Hybrid Physics-Infused Deep Learning for Enhanced Real-Time Prediction of Human Upper Limb Movements in Collaborative Robotics

Abstract Human–robot collaboration is crucial in various industries, making accurate prediction of human arm movements essential for seamless interaction. This paper presents a significant advancement in collaborative robotics by developing a hybrid model that enhances the accuracy and interpretability of human motion predictions. By integrating a Physics-Infused Model with Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks, our approach effectively captures intricate temporal dependencies while incorporating physical constraints, leading to more robust and realistic predictions. The hybrid model was successfully implemented on an ABB IRB 120 robot, demonstrating its practical applicability in real-world scenarios. Our results show that this model outperforms conventional methods, particularly in predicting human arm positions during collaborative tasks. The key contribution of this work lies in the integration of deep learning with physics-based principles, setting a new benchmark for predictive accuracy in human–robot collaboration. This research not only enhances the performance of collaborative robots but also opens the door for similar hybrid models to be applied in other fields where accurate motion prediction is critical.