In this paper, we focus on robot intelligence to generate turn-taking strategies in response to human play actions. This work builds on our previous work on play behavior recognition, and expands it to the child-robot therapeutic domain where the robot must understand and learn the play of a child and take turns manipulating the toys. The main contribution of this work is a novel attempt in applying Case-Based Reasoning (CBR) for planning human-robot turn-taking strategies. By comparing the child's play in the current scene to some past play cases stored in memory, we retrieve the best solution and adapt it to the set of toys that are available for the play scenario, bypassing a long complicated decision process. In order to ensure real-time performance, a low dimension scale invariant shape descriptor is proposed for shape matching. Turn-taking CBR (ttCBR) system is then evaluated for stacking and inserting tasks with four subjects, by comparing the decision made by the system and the actual choice of the humans.