Insights in evolutionary exploration of robot morphology spaces

Abstract

In a recent study we have encountered an unexpected result regarding the evolutionary exploration of robot morphology spaces. Specifically, we found that an algorithm driven by selection based on morphological novelty explored fewer spots in the space of morphologies than another algorithm based on a combination of morphological novelty and some behavioral criterion (speed of movement). Here we revisit these results, perform new analyses, and obtain new insights. These insights clarify the exploration behavior of these algorithms and provide guidelines for designing selection mechanisms for evolutionary robotics.

Environmental influences on evolvable robots

“…In previous work [16,[28][29][30][31] we carried out a handful of investigations regarding the same robot framework utilized in this present paper. Firstly, we evolved robots that were "isolated" from any environmental influences, using morphological novelty search as a search criterion, which means the fitness function did not account for their performance on the task (for example locomotion on a flat terrain environment).…”

Effects of environmental conditions on evolved robot morphologies and behavior

“…A group of studies [15][16][17] has carried out a handful of investigations regarding the same robot framework utilized in this present paper. They evolved robots "isolating" them from any environmental inuences, through performing morphological novelty search, having the tness function to ignore their performance on the task (locomotion on a at terrain environment).…”

“…Descriptors 1 to 4 are simple counts; for more details about the calculation of descriptors 5 to 12 we refer to [16,17]. Additionally, several analyses of our modular robot framework and its descriptors are available in [15][16][17], demonstrating the capacity of these descriptors to capture relevant robot properties.…”