UC Berkeley

A reliable, accurate, and robust robotic system is highly dependent on perception, the ability of a robot to sense and interpret its environment. A variety of sensing technologies and methods can be integrated for perception purposes depending on a particular robotic setting and the surrounding. Uncertainty, environment variability, and limited sensing capabilities are factors that pose challenges to the perception task. This dissertation focuses on exploiting geometric and probabilistic characteristics as well as hidden structural properties of robotic systems and their surroundings to address some of these challenges.\par A geometric state estimator is presented for an agent with the ability to measure single ranges to fixed points (anchors) in its environment. The state estimator is generic, and can be immediately applied to any robot with the range sensor. A greedy optimization algorithm is developed to select the best measurement in each time step. The selection algorithm is added to the extended Kalman filter, resulting in choosing the best measurement out of all the available range values. The effectiveness of the presented estimator algorithm is demonstrated through experimental setup for a flying robot. \par The estimation accuracy is improved under the assumption that the ranging infrastructure is not perfect. A real\textendash time restructure of the setup allows to enhance the localization accuracy of the ego agent. The estimator is incorporated into an adaptive algorithm. Using a mobile UWB ranging sensor, the mobile anchor moves to improve localization accuracy of the main robot. The algorithm reconstructs the range sensor network in real\textendash time to minimize the covariance matrix in the extended Kalman filter. The presented algorithm is experimentally validated in a network of range sensors. \par A probabilistic\textendash based approach for pose estimation using point clouds is presented. The point registration algorithm is based on \emph{directional statistics}, which estimates the rigid transformation (i.e. rotation matrix and translation vector) between two point cloud frames. The algorithm outputs the robot's pose estimation (location and orientation). The framework transforms the point registration task on a unit sphere, and solves the problem in two steps of \emph{correspondence} and \emph{alignment}. In particular, a mixture model (as an example of directional statistics on unit sphere in $\mathbb{R}^3$) is adopted and the process of point registration has been carried out by the two phases of Expectation\textendash Maximization algorithm. The method has been evaluated with point clouds from LiDAR sensors in an indoor environment. \par A deep graph network is presented, to improve the robustness and accuracy of point registration. The framework models the point registration task based on the flexible architecture of Graph Network (GN) blocks. Three main modules\textendash an encoder, a core, and a decoder\textendash are responsible to perform both steps of correspondence and assignment in point matching process. The experiments and examination of the proposed model shows comparable results with other state\textendash of\textendash the\textendash art geometric\textendash or learning\textendash based algorithm in terms of accuracy as well as robustness with regard to bad initial conditions and presence of outliers in data points. The flexibility and configurability of the framework allows to easily change, add, and/or combine various customized deep modules and mechanisms to the presented graph\textendash based framework. \par The last part of this dissertation, studies ReLU network architecture in the domain of control. The input/output domain and structure of the network and its proximity to explicit Model Predictive Control (eMPC). The mathematical equivalency of feedforward ReLU and piecewise affine function is presented, and we investigate the prospect of representing state feedback policy of eMPC as a ReLU DNN, and vice versa. A sampling based method has been developed to identify input\textendash space regions in ReLU networks.