Reinforcement learning control of a humanoid robotic hand actuated by shape memory alloy

Humanoid robotic hand actuated by shape memory alloy (SMA) represents a new emerging technology. SMA has a wide range of potential applications in many different fields, ranging from industrial assembly to biomedicine applications, due to the characteristic of high power-to-weight ratio, low driving voltages and noiselessness. However, nonlinearities of SMA and complex dynamic models of SMA-based robotic hands result in difficulties in controlling. In this paper, a humanoid SMA-based robotic hand composed of five fingers is presented with the ability of adaptive grasping. Reinforcement learning as a model-free control strategy can search for optimal control of systems with nonlinear and uncertainty. Therefore, an adaptive SA-Q-Learning (ASA-Q-learning) controller is proposed to control the humanoid robotic finger. The performance of ASA-Q-learning controller is compared with SA-Q-learning and PID controller through experimentation. Results have shown that ASA-Q-learning controller can control the humanoid SMA-based robotic hand effectively with faster convergence rate and higher control precision than SA-Q-learning and PID controller, and is feasible for implementation in a model-free system.