Robotic autonomous grasping of food-related objects requires a nondestructive and safe grasp system for picking up various objects. A novel underactuated flexible hand consisting of a variable palm and four soft fingers is designed and manufactured to enhance the grasp space and deformability during interaction with unknown objects. A position deviation formulation is fitted to estimate the free bend deformation of soft fingers approximately through finite element analysis. A modular convolutional neural network is presented to identify the grasp directions, shape features and anticipated input pressure levels of novel objects for achieving multitarget classification. A vision-based adaptive detection method is proposed to obtain an accurate wrist orientation and the best grasp candidate by using two means of grasp planning (i.e. cross grasp planning and equidistant optimal grasp planning). A two-stage grasp strategy combining the classification and detection methods is developed as an effective solution to estimate the grasp configuration accurately. Results show that our flexible hand achieves 91.1% success rate in a physical grasp experiment on a UR robot, thereby demonstrating the reliability and adaptability of our grasp approach. The target object can be identified and detected within 0.263 s, which indicates the suitability of our approach in real-time applications.

机器人自动抓握与食物相关的物体需要一种无损且安全的抓握系统来拾取各种物体。设计和制造了由可变手掌和四个软手指组成的新型欠致动柔性手，以增强与未知物体相互作用时的抓握空间和变形能力。拟合位置偏差公式以近似地通过有限元分析来估计软指的自由弯曲变形。提出了一种模块化卷积神经网络，以识别新颖对象的抓取方向，形状特征和预期输入压力水平，以实现多目标分类。提出了一种基于视觉的自适应检测方法，通过两种抓握计划（即：交叉把握计划和等距最优把握计划）。提出了一种结合分类和检测方法的两阶段抓握策略，作为准确估计抓握形态的有效解决方案。结果表明，在UR机器人的物理抓握实验中，我们的柔性手获得了91.1％的成功率，从而证明了我们抓握方法的可靠性和适应性。可以在0.263 s内识别并检测到目标对象，这表明我们的方法在实时应用中的适用性。在UR机器人上进行物理抓取实验时，成功率达到1％，从而证明了我们抓取方法的可靠性和适应性。可以在0.263 s内识别并检测到目标对象，这表明我们的方法在实时应用中的适用性。在UR机器人上进行的物理抓握实验，成功率达到1％，从而证明了我们抓握方法的可靠性和适应性。可以在0.263 s内识别并检测到目标对象，这表明我们的方法在实时应用中的适用性。