粘弹性农业物料机器人抓取模型辨识与主动柔性控制

Robotization is an important development direction of agricultural machineris, in which the grasping of viscoelastic agricultural materials is a fundamental key technologyin. Firstly, rheological properties of the agricultural materials will be investigated according to theory of viscoelastic mechanics, and the multi-dimensional features and bags of feartures will be mined out from abundant data obtained in real time from the grasping process by the tactile sensors, force sensors and so on. Then a weak physical model will be constructed through combining the rheological properties and the mined features with machine learning. Secondly, based on the knowledges of this physical model, hybrid control method will be applied to optimize the grasping force for the 3-dimensional viscoelastic agricultural materials,and relationship between grasping damage and grasping efficiency will be demonstrated. Thirdly, 3-dimentional point cloud data and 2-dimensional image will be fused, and the holographic discription of local space for garsping manipulation will be constructed. In this way the robot's ablities to analyze local manipulation space in multi-views and multi-sections way can be enhanced and measurement precision of the pose of the fruits and vegetables in the natural environment can be improved. Finally, experimental platform will be improved and experiments will be carried out in both the laboratory and the field. The object of this project is to take automatic harvesting of fruits and vegetables as example and get general principles in the grasping of viscoelastic agricultural materials.

机器人化是农业机械的一个重要发展方向，而粘弹性农业物料机器人抓取即是其中一项关键共性技术。本项目首先在粘弹性流变特性理论基础上，采用数据挖掘技术分析抓取过程中实时采集的接触力、位移、滑动等多传感器大数据中蕴含的多维特征信息或特征组合包，通过机器学习方法实现这些数据特征与理论模型参数之间关联，构建适宜主动柔性抓取控制的粘弹性农业物料弱物理模型。其次，根据此物理模型，探索多控制模式之间的稳定切换过程，研究粘弹性农业物料主动柔性控制方法，明确抓取损伤程度与快速抓取之间的定量联系。再次，融合三维点云空间局部法向量场数据和二维图像外观信息，建立抓取作业局部空间全息描述，增强机器人多视角、多切面分析能力，提高目标位姿参数测量精度。最后，完善实验平台，开展实验室和大田试验。总体研究目标是以果蔬自动采收为实例，探索粘弹性农业物料机器人抓取一般规律。

项目搭建了机器人两指末端执行器抓取系统，以番茄等果蔬为研究对象，选择伯格斯模型表征番茄的压缩变形特征，通过质构仪进行蠕变实验，获得不同成熟度番茄样本的粘弹性参数。构建了在线估计番茄粘弹性参数的神经网络模型，确定了网络的拓扑结构和参数，并利用质构仪试验实测数据对网络进行了训练，测试了网络模型的粘弹性参数估计性能。选择了匀减速、先匀速后减速以及变减速作为三种典型的抓取方式，并利用一次函数、巴特沃斯幅度平方函数和指数函数三种函数曲线描述机器人末端执行器实际抓取过程中不同抓取方式的速度变化，通过测量末端执行器抓取过程中力-时间、速度-时间以及位移-时间等数据，计算了三种抓取方式下产生的番茄塑性变形量。构建了机器人抓取过程中的触觉信息采集系统，利用主成分分析法进行触觉特征选择，基于KNN和SVM等算法进行了分类器设计，实现了末端机械手对不同果蔬硬度属性的识别。融合了自然环境中生长的果实的三维点云数据以及二维图像信息，运用GA-SVM等方法对果实进行了识别，实现了对果树枝条、树叶以及果实等不同对象的区分。项目的研究成果可以为开发抓取粘弹性物料的农业机器人提供理论基础，同时该成果也可以拓展到一般领域的服务机器人。