I will discuss an experimental setup where robots equipped with ways to perceive and act in the world engage in language games. This setup is used to examine what kind of cognitive architecture, interaction patterns, and sensory inputs are needed to evolve natural-language like communication systems. More concretely, I will focus on the evolution of a grammar for case, i.e., the expression of event structure using case markings or similar constructs. The robots perceive dynamic scenes (such as a hand grasping a ball) and describe to each other what they see. They start with a purely lexical language (without any grammar) and then evolve a system of case marking which is similar to natural languages, in the sense that event structures are mapped onto semantic roles (agent, patient, etc), semantic roles to cases (nominative, accusative, etc.), and cases to surface markings. Semantic roles and cases emerge as part of the grammar evolution process and are not assumed to be innate.

The basic hypothesis underlying this experiment is that agents try to optimize their communicative success while minimizing the effort involved. They do this by adopting specific behaviors, such as always use the meaning-form relation that had the highest success in the past, extend meaning-form relations by analogy, overload existing form with additional meanings, etc. The simulations show that if these behaviors are adopted, grammar spontaneously emerges and spreads in the population. Thus semantic roles emerge because the agents re-use, by analogy, mappings between specific event-object relations and their markings. By re-use, prototypes for agent, patient, etc., gradually emerge. (Surface) cases emerge because agents try to pack more information into the surface form. This requires and additional level between semantic roles and their marking which again arises gradually.