RT-HMP: A Modular and Extensible Framework for Real-Time Human Motion Prediction

Abstract

This letter presents the Real-Time Human Motion Prediction (RT-HMP) framework, a modular and open-source architecture designed to enable seamless integration of Human Motion Prediction (HMP) capabilities into a variety of application pipelines, supporting more efficient, natural, and safe human-robot interactions. The proposed architecture decouples pose estimation from motion prediction and standardizes heterogeneous sensor inputs through a unified representation, facilitating interoperability and extensibility. A model-agnostic interface is provided via standardized APIs, enabling straightforward integration of different HMP models. A comprehensive evaluation demonstrates this modularity by testing four distinct deep learning architectures on experimental human motion data under domain shift. Furthermore, physical closed-loop collaborative assembly experiments show that the framework achieves real-time compliance with end-to-end latencies under 30 ms, improves event-triggering reliability (F1-score of 0.986), and enables proactive robot behaviors that enhance workflow continuity compared to systems based solely on pose estimation data. The design of RT-HMP promotes the widespread adoption of predictive human motion modeling across domains such as autonomous robotics, collaborative workspaces, and social robot interaction.