Development of a compact rotary series elastic actuator with neural network-driven model predictive control implementation

The development and control of a rotary series elastic actuator (SEA) are investigated in this paper. First, a rotary SEA is designed with a small volume and is lightweight, where the elastic element can be used as a torque sensor. We improve the structure of this rubber material elastic element, and its characteristics are analyzed. To provide a more comprehensive description of the entire system, motor dynamics are also taken into account while establishing the entire system dynamics model. Second, a neural network-driven model predictive control (NNMPC) method is proposed for the single-link SEA system. Since a real dynamic system for the SEA is hard to establish accurately due to disturbances, uncertainties, and varying mass of the load in different applications, a simple nonlinear autoregressive neural network using the rectified linear unit as the activation function (ReLU-NARX NN) is considered to approximate the system dynamic model, based on which a model predictive controller is developed. Finally, both numerical simulations and experiments are conducted for position and torque control. The simulation and experimental results demonstrate that the proposed method is superior to the conventional PD (proportional differential) method and the traditional MPC method. For position control, the NNMPC method is shown to be more effective, that is, it can suppress residual vibrations, reduce overshoots, arrive at a steady state quickly, and robust to different loads in a range. For torque control, the control performance is also satisfactory.