Abstract: Aiming at the problem that existing research has not yet fully utilized the three-dimensional spatial and visual features of various elements in the complex assembly environment,and has not considered the rich contextual information in the assembly environment,making the accuracy of intention recognition low,a method combining the three-dimensional spatial and visual information of the elements in the assembly environment and realizes the high-precision recognition of the human worker's operation intention based on the graph attention network was proposed.The Faster R-CNN was utilized to detect various elements in the assembly scene such as human workers,robots,obtain the spatial information of each element,and extract the visual feature information of each element from the network.Then,the graph attention network was utilized to reason the human worker's interaction intention toward different parts during assembly,such as handling,assembly and dragging.A gear assembly case study was used to verify that the proposed method.The experiment result showed that the proposed method could achieve higher performance in recognition accuracy and scene generalization compared with the deep convolutional neural network.

Key words: human-robot collaborative assembly, human intention recognition, context-aware, graph attention network

摘要： 针对人机协作装配现有研究尚未充分综合利用复杂装配环境中各要素的三维空间特征与视觉特征,并考虑装配环境中丰富的上下文信息,使得复杂装配环境下人员作业意图识别精度较低,提出结合装配环境各要素的三维空间特征和视觉特征,基于图注意力网络实现人员作业意图的高精度识别方法。首先利用FasterR-CNN神经网络对装配场景中的各要素,如人员、机器人、零件等进行目标检测,得到各要素的空间位置信息,同时从网络中提取各要素的视觉特征信息;然后结合图注意力网络推理装配过程中人员对不同作业对象的作业意图,如搬运、组装、触碰等;最后通过人机协作场景下的齿轮装配实验对所提方法进行验证。实验结果表明,相比深度卷积神经网络,所提方法在识别准确性、场景泛化性等方面具有优越性。

关键词: 人机协作装配, 作业意图识别, 上下文感知, 图注意力网络