Biomedical Signal Processing and Control · Nov 4, 2026
Agile human-robot collaboration through movement anticipation: multi-task sEMG learning for knee tracking in older adults with possible sarcopenia
By Xue J, Li Z, Cheng ACY et al.
Effective human-robot collaboration (HRC) in sit-to-stand transitions (STS) requires not only accurate motion tracking but also rapid intent synchronization, especially for older adults who may have sarcopenia. Traditional control strategies often suffer from sensory delays and cannot adapt to the inherent neuromuscular delays in frail individuals. This paper proposes a novel multi-task surface electromyography (sEMG) driven framework designed to simultaneously detect pre-motion and predict continuous knee angles. By employing a dual-stream CNN-LSTM architecture, the model extracts timing and spacial features from sEMG signals to provide proactive intent gating and smooth trajectory estimation. Experimental results from 15 elderly subjects show that the model achieves a phase classification accuracy of 86.65 %±4.29 % and an R 2 value of 0.9312 ± 0.0366 for knee angle prediction. Notably, the statistical analysis revealed significant differences in muscle activation patterns among subjects with different levels of sarcopenia risk, thereby enhancing the interpretability of the proposed model. Furthermore, motion control simulations demonstrate that using pre-motion detection as an active trigger successfully translates passive latency into physical lead time (approximately 80 ms). This transformation effectively offsets hardware latency and reduces the risk of false triggering. This research provides a collaborative approach for assistive robots, bridging the gap between deep learning-based intent decoding and the clinical needs of sarcopenia rehabilitation, ultimately promoting safer and more responsive human–computer interaction.