Aplikování strojového učení při návrhu kinematických struktur robotů

Abstract

The dissertation deals with optimizations of robot kinematics for a task known in advance, using knowledge from the field of evolutionary robotics and machine learning. Evolutionary robotics aims to create robots and their controllers using methods inspired by the evolution of natural organisms, such as optimization by genetic algorithms and controlling the robots with neural network. Machine learning deals with algorithms and techniques that work with a mathematical model, which through a learning process is able to solve problems that are often difficult to define. Evolutionary robotics has been experimenting with methods for the automatic design of robots since the 1980s, but the research usually does not leave laboratories and is hardly used in practice. This may be due to the computational complexity of the methods used, where the optimization of the robot's kinematics takes too long and with uncertain results for practical use. Knowledge from the field of machine learning could speed up this robot design process and help to get closer to a practical engineering tool. The focus of the thesis is on a synthesis of a robot manipulator, optimized for a specific task. The tasks are defined as trajectories composed of target points. The opening chapters analyze the current state of the art. A summary of two published methods follows. The summary includes methods that synthesize robot kinematics, and the findings of this research are described. Following chapters define a method for a task-based synthesis of robot manipulators using a genetic algorithm in the MATLAB environment. Emphasis is placed on the correct definition of the optimized vector i.e. genotype and on the definition of the objective function, according to which the manipulator is evaluated. The results of the method are presented on several example tasks. A proposal to verify the results on a real modular manipulator is described next. Lastly, a method that uses neural networks for rapid robot kinematics design is presented.