A novel EMG-driven neuromuscular context encoding framework for real-time exoskeleton reference trajectory generation

Accurate intention-aware joint angle trajectory prediction from surface electromyography (sEMG) is central to responsive lower-limb exoskeleton reference trajectory generation, particularly under subject and protocol variability. This paper proposes NCE-Net, a novel neuromuscular context encoding architecture for real-time EMG-driven exoskeleton reference trajectory generation. NCE-Net incorporates channel-adaptive neuromuscular encoding to address nonuniform muscle relevance, temporal context aggregation to exploit short-horizon EMG activation history, and CDGE-guided learning to improve reference-trajectory fidelity. Experiments are conducted on two datasets, where NCE-Net achieves competitive accuracy against recurrent and state-of-the-art models, including a statistically significant improvement over prior work, while maintaining stable behavior relative to GRU-based baselines. On leave-one-subject-out evaluation, NCE-Net attains the lowest mean MAE among tested recurrent baselines, with the numerical differences interpreted carefully using confidence intervals and paired tests. In hardware-connected real-time validation on a private four-subject treadmill dataset, the validation achieved MAE = 2.658^∘, RMSE = 3.991^∘, and R²=0.937. The final NCE-Net prediction module required 0.726 ms mean CPU latency and 0.970 ms P99 latency for the EPIC-LAB input setting, remaining well below the 10 ms budget required for 100 Hz reference generation.