Human-Robot Collaboration through a Multi-Scale Graph Convolution Neural Network with Temporal Attention

Collaborative robots sensing and understanding the movements and intentions of their human partners are crucial for realizing human-robot collaboration. Human skele- ton sequences are widely recognized as a kind of data with great application potential in human action recognition. In this paper, a multi-scale skeleton-based human action recognition network is proposed, which leverages a spatio-temporal atten- tion mechanism. The network achieves high-accuracy human action prediction by aggregating multi-level key point features of the skeleton and applying the spatio-temporal attention mechanism to extract key temporal information features. In addition, a human action skeleton dataset containing eight different categories is collected for a human-robot collaboration task, where the human activity recognition network predicts skeleton sequences from a camera and the collaborating robot makes collaborative actions based on the predicted actions. In this study, the performance of the proposed method is compared with state-of-the-art human action recognition methods and ablation experiments are performed. The results show that the multi-scale spatio-temporal graph convolutional neural network has an action recognition accuracy of 94.16%. The effectiveness of the method is also verified by performing human-robot col- laboration experiments on a real robot platform in a laboratory environment.