EUPA: Embodied Uncertainty-Aware Multimodal Affect Perception for Consumer Electronic Systems

Multimodal sentiment analysis has achieved strong performance in controlled settings by integrating visual, acoustic, and textual signals. However, real-world consumer electronic systems operate under dynamic conditions with noisy, asynchronous, and partially observable inputs, challenging conventional models that assume reliable multimodal data. In this paper, we reformulate multimodal sentiment analysis as an embodied affect perception problem and propose EUPA, an uncertainty-aware framework that models modality reliability and performs confidence-guided temporal updates for online inference. EUPA dynamically regulates unreliable inputs and maintains a belief-based representation to support robust perception under evolving conditions. To evaluate this setting, we introduce an embodied evaluation protocol on CMU-MOSI and CMU-MOSEI with simulated modality dropout and temporal misalignment. Experimental results demonstrate improved robustness and temporal stability over existing methods, highlighting the potential of EUPA for real-world affect-aware applications.