Trajectory Tracking of Yoshimura-Origami Continuum Manipulator with LSTM-Attention Network

This paper presents a dynamic target tracking method for a Yoshimura-origami continuum manipulator (YoMo). Unlike previous studies that focus on fixed trajectory tracking, the proposed approach handles moving targets in real time. An LSTM-Attention(AttnLSTM) network is introduced as the inverse model to map desired end-effector positions to cable length commands. The YoMo’s hardware structure is briefly described, and data collection is conducted through Latin hypercube oversampling. Four modeling methods—RNN, LSTM, LSTM with Attention, and Transformer—are compared through comprehensive experiments. Results demonstrate that the AttnLSTM model achieves superior performance in both regular trajectory tracking and dynamic target following, with tracking errors as low as 1.31 mm for smooth trajectories and 4.63 mm for complex paths. The AttnLSTM model outperforms other architectures with lower RMSE and better stability in handling abrupt direction changes. The feasibility and effectiveness of the proposed approach are verified through dynamic target tracking experiments in multiple scenarios.