Abstract:
Development of modular robot control systems poses serious challenges associated with the robot’s construction subject to changes and the presence of a large number of degrees of freedom. The goal of this work was developing a versatile control system for modular hyper-redundant systems, able of independently finding ways to control robots with an arbitrary design from a certain given class. For solving the problem, a model of a control system was proposed, using logical-probabilistic knowledge discovery methods, adapted for control tasks. In accordance with the proposed approach, the task of control system training was reduced to finding patterns in an array of system’s environment interaction statistical data. For making the system independent on the chosen robot design, including modules’ spatial connection data specified in a data tree was proposed. Using this information during the training process allows the control system to independently tune in to control the robot, regardless of its design. For testing the proposed model’s performance and effectiveness, experiments in training a class of robots with different designs to move forward, which have confirmed both the learning rate and control quality being high.
Fulltext:
PDF file (390 kB)