Design of a hybrid-mode piezoelectric actuator for compact robotic finger based on deep reinforcement learning

Abstract

In this study, we propose a novel hybrid-mode piezoelectric actuator for compact robotic finger and introduce a reinforcement learning-based design approach using Double Deep Q-Networks (Double DQN) for design of the hybrid-mode piezoelectric actuator. The experiment demonstrated that the Double DQN algorithm could learn a robust policy that minimized the modal frequency difference, achieving a minimum difference of 6 Hz, which was verified through FEM simulations. Due to the broad applicability of the modal frequency degeneracy requirement, this method provides a new approach for designing various hybrid-mode piezoelectric actuators. Furthermore, the robotic finger prototype, while maintaining a compact size, achieves an angular velocity of 330 deg/s and a fingertip force of 0.32 N under a driving voltage of 400 Vpp. These experimental results validate the effectiveness of the proposed driving method. Furthermore, the proposed actuator is well-suitable for modular assembly, allowing multiple actuators to be easily connected to form multi-joint and multi-finger structures, enabling the design of dexterous robotic hands. In future research, we plan to enhance the output force by increasing the number of PZT elements or employing a sandwiched design. Our findings highlight the potential of piezoelectric actuators, combined with reinforcement learning techniques, in the development of compact robotic hands.