Developing Grammars in Embodied Situated Language Games. Luc Steels VUB AI Lab, Brussels and SONY CSL Paris. steels@arti.vub.ac.be The paper further explores the hypothesis that grammatical constructions primarily arise because a speaker seeks to maximise the chance of communicative success and the expressive power of her utterances. New lexical or grammatical constructions are introduced when parts of the meaning are not yet covered or when the available linguistic material may lead to a failure or a risk of failure in communication. When verbal interactions are grounded and situated, the hearer has a good chance to be able to infer the meaning of novel expressions, reconstruct the underlying constructions, and integrate them into her own repertoire. The paper also explores the hypothesis that meanings and their expression co-evolve. New meanings arise because the speaker needs to make distinctions which were not made before. These distinctions become lexicalised and grammaticalised to achieve success in communication. The hearer acquires new meanings by reconstructing these distinctions while trying to create hypotheses for the meaning of unknown grammatical constructions. These hypotheses sound entirely reasonable but the big challenge is to work them out in technical detail and show that their cummulative effect leads to languages with natural language-like properties. We have been constructing various computer simulations trying to do this, and engaged in experiments with robotic agents playing situated language games. Our focus has specifically been on how grammars for case and tense could self-develop in a group of autonomous embodied agents. In the present paper, I focus on the core of the cognitive mechanisms responsible for grammatical development in these experiments: (1) a mechanism used by the speaker for detecting potential uncertainty in communication, (2) a mechanism used by the speaker for inventing a new grammatical construction to alleviate such an uncertainty, )3= a mechanism used by the hearer to detect the meaning of a new grammatical construction introduced by the speaker. The main point of the paper is that the mechanisms are generic and generalise across the domains of case and tense.