Abstract
<jats:p>Low-cost motion capture may support anticipatory biomechanical surrogates, but causal deployability and target validity must be distinguished. We evaluated subject-specific 100-ms forecasting of physics-derived lower-limb torque proxies from dual-view marker video, with four-channel surface electromyography (sEMG) as an optional modality, using one 60-kg adult recording. Approximate 3D trajectories generated proxy targets through an engineering Lagrangian model; no force-plate or validated inverse-dynamics reference was available. Video-sEMG alignment was estimated on training data only and frozen at 559 frames (9.317 s). A six-channel causal gait-context receptive-field-matched temporal convolutional network (RF-TCN) achieved strict-test MAE 0.1326 Nm/kg, RMSE 0.2069, R² 0.4427, and r 0.6529 over 1564 windows; five-seed MAE was 0.1338 ± 0.0011 Nm/kg. Relative to the earlier causal-context reference, MAE decreased by 2.47%, while other metrics showed no supported improvement. A controlled K7-versus-K9 ablation showed only a small average MAE advantage for K9 and no statistically resolved overall effect. Sparse sEMG added no benefit beyond gait context. In stateful streaming replay, GPU complete-update p95/p99 latency was 21.78/25.03 ms with 0/6069 steady-state 50-ms deadline misses. These results support leakage-controlled, training-stable surrogate forecasting and GPU computational feasibility, not validated biological kinetics or hard-real-time exoskeleton control.</jats:p>