A KT-LSTM-SAC-Based Closed-Loop Temperature Control Strategy for Medical Robot-Assisted Moxibustion

Abstract Robot-assisted moxibustion remains dependent on operator experience and is challenged by delayed thermal responses and insufficient closed-loop temperature stability. To address these issues, this study proposes a kinematic--thermal dual-stream LSTM-enhanced soft actor--critic method (KT-LSTM-SAC), where LSTM denotes long short-term memory and SAC denotes soft actor--critic. The temperature-control task is first formulated as adjustment of the end-effector standoff distance along the acupoint surface normal. A temperature-feedback model is then developed by integrating the standoff distance, equivalent heat input, surface temperature, and estimated deep-tissue temperature. On this basis, a joint-state feedback framework is constructed in which KT-LSTM-SAC outputs continuous standoff-distance adjustments for closed-loop control. A simulation environment for medical robot-assisted moxibustion is implemented in MuJoCo and evaluated through tests of target-band tolerance, target-temperature adaptability, controller comparisons, and ablation of the safety constraints. At a target condition of 43.0 $\pm$ 0.5 $^{\circ}$C, the proposed method achieved a target-band occupancy rate of 94.80\%, a mean absolute error of 0.32 $^{\circ}$C, and a mean safety cost of 3.60 $\times$ 10$^{-2}$. Compared with proportional--integral--derivative control and the baseline reinforcement-learning methods, KT-LSTM-SAC provided a better balance between temperature-tracking performance and safety.