Evolving Soft Robots

Providing evolution with a larger palette of materials to work with leads to amazingly fun, quirky, lifelike virtual creatures. The bodies of the Softbots are composed of voxels with four different material properties. Voxels may be soft, rigid, or can rhythmically expand and contract with two opposing phases. Body structures formed by the voxels allow for the generation of movement through the interactions of deforming voxels.

Using the CPPN-NEAT generative encoding, the bodies are evolved to produce robots capable of walking in a physics simulation. Evolution allows for a diverse range of solutions on the walking problem by combining voxels to form different classes of body types. The phenotypes produced by CPPN-NEAT exhibit regularity in physical structure and behavior. The diverse range of phenotypes produced by evolution exhibit similarities to real-world biological structures, with organisms evolving locomotion methods such as legs. The results demonstrate the capabilities of regular evolution using CPPNs in conjunction with increasingly complex simulations.