Quadrotors flying in disturbed environments face aerodynamic effects that change as they move through the air. Predictive models trained for one set of conditions can become unreliable when winds vary in space. A new approach from an arXiv preprint introduces a disturbance-aware Koopman Control Family, or DA-KCF, which adjusts the predictive model in real time to keep it accurate during deployment.
The method builds on Koopman operator techniques, which represent nonlinear dynamics in a linear form suitable for control. DA-KCF extends this idea by making the model sensitive to disturbances, so it can account for the changing aerodynamic forces a quadrotor encounters in nonuniform winds. The authors combine this adaptive model with model predictive control, a common strategy for planning trajectories under constraints.
According to the abstract, the motivation is reliability: a quadrotor's controller should not rely on a model that drifts from reality as unmodeled effects emerge. How the method performs relative to other controllers, or in real flight tests, is not detailed in the abstract alone.