Evolutionary discovery of self-stabilized dynamic gaits for a soft underwater legged robot

Abstract

In recent years a number of robotic platforms have been developed, that are capable of robust locomotion in presence of a simple open loop control. Relying on the self-stabilizing properties of their mechanical structure, morphology assumes a crucial role in the design process, that is, however, usually guided by a set of heuristic principles falling under what is commonly known as embodied intelligence. Despite many impressive demonstrations, the result of such a methodology may be sub-optimal, given the dimension of the design space and the complex intertwining of involved dynamical effects. Encouraged by the growing consensus that embodied solutions can indeed be produced by bio-inspired computational techniques in a more automated manner, this work proposes a computer-aided methodology to explore in simulation the design space of an existing robot, by harnessing computational techniques inspired by natural evolution. Although many works exist on the application of evolutionary algorithms in robotics, few of them embrace this design perspective. The idea is to have an evolutionary process suggesting to the human designer a number of interesting robot configurations and embodied behaviors, from whose analysis design hints can be gained to improve the platform. The focus will be on enhancing the locomotion capabilities of a multi-legged, soft, underwater robot. We investigate for the first time the suitability of a recently introduced open-ended evolutionary algorithm (novelty search) for the intended study, and demonstrate its benefits in the comparison with a more conventional genetic algorithm. Results confirm that evolutionary algorithms are indeed capable of producing new, elaborate dynamic gaits, with evolved designs exhibiting several regularities. Possible future directions are also pointed out, in which the passive exploitation of robot's morphological features could bring additional advantages in achieving diverse, robust behaviors.