为了提高机器人的操纵能力，抓取方式的研究一直是机器人领域的一个重要方向。目前，已经有了很多关于刚性机器人手的抓取识别的研究，例如基于手部重建或触觉传感的方法都取得了很大的进展。由于技术进步和市场竞争，产品的生命周期正在发生变化，对机械手的需求迫切增加。同时，成本效益和简单化是工业应用的首要因素。软体机械手表现出控制简单、适应性强的特点，因此受到了广泛的关注。而软体机械手的高效掌握仍需要新的理论和方法的支持。本工作主要是针对自主学习机器人在识别抓取类型和确定抓取方向方面引入一种新的利用了红外图像的方法，特别是在软抓取方面的应用。取得的创新成果如下：我们首次提出了一种新的红外成像程序用于抓取类型和方向的识别。该方法利用手指在红外图像中留下的热印记来识别抓取类型和方向。因此，我们可以有效地避免传统图像识别方法的相互遮挡问题，实现了优异的预测计算量，解决了方法复杂度与识别效率之间的矛盾。使用Tensorflow、Keras和CUDA等深度学习框架，在GPU上对用于抓取类型识别的卷积神经网络（Convolutional Neural Networks, CNN）进行训练。结果在抓取类型检测中可以达到97%的准确率和8%的损失。最后，利用Olafenwa的YOLOv3算法框架，我们训练了一个算法来检测红外图像中的指印，平均精度达到97%，它能够提取感兴趣的区域，并用旋转最小面积矩形的方法检测角度。
In order to improve the manipulation ability of robots, the research on the grasping approach has always been an important direction in the field of robotics. At present, much research on the grasp recognition of rigid robotic hands has been realized, like methods based on hand reconstruction or tactile sensing, and have made great progress. Due to technological progress and market competition, product life cycles are changing, and the demand for manipulators is increasing urgently. At the same time, cost-effectiveness and simplicity are the primary factors for industrial applications. The soft robotic hands show control simplicity and efficient adaptation, which has received widespread attention. The efficient grasp of the soft robotic hand needs to be supported by new theories and methods. This work focuses on introducing a new approach in grasp type recognition and grasp orientation determination for autonomous learning robots via infrared images, particularly with the application in soft grasping. The achieved innovative results are as follows.For the first time, we propose a new IR-imaging procedure for identification of grasp type and orientation. This method uses the thermal imprint left by the fingers in the infrared image to identify the grasp type and orientation. Consequently, we can effectively avoid the mutual occlusion problem of the traditional image recognition method, achieve an excellent prediction calculation amount, and solve the contradiction between method complexity and recognition efficiency.The (Convolutional Neural Networks, CNN) for grasp type recognition is trained using Tensorflow, Keras and CUDA on a GPU. Results in grasp type detection up to 97% accuracy and 8% loss could be achieved. Finally, using YOLOv3 from Olafenwa, we could train an algorithm to detect a fingerprint in an infrared image with a mean average precision of 97%, extract the region of interest and detect the angle with the rotated minimum area rectangle approach.
