基于多模特征深度学习的机器人抓取判别方法

Multimodal Features Deep Learning for Robotic Potential Grasp Recognition

针对智能机器人抓取判别问题,研究多模特征深度学习与融合方法.该方法将测试特征分布偏离训练特征视为一类噪化,引入带稀疏约束的降噪自动编码(Denoising auto-encoding,DAE),实现网络权值学习;并以叠层融合策略,获取初始多模特征的深层抽象表达,两种手段相结合旨在提高深度网络的鲁棒性和抓取判别精确性.实验采用深度摄像机与6自由度工业机器人组建测试平台,对不同类别目标进行在线对比实验.结果表明,设计的多模特征深度学习依据人的抓取习惯,实现最优抓取判别,并且机器人成功实施抓取定位,研究方法对新目标具备良好的抓取判别能力.

In this paper, a multimodal features deep learning and a fusion approach are proposed to address the problem of robotic potential grasp recognition. In our thinking, the test features which diverge from training are presented as noise-processing, then the denoising auto-encoding （DAE） and sparse constraint conditions are introduced to realize the network＇s weights training. Furthermore, a stacked DAE with fusion method is adopted to deal with the multimodal vision dataset for its high-level abstract expression. These two strategies aim at improving the network^s robustness and the precision of grasp recognition. A six-degree-of-freedom robotic manipulator with a stereo camera configuration is used to demonstrate the robotic potential grasp recognition. Experimental results show that the robot can optimally localizate the target by simulating human grasps, and that the proposed method is robust to a variety of new target grasp recognition.