A Robust Myoelectric Gesture Recognition Method for Enhancing the Reliability of Human-Robot Interaction

The myoelectric gesture recognition technology based on wearable armbands provides a natural and portable solution for human-robot interaction (HRI). However, various interferences during practical interactions can severely degrade the recognition model's performance, leading to reduced interaction reliability. Therefore, this study proposes a method called Distribution Shift Online Detection and Unsupervised Domain Adaptation (DSOD-UDA), aimed at addressing two key issues in the interactive process: when the model's performance declines and how to handle it after the decline. The method utilizes a discriminator with a sliding window to monitor real-time changes in the feature space of myoelectric signals, determining whether a distribution shift has occurred. Once a distribution shift is detected, the recognition model is updated online to ensure adaptability to the current distribution. Offline validation experiments were conducted on a public dataset that includes various interference factors. Ten participants conducted online experiments, simulating practical interference factors by performing the designated task during interactions and then using recognized gestures to control a robot to complete the object transfer task. The results demonstrate that, compared to comparison methods, the proposed method significantly enhances gesture recognition performance and exhibits superior robustness to various interference factors.