Conditional Invertible Neural Networks for UAV Control

A recent research paper explores the application of conditional invertible neural networks (cINNs) to enhance the control of unmanned aerial vehicles (UAVs). Specifically, the study positions cINNs as probabilistic inverse-dynamics models for multirotor systems. In a 2-D proof-of-concept, the researchers trained a cINN to learn the complex relationship between desired control actions and the resulting system states. This learning process was guided by an incremental nonlinear dynamic inversion (INDI) teacher, utilizing rational-quadratic spline coupling and invertible linear mixing techniques. The results showed strong performance in open-loop reproduction, with high R-squared values and low mean CRPS. In closed-loop scenarios across 15 different tests, the cINN-based control achieved position RMSE comparable to the INDI teacher, with nearly half of the tracking attempts being acceptable. The primary limitations identified were attitude divergence during aggressive maneuvers and phase lag with high-frequency commands, pointing to command bandwidth and data coverage as critical factors for further improvement.