CogAAL: A Privacy-Preserving, Full-Stack Cognitive Architecture Integrating Local VLMs, Heterogeneous Robotics, and IoT for Proactive Ambient Assisted Living

We introduce CogAAL, a unified and privacypreserving cognitive architecture for proactive Ambient Assisted Living (AAL). Existing AAL frameworks are typically fragmented across perception, reasoning, and actuation layers, and frequently depend on cloud-based inference pipelines that conflict with the stringent data-sovereignty and confidentiality requirements of residential care environments. In contrast, CogAAL integrates three core modules via a shared semantic representation: (i)a four-layer Multimodal Cognitive Intelligence module that fuses data from wearable devices, Internet-of-Things (IoT) sensors, and RGB-D vision streams to enable robust health-state inference and longitudinal health monitoring; (ii)a three-tier Semantic Robotic Navigation module that semantically grounds a heterogeneous fleet of cleaning, delivery, and mobility-assistance robots in a Building Information Modeling (BIM)-based digital twin, which is continuously updated using real-time occupancy patterns and health-related events; and (iii)an Agentic IoT Orchestration module that translates natural-language intents into executable task graphs and orchestrates adaptive environmental comfort control (including lighting, HVAC, and acoustic conditions). All inference processes are executed on-premises entirely using Gemma 4 MoE models with 2-3.8B active parameters, requiring 8-16 GB of VRAM at 4-bit precision and licensed under Apache 2.0, thereby eliminating the need for cloud-based data transmission. A federated feedback loop enables continuous, privacy-preserving personalization across deployments.