Manipulator Tactile Recognition and Classification Based on LZG-Net
Accurate recognition of object categories and tactile signals is crucial for robotic hands to achieve soft grasp control.This paper proposes a lightweight pyramid neural network(LZG-Net)model for embedded devices,designed to process vibration signals during robotic hand grasping.The LZG-Net model is based on the Ghost module and adopts a convolutional strategy with progressively decreasing convolutional kernels.It addresses the issue of the attention mechanism SE module being unable to deploy on some embedded devices and improves the model's accuracy on embedded systems through knowledge distillation,operator optimization,and quantization operations.Finally,an embedded tactile recognition system is built,and the LZG-Net model is deployed within it for tactile recognition of four objects with different characteristics.Experimental results show that the model can accurately classify object categories and grasping states,achieving a classification accuracy of 90.94%.Its classification performance is superior to existing classic lightweight neural networks.
NETL
NSTL
万方数据
