Trajectory control for a quadrotor unmanned aerial vehicle: Adaptive super-twisting terminal sliding mode with adjustable recurrent neural network

This paper explores how to control the trajectory of a quadrotor UAV (Unmanned Aerial Vehicle) in unpredictable environments with external disturbances. We address the challenge of designing a controller when the UAV’s mass and inertia are unknown, which makes real-time modeling difficult. To solve this problem, we developed an adjustable recurrent neural network (ARNN) that more accurately approximates the necessary control actions. There are actually some problems when using an RNN in the design of UAV control algorithms: it produces insufficiently accurate control approximations, it is difficult to generalize across different tasks for various UAVs, and the neural network’s own gradient disappears during the training process. To improve its performance, we designed the ARNN with a flexible activation function controlled by an adjustable parameter, which improves its adaptability to different data types and reduces training problems. We also refined the self-feedback mechanism to increase the accuracy of the control approximation. The whole system combines a super-twisting sliding mode control algorithm with the ARNN. We introduce a new super-twisting algorithm that accelerates convergence and reduces the chattering problem in sliding mode controllers through an exponential nonlinear term. Using Lyapunov functions and the Lassalle invariance principle, we show that our method ensures global convergence in finite time. Simulation results confirm the effectiveness and advantages of our approach for UAV trajectory tracking.