Deep Learning Sensor Fusion Positioning of 4-wheel Robot

Abstract

This research entails the development of a deep learning-based framework to localize a steerable mobile robot using the following sensors: two-dimensional laser scanner, inertial measurement unit (IMU) and wheel encoders. Current methods for sensor fusion such as the Extended Kalman Filter (EKF) which rely heavily on linear approximations and sensor input quality are sensitive to large angular rotations and accelerations resulting in increased error. The results of this work will provide an accurate deep learning framework for precise nonlinear positioning. The localization problem is structured as a time series pose regression problem. Estimates for pose are provided using two successive laser scans and a designed Recurrent Convolutional Neural Network (RCNN) to fuse the remaining sensors which receive measurement data at varying rates. This framework is composed of one convolutional network block for the laser scanning data alongside several recurrent network blocks to accommodate the remaining sensors. Furthermore, the robot’s selected trajectory is decomposed into sequences for computational efficiency during investigation. The framework is first developed in simulation using MATLAB with PyTorch and then a physical experiment is performed using a front wheel steering, rear wheel drive robot, known as the Patrolbot outfitted with two LIDARs, wheel encoders and IMU with positioning results compared to the ground truth. It was seen that the global pose of the robot accumulates error and suffers from drift over time but can be improved with an optimization implementation comparing position points to a generated submap which is planned as a future research direction. This research has various applications and can be used to localize robots in a laboratory setting with moving obstacles (environment obstacle position varies) or for the robot to successfully park itself at a goal point with high accuracy. In addition, there minimal research investigations in the area of deep learning sensor fusion with virtually no investigations performed prior to 2021.