Abstract: To enhance safety in dynamic human-robot collaboration and address collision risks stemming from human motion uncertainty and individual differences, a collaborative safety framework based on embodied intelligence theory was proposed. The framework first fused visual information and body mass index (BMI) to construct an adaptive human bounding volume model for personalized operator representation. Subsequently, a lightweight method combining Kalman filtering and Gaussian mixture regression was employed for probabilistic prediction of upper-limb motion trajectories. Finally, guided by the ISO/TS 15066 standard, a prospective collision warning strategy was developed by integrating prediction results to calculate the real-time minimum safety distance. Simulation experiments based on a public motion capture dataset showed that after parameter optimization, the system’s perception error was reduced by 17%; in simulated close-proximity collaboration tasks, the system operated stably at 30 Hz, monitoring an average human-robot distance of 159.1 mm and triggering 29 valid warnings within 60 frames. This study provides a verifiable theoretical and methodological foundation for shifting from passive protection to proactive and adaptive safety collaboration.