M-Estimator Application in Real-Time Sensor Fusion for Smooth Position Feedback of Heavy-Duty Field Robots

Abstract

In this paper, we study the performance of a complementary filter with adaptive weights in a sensor fusion application for real-time localization of an omnidirectional field robot. The test-case robot is a large, four-wheel drive and steer (4WDS) construction vehicle with nonlinear internal dynamics and hydraulic driving and steering actuators. Our objective is to provide the vehicle's real-time controller with robust, smooth feedback that prevents unnecessary oscillations in steering, which can waste significant amounts of energy. We do so by assigning weights for measurements based on their consistency with the robot's motions. The calculations are based on two main data sources: (1) measured velocity vectors from wheel driving (odometer) and steering of the 4WDS test-case robot; and (2) data obtained from a differential global navigation satellite system on the absolute pose of the robot. We show that the sensor fusion is robust to the noise and single point failures of the sensors, while the maximum heading oscillations are reduced by 70%-95%, preserving the accuracy of the global positioning system. Moreover, we demonstrate the feasibility and efficacy of the real-time implementation of this filtering method in path-following control of the robot.