1 Abstract—The endpoint toward which reconfigurable systems should develop is programmable matter, that is, complex systems whose physical properties and structure can be controlled in a systematic way. This can be accomplished by recognizing that computational processes can be used to assemble and reconfigure complex, hierarchically structured systems. Programmable matter may be programmatically controlled externally or internally, which includes self-assembly. The best approach to the self-assembly of complex, hierarchical systems, such as future robots with capabilities comparable to those of animals, is by artificial morphogenesis, which adapts embryological morphogenesis to artificial systems. We review the requirements of self-assembling morphogenetic components. Index Terms—artificial morphogenesis, assembly systems, microassembly, micromechanical devices, molecular communication, multi-agent systems, multi-robot systems, nanofabrication, programmable matter, reconfigurable architectures, robot control, self-assembly, synthetic biology. I. MOTIVATION n this paper we consider programmable matter as the natural extension and goal of reconfigurable systems. By pursuing this ultimate goal, we will achieve more widely reconfigurable systems along the way. What could we accomplish with programmable matter? One application of programmable matter is the assembly of complex, hierarchical systems, that is, systems with levels of complex structure from the micro (or even nano) scale up to the macro scale. The need to assemble such complex, hierarchical structures is illustrated by future robots with cognitive, perceptual, and physical competence comparable to that of animals. Consider the challenge of building a cat-size robot with the behavioral competence of a cat: able to run, leap, stalk and catch prey, etc. Among other problems, consider the sensors, effectors, and artificial nervous system of such a robot. The human retina has about 100 million sensory cells, intricately wired to reduce the dimension of the input to about one million for transmission on the optic nerve. Therefore, to produce an artificial retina we might want to assemble and interconnect some millions of simple sensors. B. J. MacLennan is with the Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN 37996 USA (e- mail: maclennan@utk.edu). * Accepted version of MacLennan, B. J., “The Morphogenetic Approach to Programmable Matter,” Proc. IEEE, Vol. 103, No. 7 (July 2015): 1226–1232. ©2015 IEEE. DOI: 10.1109/JPROC.2015.2425394 Sophisticated robots with dexterous manipulators will require a delicate sense of touch and other bodily senses. Therefore, we will want to populate their surfaces (“skins”) with dense arrays of sensors and to connect them to their central processors (“brains”). Animals’ physical competence arises in part from their complex systems of muscles and tendons, which work synergistically to permit a wide variety of fluent, rapid, and robust physical actions. To accomplish this, we need to be able to assemble complex arrays of actuators and to connect them to facilitate synergistic control. Such sophisticated sensors and effectors must be controlled in real time to achieve competent behavior, and one promising approach is massively parallel neural-network-style computing. (For example, brain-scale low-power electronic neuromorphic computers is the goal of a large DARPA program called SyNAPSE for “Systems of Neuromorphic Adaptive Plastic Scalable Electronics.”) But the scale of parallelism is significant: 86 billion neurons in a human brain, 763 million in a cat brain. Furthermore, these neurons are intricately connected, with many neurons having tens of thousands of connections, and some with several hundred thousand. For achieving brain-scale intelligence, we may need to assemble artificial neural networks of comparable complexity. The point of this example is that we would like to be able to assemble complex systems with millions (or more) components in a hierarchical structure (that is, not simple homogeneous or periodic structures). Not only would we like to assemble such systems from scratch, we would also like to be able to reconfigure complex systems, rearranging their components to address a new mission, recover from damage, improve their capabilities, etc. Typically reconfiguration entails changing the connections among a fixed set of elementary components. Radical reconfiguration goes further, altering the components themselves to create different hardware resources. If programmed assembly is like the development of a fetus, then radical reconfiguration is like the metamorphosis of a caterpillar into a butterfly. Both assembly and reconfiguration are facilitated by programmability, by which we mean the ability to govern a large class of processes in some uniform way. In the first case, assembly, a program is used to control a general-purpose mechanism for assembling some complex, hierarchical system. Hierarchical structure in the program may be used to organize hierarchical structure in the product. In effect, the program “computes” a physically instantiated structure The Morphogenetic Path to Programmable Matter * Bruce J. MacLennan, Member, IEEE I