An Improved System for Artificial Creatures Evolution

We present our complete reimplementation of Karl Sims’ system for evolving and coevolving autonomous creatures in a physically realistic three-dimensional (3D) environment. Creatures are articulated structures composed of rigid blocks and controlled by embedded neural networks. The main differences with Sims are, first, the use of standard McCullochPitts neurons (instead of a set of ad hoc, complex functional neurons) and, second, an improved genetic encoding and developmental system (allowing fine-grained control of neural connections in duplicated morphological features, and replication-exaptation processes). This paper expands upon a previous version of our system (Miconi and Channon, 2005) which implemented a subset of features present in Sims’ system, and dealt with simple evolutionary experiments based on external fitness functions only: the present paper extends the feature set proposed by Sims, and describes the results of experiments based on the ‘box-grabbing’ coevolutionary task introduced by Sims. We provide a detailed description of our model and freely accessible source code. We describe some of our results, including an analysis of evolved neural controllers. To the best of our knowledge, our work is the first replication of Sims’ efforts to achieve results comparable to Sims’ in efficiency and complexity, with standard neurons and realistic Newtonian physics.