Human-Robot Interaction (HRI) is currently undergoing a paradigm shift from discrete command-based control to seamless, symbiotic communication. However, achieving this symbiosis requires overcoming significant challenges in temporal alignment and computational efficiency, particularly when processing conflicting multimodal signals. This paper presents a hybrid study combining a semi-systematic literature review (2020–2025) with a novel conceptual framework. The study critically evaluates the architectural evolution from standard Transformers to linear-complexity State Space Models (SSMs), such as Mamba, highlighting the trade-offs between long-term semantic context and real-time responsiveness. Furthermore, the Multimodal Perception-Driven Decision-Making (MPDDM) framework is introduced as a conceptual model designed to resolve conflicts between explicit commands and implicit physiological cues via a dynamic confidence-weighting mechanism. To address the limitations of traditional latency-based benchmarks, a new set of evaluation metrics, specifically Modality-Specific Fluidity (MSF), is proposed to quantify the smoothness of interaction. Finally, recent integration paradigms are categorized into modular and end-to-end Vision-Language-Action (VLA) models, offering a critical synthesis of their respective safety and efficiency profiles. This work provides a roadmap for developing verifiable, low-latency HRI systems capable of operating in dynamic, unstructured environments.