Limb Position Effects in Multimodal Sensing for Hand Gesture Recognition Under Static, Dynamic, and Unseen Arm Positions

Hand gesture recognition (HGR) is essential for applications like prosthetic control and wearable human-machine interfaces, where achieving reliable performance across varying arm positions remains a significant challenge. This study presents a systematic investigation of limb position effect in HGR using multi-modal sensor fusion with co-located surface electromyography (sEMG) and pressure-based force myography (pFMG) signals. We examine different training settings, including static-only, dynamic-only, and combined static-dynamic training. A dual-path temporal fusion network is used as a consistent modeling framework to combine sEMG and pFMG signals, allowing fair comparison across different conditions. Experimental results from ten participants performing nine functional hand gestures under multiple static and dynamic arm positions show that multi-modal sensor fusion and the inclusion of dynamic data during training clearly improve the HGR performance under limb position variation. When trained on combined static and dynamic data, the system achieves average accuracies of 96.50% and 92.02% when evaluated under static and dynamic arm positions, respectively. More importantly, this training approach improves generalization to arm-position variation, yielding average accuracies between 88.94% and 90.54% when tested on unseen arm positions. In contrast, models trained only on static data show a clear drop in performance when tested under dynamic and unseen conditions. These results highlight the importance of dual-path multi-modal sensor fusion with combined static and dynamic training data for improving robustness to limb position effect in HGR. The findings provide practical guidance on training data design and evaluation strategies to establish wearable gesture recognition systems that can operate reliably in real-world conditions.