A new improvisation algorithm based on parameter estimation process for the imitation control of a robotic acoustic musical device '' ROMI '', is presented in this paper. The musical state representation we have developed for controlling ROMI is a feature extractor for learning to imitate human players in a duet. ROMI ' s intelligent control architecture also has the ability to provide player identification and performance training. In this paper we introduce the robotic device ROMI together with its control architecture, the musical state representation and focus on parameter estimation for imitation of duo players by ROMI. ROMI is aimed at jointly playing two instruments that belong to two different classes and improvises while assisting others in an orchestral performance.