毕业生知识库

      对于机器人来说，如何进行抓取是一项重要的能力。机器人抓取任务经过数十年的研究，已经取得了很多成果，并在工业、探索、服务、军事等领域中得到了广泛的应用。然而，机器人进行抓取时远不像人一样随心所欲、灵活自如，仍然具有很大的局限性。受到目标物体各种特性如形态、材质、重量等因素以及复杂多变的环境因素的影响，机器人抓取研究仍面临着严峻的挑战性。
      本文以人机协作型机器人作为智能抓取研究的对象，控制其在特定的工作环境下进行智能分类抓取。在智能抓取研究中，本文以视觉信息为引导，通过研究基于深度学习的机器人抓取的图像特征表达方式，判断待抓取目标的类别与位置，并在待抓取目标上检测出适合机器人抓取的位置。

      机器人智能抓取系统涉及到的问题主要包括抓取目标的分类与定位、抓取位置检测和抓取位置的位姿估计：

      1. 在抓取目标的分类与定位方面，采用基于深度学习的目标检测方法。通过建立有关抓取目标的目标检测数据集，训练用于目标检测的神经网络，并对训练所得网络模型的性能进行分析。

      2. 在目标物体上的抓取位置方面，提出基于深度学习的抓取位置检测方法。同样地，通过建立有关抓取目标的抓取位置检测数据集，训练用于抓取位置检测的神经网络，将所提方法与其他方法作对比，分析所提方法的有效性。

      3. 在抓取位置的位姿估计方面，采用坐标转换的方法，将抓取位置在图像上的二维坐标转换到在机器人下的三维坐标，并对抓取位置的姿态进行估计。
      最后，本文将基于深度学习的目标检测方法与抓取位置检测方法以及抓取位置的位姿估计相结合，构建一种基于视觉的机器人智能抓取系统，并将该系统应用到实际的机器人上进行抓取实验。实验结果表明，所提抓取系统能够对待抓取目标进行快速的分类、定位与抓取。本系统可以应用于与抓取相关的机器人任务中，并为后续的机器人抓取研究提供参考依据。

      For robots, grasping objects is an important ability. After several decades, robotic grasping has made great achievements and been widely used in industry, exploration, service, military and other fields. However, because of limitations, robots are not as flexible as humans to grasp objects. Affected by various factors of the objects, such as forms, materials and weights, and the complicated and changeable environment, the robotic grasping tasks still faces severe challenges.

      In this paper, the man-machine collaborative robots are used as the targets on intelligent grasping studies and controlled to grasp classified objects in the specific working environment. The visual information is regarded as a guidance for the robots. And through the deep learning based studies of the image feature representation, the robots can estimate the categories and positions of the objects, and can detect the suitable grasping positions on these objects.

      The issues involved in the proposed robotic intelligent grasping study include the object classification and location, the grasping position detection and pose estimation:

      1. In the aspect of the object classification and location, the object detection method based on deep learning is adopted. We create the data set of it and use the data set to train the deep neural network. After getting the trained model, we test its performance and analyze the test results.

      2. In the aspect of the grasping position detection, the grasping position detection method based on deep learning is proposed. Similarly, we also create the data set of the grasping position detection, and use it to train another neural network and get a trained model. We compare the test results with other results generated by other methods, and analyze their effectiveness.

      3. In the aspect of the pose estimation of the grasping position, we use the coordinate transformation method to convert the two-dimensional coordinates in the images to three-dimensional coordinates under the robots, and estimate their axes’ orientations.

      Finally, we combine the deep learning based object detection, the deep learning based grasping position detection and the pose estimation of the grasping position to a visual information based intelligent grasping system of the robot in this paper. And we apply it to a real robot to do some experiments. The experimental results show that, the proposed grasping system can classify, locate and grasp objects rapidly. The proposed system can be applied to the grasping-related robotic tasks, and can be regarded as a reference for subsequent robotic researches.