A Comparison of Multiobjective Algorithms in Evolving Quadrupedal Gaits

Robotic systems, whether physical or virtual, must balance multiple objectives to operate effectively. Beyond performance metrics such as speed and turning radius, efficiency of movement, stability, and other objectives contribute to the overall functionality of a system. Optimizing multiple objectives requires algorithms that explore and balance improvements in each. In this paper, we evaluate and compare two multiobjective algorithms, NSGA-II and the recently proposed Lexicase selection, investigating distance traveled, efficiency, and vertical torso movement for evolving gaits in quadrupedal animats. We explore several variations of Lexicase selection, including different parameter configurations and weighting strategies. A control treatment evolving solely on distance traveled is also presented as a baseline. All three algorithms (NSGA-II, Lexicase, and Control) produce effective locomotion in the quadrupedal animat, but differences arise in performance and efficiency of movement. The NSGA-II algorithm significantly outperforms Lexicase selection in all three objectives, while Lexicase selection significantly outperforms the control in two of the three objectives.