To meet the requirements of compliant interaction control and stable operation for robotic manipulators in decontamination tasks, a compliant control method based on Long Short-Term Memory (LSTM) Network state identification is proposed. A hybrid force/position impedance control model is established based on the Newton–Euler dynamic method combined with Coulomb and viscous friction models, and a selection matrix is adopted to decouple the force-control and position-control subspaces. An LSTM-based temporal state identification model is constructed using position, contact force, and joint torque observations to achieve real-time task-state recognition and smooth control-mode switching, where the LSTM network is mainly used for state identification and mode-switching decision-making. Simulation and experimental results demonstrate that the proposed method can effectively suppress transient vibration during mode switching. Compared with the traditional hybrid force/position control strategy, the peak transient force error is reduced by 61.53%, while the dynamic response performance and contact stability are improved.