Transformer-Based Soft Robot Constrained Position Recognition with Multimodal Sensing

This study initially investigates the potential issues associated with single-modality sensing in soft robotic systems and the variations in multimodal sensing weights for robot state recognition under different conditions. Subsequently, we construct a transformer-based model for the recognition of constrained states in robots and compare its performance with several alternative methods, including LSTM, GRU, BiLSTM, and CNN-Att-LSTM.Experimental results show that the Transformer model outperforms methods such as LSTM, GRU, BiLSTM, and CNN-Att-LSTM in terms of accuracy and reliability, with an accuracy rate of over 90% in distinguishing between constrained and unconstrained states of soft robots.