Temporal Intent Reconstruction for Long-Horizon Agentic Predictive Control

Temporal Intent Reconstruction framework integrated with a Masked Cognitive Predictor to improve predictive control under changing goals and dynamic conditions.Using real multimodal data from HARMONIC, RoboMind, RoboNet, and Open X-Embodiment, the model reconstructs past intent trajectories and embeds misalignment signals into the control objective for long-horizon adaptation.Experiments showed stable reconstruction across embodiment and modality variations, reduced goal divergence by 31.4%, and improved tracking behaviour by 78% during transitions.The framework improved accuracy, RMSE reduction, and tracking behaviour compared with baseline MPC, inverse learning, and reinforcement-based controllers.These results indicate that temporal intent reconstruction enhances consistency and long-range predictive capability in systems operating under varied sensing, morphology, and task settings.