| description abstract | Abstract. In onsite assembly tasks involving large-scale components, particularly in construction and aerospace, ensuring high precision and operational safety is crucial. Traditional manual operations face limitations such as excessive labor intensity and insufficient stability, while existing automated systems rely on predetermined processes, making them unsuitable for flexible, non-standard, and small-batch assembly tasks. Physical human–robot collaboration offers a promising solution by combining human flexibility with the high load capacity of robots. However, in expansive workspaces, operator movement introduces force fluctuations that affect operation stability and task performance. To address this, a stability intent understanding method is proposed based on multi-position sensor fusion and time-series prediction. Human motion characteristics during one-handed tasks are analyzed, integrating hand and foot sensor data. Fuzzy weight distribution and Dempster–Shafer evidence theory are employed to fuse multi-source intent information, enabling accurate identification of primary motion direction and suppression of irrelevant fluctuations. An enhanced long short-term memory network, incorporating an attention mechanism and penalty loss function, predicts operational force sequences to improve smoothness and stability. Finally, admittance control optimizes interaction compliance. Experimental results show the proposed method reduces path deviation by 6.7% and 16.8%, and shortens task time by 23.1% and 10.6%, respectively, compared with traditional methods. These improvements enhance collaboration efficiency and operational precision in large-scale component assembly. | |
