A methodological framework and experimental protocol for proactive human–robot collaboration with multimodal intention prediction and adaptive control

Industry 5.0 requires collaborative robots that can anticipate operator needs to improve fluency and safety in assembly. However, many human-robot collaboration (HRC) systems still treat perception, intention inference, and control as separate components. This study presents a theoretical perception-cognition-action framework that explicitly couples multimodal intention prediction with proactive and adaptive control. Multimodal observations such as RGB-D vision, gaze, wrist force/torque, robot joint state, and previous robot action are encoded by a hybrid Convolutional Neural Network (CNN)- Long Short-Term Memory (LSTM)-Transformer to estimate (i) a probability distribution over future human intentions and (ii) a short-horizon motion trajectory, trained with a composite loss that jointly optimizes classification and regression with kinematic coherence. The predicted intention probability is embedded into an augmented Markov Decision Process state, enabling a Soft Actor-Critic agent to learn continuous policies with rewards designed for synergy, efficiency, safety, and fluency. The main contributions of this study are the formal probabilistic linkage from intention prediction to adaptive control, the definition of a multi-output cognitive objective, and the design of an implementation-ready experimental protocol for future empirical validation. Overall, the proposed methodological framework and experimental protocol provide a reproducible basis for future empirical validation of proactive human-robot collaboration in industrial assembly tasks.