A graph-based reinforcement learning-enabled approach for adaptive human-robot collaborative assembly operations

In today's prevailing manufacturing paradigm of mass personalization, neither human operators nor robots alone can perform all assembly tasks efficiently. To overcome it, human-robot collaborative assembly shows its great potentials to ensure the flexibility of human operations with high reliability of robot assistance. However, it is often challenging to achieve harmonious coexistence between humans and robots to complete the tasks safely and efficiently. In this regard, this research provides a detailed description of the human-robot coexisting environment and further introduces key issues in collaborative assembly. A part-behavior assembly and/or graph based on process requirements is proposed to represent the assembly task of complex products. Moreover, the human behavior prediction network based on self-attention can achieve higher accuracy. Combined with the robustness of Soft Actor-Critic (SAC), the collaborative system improves the self-decision ability of the robot in the dynamic scene. Finally, the effectiveness of the method is verified through experimental analysis. The results indicate that the accuracy of the proposed behavior recognition based on self-attention method is 91%. At the same time, it is proved that the reinforcement learning method is theoretically feasible to provide adaptive decision-making for robots in human-machine collaboration. In addition, the convergence speed of the reward function proves the feasibility of SAC for adaptive decision-making in a human-robot collaborative environment.