基于语义分割的不规则目标抓取策略研究

Research on Irregular Object Grasping Strategies Based on Deep Learning

人工智能和机器人是人类科技在软硬件领域的巅峰,因此本文开展了将深度学习模型用于机器人抓取策略的研究,不仅能为机器人赋予智能属性,还能让以深度学习为代表的人工智能模型落地应用。本文主要研究了将语义分割模型改进后用于不规则目标的三指抓取策略研究。首先,本文分析了DeepLab模型架构的合理性以及其包含的空洞卷积的强大性能,并搭建用于模型训练和验证的数据集。其次,本文以DeepLabV3+算法为基础,将其改进后用于不规则目标的抓取策略生成研究。最后通过实验研究发现本文提出的适用于三指灵巧手的三角形抓取策略具有通用性强和成本合理的优点,为人工智能算法的落地应用和机器人抓取研究作出了有益探索。

Artificial intelligence(AI)and robots are the pinnacle of human technology in the field of software and hardware.Therefore,this paper conducts research on applying deep learning models to robot grasping strategies,which not only endows robots with intelligent attributes but also enables AI models represented by deep learning to be applied.This paper mainly studies the three fingers grasping strategy for irregular objects after improving the semantic segmentation model.Firstly,this paper analyzes the rationality of the DeepLab model architecture and the powerful performance of the dilated convolutions it contains,and builds a dataset for model training and validation.Secondly,this paper will be based on the DeepLabV3+algorithm and improve it to study the generation of grasping strategies for irregular objects.Finally,the research results has been found that the triangle grasping strategy proposed in this paper for three fingers dexterous hands has the advantages of strong universality and reasonable cost,which makes a beneficial exploration for the practical application of AI and robot grasping research.