the system immediately notices that the existing model no longer matches the input to the output, and it re-computes the model. It is all very quick.

I agree, this is very important stuff for both the hobby-level and the full-sized aircraft level.

Tom

Comment