What Can Collective Construction Learn from Neural Cellular Automata?

Abstract

Neural Cellular Automata (NCA) have been trained to produce target images and shapes and even to regenerate after damage. These are highly attractive properties that can inform work on collective robotic construction. We discuss concepts from NCA that may be useful for collective robotic construction and discuss how the problems of morphogenesis and construction differ. As a concrete first step, we propose a simplified variant of an existing NCA model to explore the consequences of reducing the number of state channels encoded. We find that the NCA can still reproduce trained images. This bodes well for translating ideas from N CAs to collective robotic construction.