<b>Predictive Modeling of a Compliant Mechanism based Gripper using Machine Learning methods</b>

<p dir="ltr">Traditional force sensing in tele-operated robotic systems faces significant limitations—including high costs, fragility and integration challenges—that can severely restrict their use in precision manipulation tasks. These constraints present major obstacles to achieving safe and effective force control in demanding applications such as Minimally Invasive Surgery (MIS) and space exploration, where direct force measurement is often impractical. To address these challenges, this work considers a CM based gripper system as a testing apparatus for providing real-time, force feedback via structural deformation as a visual cue, thereby communicating force information without embedded sensors. Thus, this research characterizes the CM-based force feedback capabilities and subsequently develops a Machine Learning (ML) model to predict gripping force information, an important factor to advance the CM-gripper towards a future computer-assisted force feedback system.</p><p dir="ltr">The study adopted a two-phase approach to bridge mechanical characterization with intelligent force prediction of a compliant mechanism (CM) gripper. In the first phase (Paper 1), the force-deflection properties of a CM gripper are experimentally characterized, revealing non-linear relationships between structural deformation and pinch forces. The nonlinearity is attributed to plastic deformation that was observed in the experiments due to excessive actuation. This finding redefined the experimental conditions to remain within the elastic region. In a separate work, the experiments were repeated under these requirements and showed linear agreement, which enables a linear based visual force feedback system.</p><p dir="ltr">These findings directly inform the second phase (Paper 2), where the established relationships guide the selection and training of machine learning models. The experimental data from Paper 1 served as the foundation for understanding the design parameters for ML model development. In the work presented in Paper 2, ML models were systematically developed and verified against experimental data for pinch force prediction. The ML based research also employed a multi-metric evaluation framework - combining performance, absolute error and execution time visualized via radar charts to address critical gaps left by prior studies that rely on single performance indicators. Out of the seven ML models tested, the Weighted Regression demonstrated the best overall performance (Radar Chart Area = 0.95: R² = 96.27%, Mean Absolute Error = 0.51N, Execution Time = 1.41 s).</p><p dir="ltr">This research developed a deformation-based ML-driven model to translate visual cues from a compliant mechanism gripper to pinch force prediction. This work forms a critical component of a larger research initiative aiming to integrate the ML model with image processing software to create a CM-gripper system equipped with real time computer-assisted force feedback capabilities. Such advances pave the way for future evaluations of two distinct force feedback systems using the CM based gripper system as a testing apparatus, specifically (a) human-assisted gripper manipulation, where force predictions are graphically overlaid to guide users and (b) fully automated ML-controlled gripper system that autonomously adjusts and maintains the required pinch force without user intervention. Collectively, this research moves this field closer to an intelligent, sensor-less robotic manipulation system.</p>