Qian Xiaoliang, Li Erkai, Zhang Jianwei, Zhao Su-Na, Wu Qing-E, Zhang Huanlong, Wang Wei, Wu Yuanyuan
School of Electrical and Information Engineering, Zhengzhou University of Light Industry, Zhengzhou, China.
Front Neurorobot. 2019 Sep 6;13:73. doi: 10.3389/fnbot.2019.00073. eCollection 2019.
The hardness recognition is of great significance to tactile sensing and robotic control. The hardness recognition methods based on deep learning have demonstrated a good performance, however, a huge amount of manually labeled samples which require lots of time and labor costs are necessary for the training of deep neural networks. In order to alleviate this problem, a semi-supervised generative adversarial network (GAN) which requires less manually labeled samples is proposed in this paper. First of all, a large number of unlabeled samples are made use of through the unsupervised training of GAN, which is used to provide a good initial state to the following model. Afterwards, the manually labeled samples corresponding to each hardness level are individually used to train the GAN, of which the architecture and initial parameter values are inherited from the unsupervised GAN, and augmented by the generator of trained GAN. Finally, the hardness recognition network (HRN), of which the main architecture and initial parameter values are inherited from the discriminator of unsupervised GAN, is pretrained by a large number of augmented labeled samples and fine-tuned by manually labeled samples. The hardness recognition result can be obtained online by importing the tactile data captured by the robotic forearm into the trained HRN. The experimental results demonstrate that the proposed method can significantly save the manual labeling work while providing an excellent recognition precision for hardness recognition.
硬度识别对于触觉感知和机器人控制具有重要意义。基于深度学习的硬度识别方法已表现出良好性能，然而，深度神经网络的训练需要大量人工标注样本，这需要大量时间和人力成本。为缓解此问题，本文提出一种需要较少人工标注样本的半监督生成对抗网络（GAN）。首先，通过GAN的无监督训练利用大量未标注样本，为后续模型提供良好的初始状态。之后，分别使用与每个硬度级别对应的人工标注样本训练GAN，其架构和初始参数值继承自无监督GAN，并由训练后的GAN的生成器进行增强。最后，硬度识别网络（HRN）的主要架构和初始参数值继承自无监督GAN的判别器，先用大量增强后的标注样本进行预训练，再用人工标注样本进行微调。将机器人前臂采集的触觉数据导入训练好的HRN即可在线获得硬度识别结果。实验结果表明，所提方法在显著节省人工标注工作的同时，为硬度识别提供了优异的识别精度。
