<p>Human-robot interaction (HRI) is inherently challenging due to the dynamic and unpredictable nature of human motion, which complicates the determination of appropriate timing and location for interaction. To address this challenge, we propose a novel framework that integrates skeleton-based action recognition with motion prediction, enabling safe and adaptive trajectory generation. Action recognition is used to determine when to interact, while motion prediction is used to determine where to interact. Since human actions inherently influence subsequent motion patterns, the combination of these two components yields a more comprehensive understanding of human intent, thereby improving both prediction accuracy and interaction timing. A dual-level deep learning architecture is introduced to implement the proposed method. The high-level module performs human action recognition for interaction timing by capturing spatial features, whereas the low-level module focuses on human motion prediction by modeling temporal dynamics. The fusion of these features reduces uncertainty in future motion prediction. Based on the predicted human motion, the robot generates collision-free trajectories that dynamically adapt to human actions. The proposed method was evaluated against state-of-the-art methods and further validated through human-to-robot handover experiments. Experimental results demonstrate that the method achieves superior prediction accuracy, enables real-time trajectory adaptation, and enhances the adaptability of robots in interactive scenarios.</p>