[tobinfricke]

The top-level autopilot used a nonlinear path controller, which produced position, attitude, and rate commands for an inner-loop controller that used linear feedback and gains computed using LQR and a linearized model of the system.

Because the maneuver the kite must fly is very dynamic (always turning, but not at constant rates), control with feedback alone would lead to pretty sloppy flight, since feedback requires an error to develop before the system can respond to it. Good feed-forward was crucial to flying well; from the desired flight path, wind speed/direction, and desired aerodynamic quantities (airspeed, angle of attack, and angle of sideslip), we could compute the required rates and feed-forward actuator displacements needed to execute the trajectory.

For more details, check out my paper titled "The Makani Autopilot: A Critical Retrospective", available here: https://storage.googleapis.com/x-prod.appspot.com/files/Maka...