This paper proposes a novel method to refine the 6D pose estimation inferred by an instance-level deep neural network which processes a single RGB image and that has been trained on synthetic images only. The proposed optimization algorithm usefully exploits the depth measurement of a standard RGB-D camera to estimate the dimensions of the considered object, even though the network is trained on a single CAD model of the same object with given dimensions. The improved accuracy in the pose estimation allows a robot to grasp apples of various types and significantly different dimensions successfully; this was not possible using the standard pose estimation algorithm, except for the fruits with dimensions very close to those of the CAD drawing used in the training process. Grasping fresh fruits without damaging each item also demands a suitable grasp force control. A parallel gripper equipped with special force/tactile sensors is thus adopted to achieve safe grasps with the minimum force necessary to lift the fruits without any slippage and any deformation at the same time, with no knowledge of their weight.

翻译：本文提出了一种新颖的方法，用于优化由仅基于合成图像训练的实例级深度神经网络从单张RGB图像推断出的6D姿态估计结果。所提出的优化算法有效利用了标准RGB-D相机的深度测量值，以估计目标物体的实际尺寸，尽管该网络仅基于同一物体具有给定尺寸的单一CAD模型进行训练。姿态估计精度的提升使得机器人能够成功抓取不同类型且尺寸差异显著的苹果；而标准姿态估计算法仅能成功抓取尺寸与训练所用CAD模型非常接近的水果。为在无损伤前提下抓取新鲜水果，还需合适的抓取力控制。因此，本文采用配备专用力/触觉传感器的平行夹爪，在无需获知水果重量的情况下，以托起水果所需的最小力实现安全抓取，同时避免滑动与变形。