Deep Reinforcement Learning for Temperature Control of a Two-Way SMA-Actuated Tendon-Driven Gripper

Shape Memory Alloy (SMA) actuators offer strong potential for compact, lightweight, silent, and compliant robotic grippers; however, their practical deployment is limited by the challenge of controlling nonlinear and hysteretic thermal dynamics. This paper presents a complete Sim-to-Real control framework for precise temperature regulation of a tendon-driven SMA gripper using Deep Reinforcement Learning (DRL). A novel 12-action discrete control space is introduced, comprising 11 heating levels (0–100% PWM) and one active cooling action, enabling effective management of thermal inertia and environmental disturbances. The DRL agent is trained entirely in a calibrated thermo-mechanical simulation and deployed directly on physical hardware without real-world fine-tuning. Experimental results demonstrate accurate temperature tracking over a wide operating range (35–70 °C), achieving a mean steady-state error of approximately 0.26 °C below 50 °C and 0.41 °C at higher temperatures. Non-contact thermal imaging further confirms spatial temperature uniformity and the reliability of thermistor-based feedback. Finally, grasping experiments validate the practical effectiveness of the proposed controller, enabling reliable manipulation of delicate objects without crushing or slippage. These results demonstrate that the proposed DRL-based Sim-to-Real framework provides a robust and practical solution for high-precision SMA temperature control in soft robotic systems.