We present a comparative analysis of two distinct machine-learning models designed to detect asynchronous errors during Human-Robot Interaction (HRI). The models under scrutiny are a customized ResNet model and an ensemble model, both trained and validated using EEG data. The ResNet model is a unique adaptation of the Residual Network, comprising a one-dimensional convolutional layer followed by batch normalization and ReLU activation. It also features a custom Residual Block structure and an adaptive average pooling layer, concluding with a fully connected linear layer for binary classification. The ensemble model, on the other hand, incorporates various machine learning models such as MLP, logistic regression, SVM, random forest, and XGBoost, unified in a pipeline with feature extraction and transformation steps. A critical component of our research is the innovative probability map, which maintains a granularity of 0.1 seconds. This map forecasts the likelihood of forthcoming one-second intervals being classified as either Error or Non-error. Our comparative analysis reveals significant variations in the performance of the two models, both of which exhibit promising results in detecting erroneous behaviors during HRI. We provide detailed validation results, including the accuracy, F1 score, and confusion matrix for each model. This study offers valuable insights into the potential of machine learning in enhancing HRI efficiency and accuracy, indicating promising directions for future research.
