Submodel 1 CNN Block 1 CNN Block 2 Submodel 2 CNN Block 1 CNN Block 2 (a) Ultrasound CNN Block 1 CNN Block 2 (b) CNN Block 1 Conv-6 Conv-1 Conv-2 Pooling-1 Pooling-2 Conv-3 Pooling-3 Conv-4 Pooling-4 Conv-5 Pooling-6 Conv-7 Pooling-7 Conv-8 Pooling-8 Conv-9 Pooling-9 Pooling-i Pooling-5 Conv-10 Pooling-10 MaxPooling2d (c) (d) (f) Figure 6. (a) The MMCNN model structure for extracting deep features of sEMG signals; (b) the MCNN model for extracting deep features of ultrasound signals; (c)-(f) show the detailed structure of CNN blocks 1 and 2, and the convolution and pooling layers, respectively. CONV: convolutional. 16 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022 Dropout CNN Block 2 Conv-i Separable Conv2d (e) BatchNorm2d + ELU Spectrogram sEMG Time-Domain sEMG FFT Concatenaten Conv-12 Pooling-12 Conv-13 Pooling-13 LR-UI FC1 FC2 FC1 FC2 Conv-12 Pooling-12 Conv-13 Pooling-13 LR-EMG Conv-11