Hierarchical Recurrent-Inception Residual Transformer (HRIRT) for Multidimensional Hand Force Estimation Using Force Myography Sensor

In this letter, we present the hierarchical recurrent-inception residual transformer (HRIRT), an innovative deep neural network architecture designed for accurate hand force estimation in human-robot collaboration (HRC). The HRIRT combines recurrent layers, inception modules, residual connections, transformers, and Time2Vec feature engineering within a hierarchical framework to adeptly capture the complex spatiotemporal dynamics of hand force data. Our evaluation spans three dimensions of HRC-1-D, 2-D, and 3-D hand force estimation-leveraging data from force myography (FMG) sensors to train and test the model's performance. The HRIRT demonstrates exceptional accuracy and robustness across varied interaction scenarios with the Kuka Robot. The 1-D interactions focus on linear force applications, while 2-D and 3-D interactions involve more complex spatial movements, showcasing the model's capability to generalize across different force interaction contexts. In 1-D scenarios, HRIRT achieved a 93.76% R-Square (R2) score, significantly outperforming transfer learning with cross-domain generalization and stacked convolutional neural network (CNN) models. In addition, in 2-D and 3-D force estimations with R2 scores of 94.25% and 91.61%, respectively, the HRIRT showcased exceptional accuracy and maintained low error rates across root-mean-square error, normalized mean square error, and mean absolute error metrics. These results highlight HRIRT's potential as a powerful tool for real-time precise hand force estimation in diverse HRC applications.