基于力-位图像学习的工业机器人柔顺装配方法研究

利用机器人进行自动装配操作时，控制装配过程的接触力和柔顺性对于保证装配质量具有重要意义。为此提出一种基于力-位图像学习的柔顺装配方法，将装配过程中的位姿和接触力信息转化为力-位图像，然后通过对力-位图像的分类学习，获得不同初始位姿情况下的柔顺装配动作策略，从而控制机器人实现柔顺装配。首先，控制机器人完成多次装配，并在装配过程中收集机器人位姿以及装配力和力矩信息；然后，利用这些信息绘制力-位曲线并将其组合成为力-位图像，基于运动方向判定算法为力-位图像自动标记装配动作标签，以构建力-位图像数据集；最后，在力-位图像数据集上对深度学习模型进行训练，并基于深度学习模型控制机器人进行柔顺装配。为了验证该方法的有效性，以RJ45连接头和连接口间的装配为例，采集了2 500次装配操作的力-位数据，共生成92 328张力-位图像及对应标签，然后基于ResNet50网络训练力-位图像分类模型，并基于该模型控制机器人进行装配实验，装配成功率达到96.7%。

Research on Flexible Assembly Method of Industrial Robot Based on Force-pose-image Learning

When using robots for automatic assembly, it is important to control the contact force and compliance in assembly process for ensuring assembly quality. Therefore, a flexible assembly method based on force-pose-image learning is proposed, which transformed the information of pose and force/torque into force-pose-image, and obtained the flexible assembly strategy under different initial pose conditions by classifying and learning the force-pose-image. The whole scheme is as following:Firstly, the assembly operations are completed many times, the data of pose, force and torque are collected during assembly. Then, the force-pose-curves are drawn and combined into force-pose-image, based on the motion direction determination algorithm, the force-pose-image are marked with motion labels to construct the force-pose-image data set. Finally, the deep learning model is trained on the force-pose-image data set, and the robot is controlled for flexible assembly based on the trained model.In order to verify the method, the assembly experiment of RJ45 connector and port is conducted. 2 500 times of assembly operations are performed, as the result, 92 328 force-pose-images and corresponding labels are generated. The force-pose-image classification model is trained based on ResNet50 network, and the robot is controlled to conduct assembly experiments based on the model. The assembly success rate is up to 96.7%.