Learning dynamics of drones with moving arms in real time

Aerial manipulators—drones carrying robotic arms—are difficult to control because the quadrotor and arm coupling, aerodynamic delays, and payload changes all create nonstationary, history-dependent dynamics that offline models cannot capture. This work proposes an encoder-decoder framework where a nonlinear encoder learns cross-coupled temporal dependencies from state histories, while a linear decoder enables closed-form Bayesian adaptation online. Tested on a real aerial manipulation platform, the approach improves residual dynamics prediction, adapts quickly to changing operating conditions, and enhances MPC-based trajectory tracking—all while remaining compatible with real-time control.