Evaluation of UWB positioning on a mobile robot using precise optical tracking system

Knowing the pose of a robot is essential for the successful execution of most of the tasks it usually performs. Consequently, mobile robot localization becomes a crucial problem to address. Odometry is a commonly used technique for the localization of Unmanned Ground Vehicles (UGVs). It exploits the information coming from the wheel encoders and the inertial measurement unit (IMU) to estimate the pose of the robot. However, it is very sensible to slippage, and, since it calculates the position by integrating the encoders' data, it accumulates errors, becoming more and more inaccurate in short time. For GPS-denied zones and indoor environments, Ultra-Wideband (UWB) seems to be a promising technique for the precise tracking of moving objects. The huge bandwidth (>500 MHz) allows high resolution in time and consequently in range too, leading to a positioning error of less than 30 cm. This study aims to implement and analyze different localization algorithms that rely only on UWB data for the estimation of the position of a rover. Three different algorithms for nonlinear problems have been implemented, ranging from standard approaches to more innovative ones. Firstly, two of the mainly used probabilistic algorithms have been developed, the Extended Kalman Filter (EKF) and the Unscented Kalman Filter (UKF). While the EKF linearizes the state and the measurement model at the point of current estimate, the UKF uses a set of sigma points to better capture the nonlinearity present in the models, making it more robust against high nonlinearity. An alternative to the more classical and widely accepted probabilistic algorithms is represented by the combination of a Neural Network (NN) with the EKF. The main idea is to make the EKF adaptive by adding a NN to better estimate the error on the measurements coming from the UWB antennas at each time step, and then use this estimation, converted into variance, in the filter. Thus, using different variance values for the various situations allows to better understand which measurements to trust more and which not, potentially leading to better localization results for the robot. A fundamental step for the performance evaluation of the different algorithms and the training of the NN is the creation of a new dataset. Therefore, all experimental tests were conducted using the Jackal UGV, equipped with several sensors, including an IMU, to get information about the orientation, acceleration, and angular velocity of the rover, the wheel encoders for the linear and angular velocity, and a UWB antenna, to get the distances between the tag and each of the four anchors. Also, a Vicon system was used as ground truth. The data coming from the sensors was collected both under Line of Sight (LOS) and Non-line of Sight (NLOS) conditions, where different obstacles in different positions were used in order to test their effect on the radio performance.