Autonomous grasping is a complex task for robots. It is a high dimensional problem since it involves controlling for the hand position, orientation and joint angles to successfully grasp an object. In order to reduce the control complexity, we adopt a 3-step approach. In the first step, we compute several stable grasps that are adapted to the robotic hand using an optimization technique. In a second step, we extract postural synergies from this grasping data, project the grasps into these synergies subspace, and use this data representation to learn a distribution of the feasible grasps. The third step uses the learned model to generate quickly new grasps for the given object. Our approach was validated on the four degrees of freedom Barrett hand.