Thermomechanical characterization and hybrid LSTM–XGBoost force sensing of a two-way SMA spring actuator for robotic grasping
A two-way SMA spring gripper uses infrared thermal imaging and hybrid LSTM-XGBoost models to estimate gripping force without physical sensors, achieving 73.45 mN RMSE and 82.5% grasp success.
Published Oct 2, 2026Paper ↗
Only vote on papers you've read. Sign in to vote.
Abstract
Shape memory alloy (SMA) actuators exhibit pronounced thermomechanical hysteresis and history-dependent behavior, making gripping-force estimation challenging without dedicated force sensors. This study presents a compact robotic gripper actuated by a single two-way shape memory effect (TWSME) NiTi spring and a non-contact virtual force-sensing framework based on infrared thermal imaging and RGB-based object-width measurement. The actuator was experimentally characterized through gripping-force hysteresis, free-stroke thermomechanical behavior, transient thermal response, and repeated heating–cooling tests. A closed-loop thermal-regulation system combining PWM-controlled Joule heating and pressurized-air cooling was implemented. Compared with passive cooling, the 100-kPa pressurized-air condition reduced the measured cooling time to 40 °C by approximately 71–76% over initial temperatures of 65–80 °C. For virtual force sensing, synchronized sequences of SMA temperature and object width were processed using a hybrid LSTM–XGBoost model. A pretrained LSTM encoder extracted history-dependent thermomechanical features from an approximately 2-s observation window, which were subsequently mapped to gripping force using XGBoost regression. Under a leakage-free multi-seed leave-one-condition-out evaluation, the proposed hybrid model achieved the lowest mean RMSE of 73.45 ± 38.88 mN, with an MAE of 57.38 mN and R 2 = 0.9267. Paired fold-level comparisons showed significantly lower RMSE than the DNN and standalone XGBoost baselines, while differences from the standalone LSTM and GRU models were not statistically significant. Robotic grasping experiments with a UR3 manipulator achieved 66 successful grasps out of 80 trials (82.5%; Wilson 95% CI: 72.7–89.3%) across eight objects without a force sensor mounted on the gripper. These results demonstrate the feasibility of non-contact virtual force estimation for the investigated TWSME spring–gripper under low-speed laboratory operating conditions.