Abstract
Background: As human–robot collaboration (HRC) expands in industrial and service domains, accurate short-term prediction of human motion is essential for improving safety and task efficiency. Existing frameworks combine reactive safety measures with limited predictive capabilities, leading to conservative robot behaviors and reduced productivity. Methods: This study proposes a hybrid predictive human motion modeling framework that integrates probabilistic intent recognition, model predictive control (MPC)-based robot adaptation, and musculoskeletal-informed motion priors. The framework fuses multimodal sensing (depth, vision, and wearable inertial measurement) and performs online adaptation using recursive Bayesian updating. We validate the approach in simulated and pilot physical collaborative cell experiments involving reach-and-transfer tasks with variations in human intent and speed. Results: Compared to a baseline reactive safety controller, the predictive framework reduced time spent in reduced-speed safety mode by 42% (p < 0.01) and decreased near-miss events by 63% while maintaining normative safety clearances. Mean short-term (0.5 s) position prediction error dropped from 78 mm to 34 mm. A mixed-effects regression showed that fused sensor confidence and intent certainty were significant predictors of prediction accuracy (p < 0.001). Tables present descriptive statistics and regression coefficients. Conclusions: Incorporating short-horizon predictive models into HRC controllers substantially improves both safety and efficiency by enabling graded, confidence-informed robot responses rather than binary reactive measures. The approach is computationally tractable for industrial deployment and complements formal verification and safety assessment methods. Future work will extend horizon length via hierarchical prediction and integrate ergonomic cost models for worker comfort and fatigue mitigation.