In order to adapt to complex and changeable mechanical conditions and make the deformable mechanisms perform well statically and dynamically, variable stiffness joints have been studied extensively. The variable stiffness actuator is the key driving component to adjust stiffness of the joint. By inserting flexible elements between the driving and driven ends of rigid motion, the variable stiffness actuator makes the joint move precisely and allows humans to interact with machines safely. At present, many kinds of variable stiffness actuators have been applied, among which the way of changing the length of the force arm of leaf springs has obvious advantages. However, overall configuration design, accurate stiffness model, mechanical characteristics and safety analysis have not been studied in depth. This paper investigates a variable stiffness actuator based on leaf spring by design, model and mechanical analysis. The composition and configuration of the actuator is analyzed and optimized. Using the deflection theory of the beam, a new rotational stiffness model of the actuator is established, and a safe position criterion is set up upon the deformation constraint conditions. The variation law of stiffness and the influence of parameters on mechanical characteristics are studied. The finite element analysis method verified the rotational stiffness model, and static test proved that the actuator could effectively work in the joint.

Shape memory alloy (SMA) has been widely used in different applications due to its unique shape memory property. However, when used as an actuator, it exhibits a hysteresis behavior in its relation between temperature and strain, which is highly nonlinear and difficult to control. Although studies have been conducted on establishing various constitutive models of SMA, it is still difficult to achieve the precise control of the SMA wire with the existing models. In this work, a new promising approach regarding the SMA control task as a reinforcement learning (RL) problem is proposed to address this issue, which does not require accurate mathematical models. Both RL and an improved method named deep reinforcement learning (DRL) are used to solve the problem of precise control of a 1D SMA wire actuator, respectively. The simulation results indicate that with the DRL method, the agent can precisely control the output deformation of the SMA wire after only ten episodes of training. Compared with the DRL method, the RL agent can also achieve the same training target but with hundreds of training.