| Title |
Analysis of AI Model Performance for EMG-Based Human-Robot Handover State Recognition |
| Authors |
Kim, Taeeun ; Yang, Kanghyeok |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2025.26.1.067 |
| Keywords |
Human-Robot Collaboration; Handover; Electromyography Signal; Robotic Arm; Deep Learning |
| Abstract |
The study developed an approach to recognize handover tasks required for collaborative construction work with robots using a worker’s electromyography (EMG) signals. The study investigated the recognition performance based on different artificial intelligence algorithms. The handover task was divided into three stages (Pre-grasp, Half grasp, and Full grasp) depending on the degree of object grasp. The EMG signals of each grasp state were collected in a laboratory environment. The collected data were visualized in time and frequency domains, and recognition performance was evaluated using convolutional neural networks (CNN) and long short-term memory (LSTM) networks for each data domain to derive the optimal AI model. The analysis results showed that the CNNbased model exhibited superior performance with an accuracy of 0.99 for time domain data, while the LSTM-based model achieved better performance with an accuracy of 0.98 for frequency domain data. Furthermore, the leave-onesubject-out cross-validation approach demonstrated that the LSTM model achieved a notably higher performance with an accuracy of 0.69 compared to the CNN model. The results of the study serve as foundational research for developing technologies for human-robot collaboration in construciton are expected to contribute to improvement of the safety and the productivity through collaborative construction robots. |