Impact of Morphology Variations on Evolved Neural Controllers for Modular Robots

Modular robots, in particular those in which the modules are physically interchangeable, are suitable to be evolved because they allow for many different designs. Moreover, they can constitute ecosystems where “old” robots are disassembled and the resulting modules are composed together, either within an external assembling facility or by self-assembly procedures, to form new robots. However, in practical settings, self-assembly may result in morphologies that are slightly different from the expected ones: this may cause a detrimental misalignment between controller and morphology. Here, we characterize experimentally the robustness of neural controllers for Voxel-based Soft Robots, a kind of modular robots, with respect to small variations in the morphology. We employ evolutionary computation for optimizing the controllers and assess the impact of morphology variations along two axes: kind of morphology and size of the robot. Moreover, we quantify the advantage of performing a re-optimization of the controller for the varied morphology. Our results show that small variations in the morphology are in general detrimental for the performance of the evolved neural controller. Yet, a short re-optimization is often sufficient for aligning back the performance of the modified robot to the original one.