YOLO-Based Human Pose Awareness for Occlusion-Challenged Human–Robot Collaborative Assembly

Human–robot collaborative assembly (HRCA) has become an important paradigm in modern intelligent manufacturing, where robots must accurately perceive human posture and motion in order to ensure both operational safety and efficient cooperation. However, reliable human pose perception in real assembly environments remains challenging due to frequent occlusions, cluttered backgrounds, and complex human movements. To address these challenges, this paper proposes a human pose awareness framework for HRCA based on an enhanced YOLO11-Pose architecture, termed YOLO11-CKPose. The proposed framework integrates skeleton-based pose estimation with improved feature extraction and occlusion-aware detection mechanisms to enhance keypoint localization accuracy under partial visibility conditions. In addition, a dedicated human–robot collaborative assembly dataset is constructed using RGB-D sequences collected from 15 participants performing representative assembly tasks under multiple interference scenarios. Transfer learning is employed to adapt the model from the COCO dataset to the assembly-specific environment. Experimental results in both simulated and real assembly scenarios demonstrate that the proposed method achieves over 95% recognition accuracy, outperforming baseline YOLO-based pose models in terms of robustness and inference efficiency. The proposed framework provides reliable human pose perception for behavior recognition and safety-aware decision making, thereby supporting more adaptive and intelligent human–robot collaboration in complex assembly environments.