Investigating Modular Coupling of Morphology and Control with Digital Muscles

Abstract

The musculoskeletal systems of animals are governed by a complex network of neurons that define both high- and low-level control. Individual joints are manipulated by multi-ple muscles acting as effectors for both movement and sta-bilization. We previously proposed a digital muscle model (DMM), where the morphological and control aspects of sim-ulated joints evolve concurrently. The resulting solutions can provide insight into the evolution of natural organisms as well as possible designs for engineered systems. In this paper, we explore the integration of this model with an arti-ficial neural network (ANN), focusing on the communication connections between the two. In the singly-connected strat-egy, a single ANN output is delivered to a joint; each con-stituent muscle responds to the signal according to an evolved function. In the individually-connected strategy, a unique ANN output is delivered to each simulated muscle. Results indicate that for low degree-of-freedom (DOF) robots, the individually-connected systems exhibit higher fitness than the singly-connnected systems. However, in larger DOF robots, the two strategies perform comparably, despite the fact that evolved ANNs for the singly-connected system are consid-erably simpler in terms of the number of connections in the network.