In recent years, soft robotics has emerged as a key area of research in modern robotics. Among structural solutions, robotic arms play a key role. From the perspective of actuation, soft arms can be classified into two main types: cable-driven and pressure-driven arms. The latter often utilize pneumatic artificial muscles, whose characteristics — such as hysteresis and high degrees of freedom—make traditional dynamic modeling methods ineffective. This work focuses on the application of recurrent neural networks (LSTM, GRU), which are well-suited for processing time-series data. To improve model accuracy, a Curriculum Learning strategy with varying training block sizes is applied, enabling gradual learning. The proposed methods are evaluated on the task of predicting the X-coordinate of the soft arm’s endpoint. Results confirm the effectiveness of the selected approach and demonstrate the positive impact of Curriculum Learning on the ability of RNNs to model the dynamic behavior of soft robotic arms.
Science Journal 