Time-Advancing Multimodal Motion-State Estimation for Soft Lower-Limb Exoskeletons Using sEMG-IMU Fusion

Latency in the sensing-estimation pipeline can make exoskeleton assistance arrive late. We study an offline time-advancing estimator that fuses sEMG and IMU signals to predict future gait phase, bilateral hip angles, and walking speed at t+δ. The model uses a lightweight dual-stream architecture with a CNN sEMG encoder, a GRU IMU encoder, and a channel-wise gating module. Evaluation on a synchronized sEMG-IMU-MoCap dataset from eight participants under six treadmill conditions (48 trials) showed that, at δ=100 ms, the fusion model achieved NRMSE 0.082 ± 0.017 for phase, 0.060 ± 0.013 for hip angle, and 0.183 ± 0.027 for speed, with correlations of 0.970 ± 0.014,0.984 ± 0.010, and 0.867 ± 0.037. Fusion also degraded more gracefully than unimodal baselines as the horizon increased to 250 ms, supporting its use for offline future-state estimation.