Development of a Four-Axis Force Sensor for Center of Gravity Estimation Using Tree-Based Machine Learning Models

State-of-the-art center-of-gravity (CoG) estimation methods often face accuracy limitations due to significant errors introduced by commercial force sensors. In addition, these methods typically require more than three poses and rely on complex matrices and equations, leading to prolonged computational times. This study introduces an advanced sensor system for precise CoG determination that requires only two poses, integrating a novel four-axis force sensor with a machine learning (ML) model. The sensor design is optimized based on the Timoshenko beam theory and validated through finite element analysis (FEA) and experimentation, achieving average measurement errors below 0.82% and interference errors under 0.71%, demonstrating its accuracy and reliability. The ML dataset comprises approximately 391000 data points, utilizing 12 input variables to predict CoG represented by four variables. Various tree-based ML models - including decision tree (DL), random forest (RF), extra trees (ETs), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM) - were evaluated, with hyperparameter tuning performed using Optuna and Bayesian optimization. The results demonstrated that all the developed ML models effectively captured the underlying patterns in the data, as evidenced by evaluation metrics. Among these, the ET model outperformed the others in terms of accuracy and reliability, as indicated by its lowest mean absolute error (MAE) of 0.000 N for weight estimation and 0.009 mm for position estimation. These findings underscore the potential of the proposed system and model combination as a highly efficient and scalable solution for precise CoG estimation, with significant implications for industrial and robotic applications.