We developed a new experimental and simulative paradigm to study the establishing of compositional grounded representations for motor patterns. Participants learned to associate nonsense arm motor patterns, performed in three different hand postures, with nonsense words. There were two group conditions: in the first (compositional), each pattern was associated with a two-word (verb-adverb) sentence; in the second (holistic), each same pattern was associated with a unique word. Two experiments were performed. In the first, motor pattern recognition and naming were tested in the two conditions. Results showed that verbal compositionality had no role in recognition and that the main source of confusability in this task came from discriminating hand postures. As the naming task resulted too difficult, some changes in the learning procedure were implemented in the second experiment. In this experiment, the compositional group achieved better results in naming motor patterns especially for patterns where hand postures discrimination was relevant. In order to ascertain the differential effect, upon this result, of memory load and of systematic grounding, neural network simulations were also made. After a basic simulation that worked as a good model of subjects performance, in following simulations the number of stimuli (motor patterns and words) was increased and the systematic association between words and patterns was disrupted, while keeping the same number of words and syntax. Results showed that in both conditions the advantage for the compositional condition significantly increased. These simulations showed that the advantage for this condition may be more related to the systematicity rather than to the mere informational gain. All results are discussed in connection to the possible support of the hypothesis of a compositional motor representation and towards a more precise explanation of the factors that make compositional representations working.