Geometric Adaptive Neural Controller and Optical Flow-Based Invariant Extended Kalman Filter for Mars Quadrotor Under Disturbance

Abstract

With the continuous progress of deep space robotics technology and the deepening of human understanding of the Martian surface environment, the rotorcraft is expected to overcome the limitations of Mars orbiters and rovers in terms of exploration accuracy, range, and flexibility. Rotorcraft are poised to become essential vehicles for future deep space exploration missions. In this paper, an invariant extended Kalman filter (IEKF) and geometric adaptive neural controllers (GANC) are introduced for the state estimation and trajectory tracking of the Mars quadrotor. The IEKF fuses the IMU and depth camera information to estimate the optimal states of the quadrotor, which are used as the inputs of the trajectory tracking controller. Multilayer perceptron networks (MLP) and temporal convolutional networks (TCN) are designed and trained to predict the force and torque disturbances to improve the tracking accuracy and robustness of the proposed geometric adaptive neural controllers, including the feedforward (FF) proportional-differential (PD) neural controller on SE(3) for large-angle maneuver and the differential-flatness based neural controller (DFBC) for flight in strong winds. Especially, for underpowered situations, an event-trigger neural model predictive contouring controller (ET-NMPCC) is designed to optimize the control inputs, foresee and adapt to potential future external forces, and ultimately achieve higher trajectory tracking accuracy. Physical simulation systems are established to mimic the terrain and atmosphere of the Martian surface in Webots and Airsim simulator. The simulation and real-world experimental results show the effectiveness and superiority of these methods in flight navigation and control performances of the quadrotor on the Martian surface.