| Abstract: |  In mobile robotics it is necessary to predict a robot pose on a terrain to guarantee its stability when traversing an obstacle. Usual methods are based on an exact simulation of a robot-surface interaction, but this requires a precise physical model, which can be hard to solve or can be too much complex. The aim of this thesis is to propose and experimentally evaluate an algorithm, based on machine learning methods, which predicts attitude of the robot (roll and pitch angles) in natural environment. The main contribution of this work lies in development and evaluation of models, which can be used for predicting the safety of robot states and actions while interacting with the environment. Three models based on different multidimensional regression methods (linear, piecewise constant and Gaussian process) were trained and compared. As a part of this work, testing dataset was created and will be relased for the robotic community.
|
|---|