QMIX Algorithm for Coordinated Welding of Multiple Robots

This paper considers the coordinated welding control based on deep multi-agent reinforcement learning. The discrete-time states and actions with local observation for the welding robots and the characteristics of the weld lines (e.g., the starting points of the weld lines and the ending points of the weld lines) are defined, which is suitable to use the monotonic value function factorisation for deep multi-agent reinforcement learning (QMIX) algorithm. A novel reward composed of trajectory optimization, collision avoidance and the task done is designed, which is proved by the simulation in the grid-world environment.