Aerial manipulators—robotic arms attached to flying platforms—pose a difficult dynamics-modeling problem. The platform and arm interact strongly, the system's behavior depends on its history, and payload or configuration changes make the residual dynamics nonstationary. Standard models often fail to capture these cross-coupled, time-varying effects.

The authors propose a structured encoder–decoder framework. A nonlinear history encoder learns temporal and cross-variable dependencies in the residual dynamics, while a lightweight linear latent decoder provides a linear-in-parameter structure. That structure allows closed-form Bayesian updates with covariance inflation, enabling fast adaptation at deployment and straightforward integration with real-time model predictive control (MPC).

Experiments on a real aerial-manipulation platform showed improved residual-prediction accuracy, rapid adaptation after dynamics changes, and lower MPC tracking error under varying payload conditions. The results indicate that combining history-dependent representation learning with lightweight online adaptation is an effective approach for these coupled, nonstationary systems.