One of my favorite research topics is artificial life. Although this paper is not recent, Karl Sims published Evolving virtual creatures at SIGGRAPH ’94 and I still enjoy watching the video of these body shapes:
I don’t think it’s technically legal for me to post parts of the paper here without permission, so I hope you have a subscription to ACM. Anyways, he uses some relatively basic directed graphs to model the creature’s body. Each node is represented by a rectangular prism, and edges are the physical connections between them. The mutations can occur by altering the internal properties of a node, adding/removing a node in the graph, altering the properties of connections between nodes and adding/removing connections between nodes. The vestigial components–unconnected nodes–are discarded. The resulting graph is the creature’s genotype.
Mating also occurs between the creatures, where the graphs are combined by simulating the biological concepts of crossover and grafting. When represented in a virtual environment, the graphs become “physical” creatures, the body phenotype. Further, the control of these mechanics are determined by artificial neural networks, also fueled by the evolutionary process, the brain phenotypes. Sims uses basic fitness functions to choose for species that can swim quickly or grab an object. The result, though over sixteen years old, is a very cool adaptive process to both “aquatic” and “terrestrial” environments.