Learning to Grasp Unknown Objects using Force Feedback

Most robotic manipulator includes end effector that grasp object securely and perform useful tasks. Modern gripper designs employ force sensor arrays for force feedback. In order to control a manipulation task, the contact kinematics, applied forces, and dynamics of rigid bodies must be understood and defined with a model. In this paper, we propose the use of machine learning algorithm for gripper controller, without a predefined grasping model. Q-learning algorithm is applied to a gripper controller whereby the agent learns to grip or release based on the force sensing signals. We propose different approaches to formulate the Markov states in this problem. A general and progressive approach is chosen based on our gripper hardware. Experimental results show that it takes few minutes to learn an optimal grasping behavior on-line. Several new objects were successfully tested using the Q-learning controller to evaluate its grasping stability.