In recent years, machine learning techniques have been increasingly applied to develop robust human character animation controllers. We present a novel application of reinforcement learning for the task of grasp synthesis in a physics-based virtual environment. A collection of low-dimensional controllers is employed to actuate the joints of a virtual hand model in coordinated motions. These controllers may be derived manually or from motion capture data. An exploration stage allows the agent to evaluate the state-action space of a dynamic environment, guided by an initial set of user-defined states. The approach is straightforward, interactive, and results in a control policy capable of synthesizing stable grasps. This talk will present preliminary results obtained by our method, followed by a discussion of ongoing and future work.