A Self-Adaptive Agent for Flexible Posture Planning in Robotic Milling System
Shengqiang Zhao,Fangyu Peng,Juntong Su,Hao Sun,Rong Yan,Xiaowei Tang,Teng Zhang,Zepeng Li
DOI: https://doi.org/10.1016/j.jmsy.2024.06.009
IF: 12.1
2024-01-01
Journal of Manufacturing Systems
Abstract:In recent years, industrial robots are widely used in the milling processes of the fundamental components such as aeronautic parts, aerospace cabins and blades on the marine propellers of the worships, because of their flexible structure, large operating space, and robust capacity of rapid reconfiguration. Along with the robotic machining process, multiple types of geometrical and physical constraints display the time-varying characteristics, which are difficult to be considered simultaneously by the traditional static planning methods on the robotic milling posture. Therefore, a deep reinforcement learning (DRL) enhanced virtual repulsive potential field flexible model is proposed for the robotic milling posture dynamic planning. The preliminary planning of static milling posture is realized based on the fixed setting of the virtual modal parameters in the posture planning model, which is treated as a theory preparation for training the deep reinforcement learning agent. To develop the static planning model, multiple constraints are established by considering the performances of geometric, kinematic and machining precision simultaneously. Furthermore, a reward function is constructed for the dynamic agent, comprehensively combining the long and short reward benefits in nonlinear correlation with virtual repulsive forces. By taking the virtual modal parameters as the action space of the dynamic agent, a DRL enhanced agent VRPF-SAC is created and trained under the dynamic milling environment from virtual repulsive planning field. Four tasks of large size robotic milling simulations and experiments are designed and developed to validate the effectiveness of the flexible planning model. In comparison with the static model, the flexible planning model displays the excellent global effects covering the kinematic, and precision performance, with 65.78 %, 72.04 %, and 70.25 % reduction in velocity, acceleration, and jerk of robotic joints, and ensuring the effective precision performance from -0.365 to 0.283 mm of large size workpiece with large curvature changes. The proposed flexible posture planning architecture can provide a basic for robotic dynamic milling of core parts with large size, narrow space and sharp curvatures of the aeronautic parts, aerospace cabins and worship propellers.