Compliant assembly of complex aviation plugs combining multi-agent search and variable parameters

Purpose In response to the complex external structure of high-precision aviation plugs, which makes it difficult to search outside the hole and adjust inside the hole during automated assembly. This paper aims to propose an assembly framework that combines multi-agent search and variable parameter compliant control to solve this problem. Design/methodology/approach First, a multi-agent search strategy (MAS) based on Gaussian Mixture Model and Deep Q-Network was proposed to optimize displacement direction and actions, thereby improving search speed and success rate. Then, a variable parameter admittance control method (RL-VPA) based on dual delay depth deterministic policy gradient (TD3) was proposed, which dynamically optimized the internal parameters of the admittance controller and adopted state space discretization to improve convergence speed and assembly efficiency. Findings Compared to spiral search and single-agent search, the average search success rate has improved by approximately 10% and 6.6%. Compared to fixed admittance control and other RL-based methods, the average assembly success rate has increased by approximately 38.6%, 22% and 8.6%. Compared with the training results of the model without state discretization, it was found that state discretization helps the model converge quickly. To verify the generalization ability of the assembly framework, experiments were conducted on three different pin counts of aviation plugs, the assembly success rate reached 86.7%, all of which showed good assembly results. Finally, combining state space discretization to reduce the impact of environmental noise, improve training effectiveness and convergence speed. Originality/value MAS has been proposed to optimize displacement direction and action, improving search speed and success rate. RL-VPA is designed to dynamically optimize the internal parameters of the admittance controller, enhancing the robustness and generalization ability of the model. Additionally, state space discretization is combined to improve training effectiveness and convergence speed.