Deep Nonlinear Adaptive Control for Unmanned Aerial Systems Operating Under Dynamic Uncertainties

Recent literature in the field of machine learning (ML) control has shown promising theoretical results for a Deep Neural Network (DNN) based Nonlinear Adaptive Controller (DNAC) capable of achieving trajectory tracking for nonlinear systems. Expanding on this work, this paper applies DNAC to the Attitude Control System (ACS) of a quadrotor and shows improvement to attitude control performance under perturbed flying conditions where the model uncertainty is high. Moreover, these results are noteworthy for ML control because they were achieved with no prior training data and an arbitrary system dynamics initialization; simply put, the controller presented in this paper was designed with no knowledge about the underlying system, yet yields excellent trajectory tracking performance whilst rejecting undesirable model disturbances learned through a DNN. The combination of ML techniques to learn the system's dynamics and the Lyapunov analysis required to provide stability guarantees leads to an effective adaptive controller with applications to safety-critical systems that may undergo uncertain model changes, as is the case for most aerial systems. Experimental findings are presented in the final section of this paper, where DNAC is shown to outperform the trajectory tracking capabilities of PID, MRAC, and the recently developed Deep Model Reference Adaptive Control (DMRAC) schemes.