Control of quadrotor trajectory tracking with sliding mode control optimized by neural networks

Abstract

Quadrotor or unmanned helicopter is a mobile robot which often flies in unknown environment to perform special missions. In navigational tasks, the robot is intended to fly autonomously toward a target position by following an optimum trajectory. For a successful navigation, controlled attitude, minimum position and velocity error and obstacles collision avoidance are often considered during trajectory tracking procedure. By considering environmental variabilities and due to the existence of noises, uncertainties and unpredictable factors, it is indispensable to steer and control moving robots using sophisticated autonomous algorithms. In this work, a nonlinear model of four-rotor helicopter is simulated. An optimized terminal sliding mode control is then designed to control trajectory tracking. In order to improve the time indices for sliding mode controller, this controller is modified with neural networks. The idea is to optimize the controller parameters through a network learning process which is based on the control process error. The proposed method is evaluated with simulated and real-world indoor navigation tasks. Trajectories that are tracked by quadrotor are generated by a simultaneous localization and mapping algorithm and refined with an optimization technique. A well-known simultaneous localization and mapping technique (a camera-based extended Kalman filter-simultaneous localization and mapping) is employed to generate maps, and a path planning algorithm (particle swarm optimization) is utilized to optimize a collision-free flight path using the probability-based maps generated by simultaneous localization and mapping. Simulations and experiment are done in unknown but structured indoor environments containing a number of obstacles. The steady state error, the reaching and settle time and the chattering effect are all quantified and assessed. The controlled experimental flight robustness and sensitivity are further verified for noises occurred on vision and data acquisition system. Results indicate suitable performance for the proposed neural network-sliding mode controller. Less error and more stability were achieved comparative to the conventional sliding mode controllers.