Evolutionary Legged Robotics

Due to the technological advance, robotic systems become more and more interesting for industrial and home applications. Popular examples are given by robotic lawn mower, robot vacuum cleaner, and package drones. Beside the toy industry, legged robots are not as popular, although they have some clear advantages compared to wheeled systems. With their flexibility concerning the locomotion, they are able to adapt their walking pattern to different environments. For instance they can walk over obstacles and gaps or climb over rubble and stairs. Another possible advantage could be a redundancy for locomotion. A faulty motor in one limb could be compensated by other motors in the kinematic chain. As well, multiple failing legs can be compensated by an adapted walking pattern. Compared to this, the more complex mechatronic systems represent a major challenge to the construction and the control. This thesis is dedicated to the control of complex walking robots. Genetic algorithms are applied to generate walking patterns for different robots. The evolutionary development of walking patterns is done in a simulation software. Results of various approaches are transferred and tested on existing systems which have been developed at RIC/DFKI. Different robotic systems are used to evaluate the generality of the applied methods. Eventually, a method is developed that can be utilized, with a few system specific modifications, for a variety of legged robots. As basis for the development and investigation of several methods, software tools are designed to generalize the application of applying genetic algorithms to legged locomotion. These tools include a simulation environment, a behavior representation, a genetic algorithm and a learning and benchmark framework. The simulation environment is adapted to the behavior of real robotic systems via reference experiments. In addition, the simulation is extended by a foot contact model for loose surfaces. The evaluation of the genetic algorithm is done on several benchmark problems and compared to three existing algorithms. This thesis contributes to the state of the art in many areas. The developed methodology can easily be applied to several complex robotic systems due to its transferability. The genetic algorithm and the hierarchical behavior representation provide a new opportunity to control the generation of the offspring in an evolutionary process. In addition, the developed software tools are an important contribution for their respective research fields.