The Intelligent Home: A Systematic Review of Technological Pillars, Emerging Paradigms, and Future Directions

Abstract

Home automation is undergoing a paradigm shift from connected IoT environments with rule-based control to intelligent homes exhibiting ambient intelligence and proactive adaptation. Artificial intelligence, privacy-preserving sensing, and converging connectivity standards are the primary forces driving this transition. This systematic literature review synthesizes the technological foundations, architectural developments, emerging paradigms, and socio-technical challenges characterizing the next generation of smart homes, evaluated against the original Ambient Intelligence (AmI) vision. Following PRISMA 2020 guidelines, searches were conducted across four databases—IEEE Xplore, ACM Digital Library, Scopus, and Web of Science—covering studies published between January 2020 and June 2025. From 3450 records, 113 studies were selected through a two-reviewer screening procedure with inter-rater reliability assessments. Quality was assessed using a modified JBI Critical Appraisal Checklist, and findings were synthesized through thematic analysis. Three converging technological pillars were identified: multi-modal privacy-preserving sensing including mmWave radar; a hierarchical cloud-edge-TinyML intelligence engine; and unified connectivity through the Matter/Thread standard. Emerging paradigms include LLM-based cognitive orchestration, hyper-personalization, Digital Twin simulation, and grid-interactive prosumer energy management. Realizing that the intelligent home vision requires addressing the privacy–security–trust trilemma, algorithmic bias, system reliability, and human–agent collaboration, a research roadmap encompassing explainable AI, privacy-by-design, lifelong learning, and standardized ethical auditing is proposed.

1. Introduction

1.1. Background

There have been three stages of home automation. The original (1970s–1990s) one was based on electromechanical timers and the X10 powerline protocol and had isolated, unidirectional control of devices with no contextual awareness [1,2]. The second stage (2000s–2010s) brought forward the Internet of Things (IoT) with networked devices that can communicate with each other using Zigbee, Z-Wave, and Wi-Fi, allowing remote control with smartphones and cloud computing systems [2,3,4,5]. Nevertheless, these linked houses were still relying on explicit user instructions and weak when-then-that (IFTTT) rule-based logic [6,7,8].

The third stage, which represents the present era of home automation, denotes a shift to intelligent homes that display ambient intelligence (AmI): ambient-aware, proactively adaptive, learning environments that continuously adapt to the needs of their residents over time [6,9,10]. This is facilitated by the intersection of three technological trends, which are improvements in AI and machine learning (in particular, deep learning, transformer architectures, and large language models) [11,12,13,14,15], the maturation of distributed computing between cloud and edge to TinyML [16,17], and the development of standardized connectivity, chiefly the Matter protocol based on Thread networks [18,19].

At the same time, forces in society are creating the need for domestic intelligence. The elderly population needs ambient assisted living systems that can facilitate their independent lives by having unobtrusive health care surveillance [20,21,22]. The sustainability imperative is the key to the creation of AI-based home energy management systems that utilize methods to schedule loads, integrate renewable sources, and respond to demand [23,24]. The COVID-19-increased expectations of homes that support work, health, and well-being at once, such as intelligent zoning of various activities, circadian-friendly lighting that adapts to the biorhythms of occupants, and active indoor air quality control to ensure longer occupancy, are already normalized [25].

1.2. Research Gap and Rationale

Although a significant scholarly effort has been made on smart environments, there is a gap in synthesis in the literature after 2020. The Matter standard itself has been preceded by previous reviews of IoT architectures [3,26,27,28] that do not have the potential to evaluate its influence on interoperability. Before the disruption of generative AI and LLMs to arrange tasks in smart homes, machine learning surveys based on Smart Homes were published [10,13,29]. Disciplinary reviews of ambient assisted living [20,21], home energy management [23,30], and security [31] do not look across the cuts in the technologies that change more than one of the fields at once, including federated learning [32,33] and Digital Twins [34,35]. Moreover, previous challenge-based approaches concentrated on the classical security and privacy issues without sufficiently considering the issue of algorithmic bias, explainable artificial intelligence (XAI), or the collaboration between humans and AI [36,37]. No review reassesses the original AmI vision in relation to the ability of the present technologies.

1.3. Research Questions

• This review is guided by four research questions:

  ▪

  RQ1: What are the core technological foundations (sensing, intelligence, connectivity) enabling the transition from connected smart homes to intelligent homes in the post-2020 era?

  ▪

  RQ2: What emerging paradigms (e.g., LLM-based orchestration, Digital Twins, hyper-personalization) are reshaping human–home interaction?

  ▪

  RQ3: What socio-technical challenges (privacy, security, trust, bias, usability) remain unresolved, and how have they evolved?

  ▪

  RQ4: What research directions are needed to realize the original AmI vision of a collaborative, adaptive, and trustworthy domestic environment?

1.4. Contributions

This review contributes in five aspects (to the best knowledge of the authors): (i) the first post-2020 horizontal synthesis integrating the specific post-2020 instantiations of AI, sensing, and connectivity—namely Matter/Thread standardization (ratified 2022), TinyML on bare microcontrollers, federated learning for domestic privacy, and LLM-based cognitive orchestration—as a convergent system rather than parallel developments (Gap 1; confirms that none of the prior studies covers all five dimensions simultaneously; the novelty lies not in the categories but in their co-maturation and combined architectural implications, which only became assessable after 2020); (ii) cross-cutting software technologies reshaping more than one domain at once (Gap 2); (iii) the change in thinking of the A comparative analysis against seven previous review papers has been provided in (Section 2, Section 3 and Section 4) and the coverage gaps, which justify this synthesis, are explicitly documented there. This review is an integration of two complementary functions that are carefully preserved separately in all its parts: a systematic evidence synthesis (Section 3, Section 4, Section 5 and Section 6, based on the 113 included studies) and a forward-looking research roadmap (Section 7 and Section 8, extrapolating evidence on directions). The cross-domain integration and evidence maturity mapping are the two original analytical contributions of this review, distinct from its narrative synthesis function. Specifically, the Evidence Maturity Assessment framework (M1–M4)—defined by explicit pre-specified criteria, applied independently by two coders at the technology-cluster level with documented inter-rater reliability (κ = 0.78), and resolved conservatively in cases of disagreement—constitutes a structured classification instrument, not a descriptive label. Its function is to operationalize the distinction between commercially validated technologies (M1), empirically validated laboratory evidence (M2), proof-of-concept demonstrations (M3), and conceptual or simulation-only proposals (M4), thereby preventing the conflation of maturity levels that characterizes several prior reviews in this domain. All forward-looking claims in Section 7 and Section 8 are explicitly grounded in this evidence base and clearly marked with their maturity level.

1.5. Paper Organization

The rest is arranged in the following manner. Section 2 examines related work and finds gaps in the research. Section 3 is a description of the systematic review method. The findings on the three technological pillars are given in Section 4. Section 5 looks into trends emerging. Challenges are analyzed in Section 6. A synopsis discussion is given in Section 7. Future research direction is described in Section 8. Section 9 recognizes limitations, and Section 10 is the conclusion. Other materials are included in Appendix A, Appendix B, Appendix C and Appendix D.

• Key Terminology: To ensure clarity throughout this review, the operational definitions for core concepts adopted in this work are presented in Table 1.
  Table 1. Operational definitions of key terminology in home automation and intelligence.

  Term	Operational Definition
  Smart Home	A residence equipped with IoT devices enabling remote monitoring, control, and rule-based automation (e.g., IFTTT). Intelligence is minimal and externally programmed.
  Intelligent Home	A residence exhibiting context-awareness, proactive behavior, continuous learning, and explainable autonomy. The primary subject of this review.
  Ambient Intelligence (AmI)	A paradigm in which environments are sensitive, adaptive, and responsive to human presence, acting as an unobtrusive intelligent partner.
  Edge AI/TinyML	ML inference on local gateways (Edge AI) or microcontrollers (TinyML), enabling low-latency, privacy-preserving processing.
  Matter/Thread	Matter: open-source, IP-based application-layer protocol for smart home interoperability. Thread: low-power mesh networking protocol serving as transport layer for Matter.
  Digital Twin (DT)	A dynamic virtual replica of the physical home, continuously synchronized with real-time sensor data, used for simulation, prediction, and optimization.
  Prosumer Home	A home that both produces and consumes energy, actively participating in grid demand–response programs.

2. Related Work

2.1. IoT Architectures and Early Automation (2010–2018)

The early IoT era was concerned with connectivity, interoperability, and data management [2,3,38,39,40]. Layered architectures, which included perception, network, and application layers, dealt with a heterogeneous device ecosystem [26,27,41,42]. Interoperability has continued to be a thorn in the flesh, with vendor-based ecosystems developing disjointed user experiences [7,18]. Hub-based architectures tried to connect such silos [43,44], whereas cloud platforms allowed aggregation at the center [3,45] (see Figure 1).

Figure 1. Evolution of smart home architectural paradigms. Arrows indicate the direction of data and control flow between system components within each architectural layer. Author-generated synthesis.

Studies focused on energy-conserving MAC protocols for sensor networks to increase battery life [22,46], security architecture to secure vulnerable devices [31,47], and middleware platforms to abstract hardware heterogeneity to enable the same applications to run across heterogeneous hardware.

Deterministic IFTTT logic-controlled automation [6]. The AmI was still of interest [9,10] but was unfortunate to be held back by the limitations of the then-current computers and architectures. These connectivity-era limitations—fragmented interoperability, cloud dependence, and rigid rule-based control—created the very conditions that motivated the next wave of research: applying machine learning to transform accumulated sensor data into adaptive, data-driven home intelligence [48], as examined in Section 2.2.

2.2. Machine Learning for Smart Homes (2015–2022)

The spread of sensor data transformed homes into a data-driven environment [29,49]. HAR (Human Activity Recognition) has developed deep architectures (LSTMs, 1D-CNNs) to extract features automatically [13,29,50,51,52]. The anomaly-detection methods were used to model baselines to detect falls or cognitive deterioration. Non-Intrusive Load Monitoring (NILM) has evolved from factorial Hidden Markov Models (HMMs) to transformer-based attention mechanisms [53,54] (see Figure 2). However, privacy threats, latency problems, generalization problems, and concept drift were the weak points of the existing cloud-centric architecture [6,16,55,56,57,58]. Hence, these limitations reflected a potential threat to one-dimensional machine learning capability, which also relied on a data-centric, independent, and heterogeneous client-server architecture. The human-centered vision of an adaptive and intelligent responsive home environment was yet to be realized [6]. This led to the identification of the Ambient Intelligence paradigm discussed in Section 2.3. Instead of machine learning, many years back, there was a focus on human-centered vision to utilize technology in a responsive home environment. Unlike previously discussed technologies, Ambient Intelligence seems to demonstrate a strong linkage to today’s developments to ensure its applicability in homes as predicted previously.

Figure 2. The data-driven intelligence pipeline in smart homes. Solid arrows indicate the sequential flow of data across pipeline stages; dashed arrows represent the feedback signal for iterative model retraining; downward arrows denote internal processing hierarchies within each module.

2.3. The Ambient Intelligence Paradigm

AmI stipulated the human-centered vision guiding [9,10], according to which the environment should be sensitive, adaptive, responsive, and transparent. Instead of providing the convenience of the remote control, AmI projected the home as an invisible butler—intelligent partner that can foresee and address the new needs of the occupants [6,59]. AmI extended the idea of ubiquitous computing (see Figure 3) to include the concept of an intelligent and user-centric response: contextual intelligence, personalization, adaptiveness, and proactive response [9,10,60].

Figure 3. The confluence of driving forces in intelligent home automation. The blue arrow (PUSH) represents technology-driven forces propelling smart home adoption, while the green arrow (PULL) denotes societal demands driving technological development. Small arrows within the Societal Catalysts panel indicate interconnections between sustainability and lifestyle factors.

Context-awareness is the gist of AmI, where context moves beyond the mere sensor events of a home environment to a multi-dimensional and changing image. They have singled out several important types of context: Physical Context—location and occupancy of users through mmWave radar [61] or Wi-Fi CSI sensing [62,63], ambient temperature, lighting, and noise levels [64]; Temporal Context—time of day, day of the week and season, duration of an activity [6,65,66]; Personal Context—user identity (gait recognition) [67] or voice [68], current activity through HAR [29,69], emotional state (based on behavioral patterns).

Advanced context-aware systems usually use a central context manager or reasoning engine that combines ambiguous low-level information (e.g., “kitchen motion,” “oven on,” “water flow”) into higher-level inferences (e.g., user is making dinner) via rule-based systems, ontologies [70,71,72], and probabilistic models (e.g., Dynamic Bayesian Networks) [73,74,75]. There has been substantial interest in developing formal ontologies of home settings as a tool to develop common terms and logical structures of rooms, devices, and activities, and increase semantic interoperability [75,76].

The strong context-awareness allows active service delivery. In contrast to reactive IFTTT triggers, AmI systems can identify intent by looking at a combination of data points—such as seeing somebody headed to the bedroom, bathroom lights on, and time of the day to conclude that somebody is going to sleep and turn on sleep doors, first-floor lights, and the thermostat [6,77]. This paradigm change of reactivity to proactive and contextual anticipation is a qualitative jump in system intelligence [6,10]. Nevertheless, before actualization was technically feasible, concepts had already reached a high level of conceptual maturity, but this was initially limited by computational capability and the privacy considerations of pervasive sensing [10,58].

2.4. The Evolutionary Journey of Home Automation

Table 2 summarizes the key paradigms, characteristics, and enabling technologies across four evolutionary eras of home automation.

Table 2. The evolutionary journey of home automation—key paradigms, characteristics, and technologies across different eras.

Era	Paradigm	Key Characteristics	Enabling Technologies	User Interaction	Refs
~1970s–1990s	Simple Automation	Isolated devices Reactive, pre-programmed No context-awareness One-way communication	X10 Powerline Protocol Electromechanical Timers Infrared Remotes	Manual and Physical (Switches, Dials, Timers, Remotes)	[4,22,78]
~2000s–2010s	Connected Smart Home (IoT Era)	Remote control and monitoring Device ecosystems “Walled garden” interoperability Cloud-centric data aggregation Brittle IFTTT rule-based automation	Zigbee, Z-Wave [19,79] Wi-Fi microcontrollers Smartphones and Central Hubs Cloud Platforms	Explicit and Remote (Smartphone Apps, Web Portals, Basic Voice Commands)	[2,3,4,7,26,43,44,78,80,81,82]
~2010s–Present	Intelligent Home (Ambient Intelligence)	Context-aware and proactive Multi-modal sensor fusion ML and personalization Distributed intelligence (Cloud-Edge) Predictive behavior modeling	Multi-sensor Fusion [83,84,85] Machine/Deep Learning for HAR [13,29,50] Edge Computing and TinyML [32] Natural Language Processing	Implicit and Proactive (Environment adapts; Limited explicit input required)	[1,6,9,10,13,20,29,32,34,73,83,84,86,87,88,89,90,91]
Emerging (Present–Future)	Cognitive Home	Predictive and self-optimizing Collaborative human–AI teaming Explainable AI (XAI) Digital Twin simulation Generative task orchestration	Large Language Models [12,13,14] Digital Twins [92,93,94] Explainable AI [36,37,58] Advanced Robotics [95,96] Federated Learning [97]	Symbiotic and Collaborative (Natural Dialogue, Negotiation, Co-learning between user and system)	[6,12,13,14,58,65,74,92,93,94,95,97,98]

2.5. Identification of Research Gaps

Table 3 provides a comparative analysis of seminal review papers, identifying five research gaps that this review addresses.

Table 3. Comparative analysis of seminal review papers on smart and intelligent homes (with added columns for number of studies and quality assessment tool).

Review Focus & Citations	Temporal Scope & Paradigm	Core Technological Emphasis	Key Challenges Identified	Limitation/Gap vs. Post-2020	Studies (n)	Quality Tool
IoT Architectures and Connectivity [3,19,26,27,99]	~2010–2018; Cloud-Centric/Hub-Based	Wireless Protocols (Zigbee, Z-Wave, Wi-Fi), Layered Architectures, Middleware	Interoperability, Security, Energy Efficiency, Scalability	Predates Matter/Thread standard; lacks edge-native architecture discussion	Varies	None
ML for Smart Homes (HAR, NILM) [10,13,29]	~2012–2020; Cloud-Centric	Supervised Learning (SVM, HMM), Early Deep Learning (LSTM, CNN)	Data Availability, Model Accuracy, Computational Cost, Personalization	Misses Generative AI and LLMs for orchestration and natural interaction	Varies	None
AAL and Healthcare [20,21,96,100]	~2005–2020; Varies, often cloud-reliant	Wearable and Ambient Sensors, Activity Recognition, Anomaly Detection for Elderly Care	Privacy, Sensor Accuracy, User Acceptance, Cost	Vertical silo; does not integrate DTs for simulation or FL for privacy	Varies	JBI (partial)
Home Energy Management [23,35,72,101,102]	~2014–2021; Cloud-Centric	Optimization Algorithms, Demand Response, Appliance Scheduling	User Comfort, Grid Integration, Cost Optimization	Treats energy in isolation; lacks synthesis with edge AI or multi-agent systems	Varies	None
Security and Privacy [103,104,105,106]	~2014–2020; Cloud-Centric assumed	Network Intrusion Detection, Encryption, Vulnerability Analysis of IoT devices	Botnets (Mirai), Data Leakage, Device Vulnerabilities	Under-explores algorithmic bias, XAI for trust, physical adversarial attacks	Varies	None
Ambient Intelligence (Vision) [1,9,10,87]	~2000–2010; Conceptual/Architectural	Ubiquitous Computing, Context-Awareness, Human-Centric Design	Privacy, System Reliability, Natural Interaction, “Calm” Technology	Visionary roadmap lacking modern enablers (Transformers, mmWave, TinyML)	N/A	None
Edge AI and TinyML for IoT [107]	~2019–2025; Edge-Centric/MCU-Deployed	TinyML on Microcontrollers, Model Compression (Quantization, Pruning), Hardware-Software Co-design, Federated + Edge Inference	Power Efficiency, Memory Constraints, Benchmark Standardization, On-device Privacy	Covers general IoT/smart city deployment; does not integrate home-specific connectivity standards (Matter/Thread) or LLM-based cognitive orchestration in residential settings	66	PRISMA
Federated Learning Privacy in Edge-IoT [108]	~2020–2024; Decentralized/Edge-Distributed	Federated Averaging, Differential Privacy, Secure Aggregation, Byzantine-Robust Aggregation, Non-IID Handling	Data Heterogeneity, Communication Overhead, Model Poisoning, Privacy Leakage	Treats FL as a standalone privacy mechanism; does not synthesize with home-specific sensing pipelines, TinyML deployment on domestic MCUs, or AmI orchestration goals	Varies	PRISMA
LLMs in IoT Ecosystems [109]	~2020–2024; Cloud + Lightweight Edge Hybrid	LLM-IoT Integration, Natural Language Interfaces, Lightweight LLMs for Resource-Constrained Devices, Semantic Task Decomposition	Latency, Hallucination, Computational Cost on Edge Devices, Security and Privacy of LLM Queries	Addresses LLM-IoT broadly across smart cities, agriculture, and healthcare; does not engage with the residential AmI vision, privacy-by-design sensing, or cross-standard (Matter/Thread) home interoperability	30	Structured literature search
Human-Centered Smart Home Research [110]	~2000–2022; Technology-Push vs. User-Centric Transition	User Adoption, Behavioral Modeling, UX Design, Trust Calibration, Accessibility for Diverse Populations	User Acceptance Gap, Technology-Push Bias, Underserved Populations, Multi-Stakeholder Needs	Adopts bibliometric and UX lens; does not address the technical convergence of sensing, AI, and connectivity enabling next-generation intelligent homes, nor the AmI reassessment this review undertakes	7003 papers analyzed	Bibliometric + SLR
This Review	Post-2020 Convergence; Hybrid Edge-Cloud	Matter/Thread, Edge AI/TinyML, Generative AI/LLMs, Digital Twins, Federated Learning	XAI and Trust, Algorithmic Fairness, Human-Agent Collaboration, Socio-Technical Complexity	Fills gap: horizontal synthesis of disruptive technologies and their collective impact across all five dimensions	113	Modified JBI

The five gaps identified include the following: (1) lack of post-convergence synthesis; (2) a discrete and not integrated treatise of the cross-cutting technologies; (3) cloud-based and not hybrid architecture study; (4) technical and not socio-technical problem-setting; and (5) no reevaluation of the AmI vision in relation to modern enablers. Importantly, even the most recent post-2022 reviews do not close these gaps individually or collectively. Reviews of Edge AI and TinyML [107] address on-device inference at a general IoT level but do not connect this to home-specific connectivity standards or cognitive orchestration. FL surveys [108] treat privacy-preserving learning in isolation from domestic sensing architectures. LLM-IoT reviews [109] cover natural-language interfaces broadly but do not engage with the residential AmI vision or the trilemma of privacy, security, and trust specific to the home environment. Last but not least, human-centered evaluations [110] cover some aspects regarding the UX of these systems. However, the method they adopt is not harmonized with the technical convergence delivered by this review. As a result, there is no novelty introduced in one specific technology, but a unique synthesis that horizontally interconnects all five dimensions of the current architecture—namely, privacy-preserving ubiquitous sensing, tiered intelligence, interoperable connectivity, innovative cognitive approaches, and socio-technical ramifications—along with its corresponding state-of-the-art evidence base, following the PRISMA guidelines, in the context of intelligent home applications.

3. Methodology

This review follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [111]. PRISMA is an internationally recognized methodological framework designed to ensure transparency, reproducibility, and rigor in systematic literature reviews. It has a structured protocol that includes four steps: finding records through database searches, screening titles and abstracts, checking the eligibility of full texts, and adding final studies. Each step has clear reporting requirements. PRISMA 2020 was selected for this review for three reasons: it is the current gold standard for systematic evidence synthesis across disciplines; it accommodates heterogeneous study designs, including empirical, system design, and conceptual papers, which characterizes the intelligent home literature; and its structured reporting requirements enable readers to independently verify the review process and assess the risk of bias. The PRISMA flow diagram is presented in Figure 4, and all reporting checklist items are addressed across Section 3.1, Section 3.2, Section 3.3, Section 3.4 and Section 3.5.

Figure 4. PRISMA 2020 flow diagram.

3.1. Search Strategy

An electronic search was done in four databases, including IEEE Xplore, ACM Digital Library, Scopus, and Web of Science Core Collection (see Figure 4). The search was limited in time to January 2020–June 2025. Thematic synthesis was chosen as the method for integrating evidence instead of meta-analysis or statistical aggregation because the corpus includes a variety of study designs, such as empirical evaluations, system design papers, review articles, and conceptual frameworks, with outcome measures that cannot be compared across domains. In these circumstances, PRISMA 2020 guidelines advocate for thematic synthesis as the methodologically suitable approach; employing quantitative aggregation across such diverse designs would yield illusory precision rather than legitimate analytical conclusions [111].

The search query was based on three conceptual pillars, namely domain (“smart home” OR “intelligent home” OR “home automation”), technology (“artificial intelligence” OR “machine learning” OR “edge computing”), and scope (“trend” OR “framework” OR “review”). The pillar of scope was also added to accommodate the architectural and synthesis-based work that defines the current trend of the field; it did not include empirical papers, and those were also obtained via backward reference searching, forward citation tracking, and a manual search of the leading venues. To evaluate sensitivity, a pilot search was conducted on IEEE Xplore without the scope pillar that retrieved 2341 records, and a 50-record random sample was reviewed manually, indicating that 94% of the added records were either out-of-scope primary hardware studies or duplicates of existing records had been included, which justifies retention of the tripartite query design. All database-specific search query adoptions can be found in Appendix D. The sensitivity analysis result, which showed that 94% of records retrieved without the scope pillar were out-of-scope hardware studies or duplicates, is important because it is consistent with the corpus composition reported in Section 3.5. This shows that the tripartite query did not create a conceptually biased corpus in practice.

3.2. Inclusion and Exclusion Criteria

Inclusion: the articles published between 2020 and 2025, peer-reviewed journal articles, conference papers, or systematic reviews, English language, and devoted to the development of AI/ML, architectural innovations, context-aware sensing, or critical review of emerging issues related to intelligent home settings. Exclusions: books, editorials, patents, theses; all-hardware studies that do not involve AI elements; all tangential aspects of smart home; smart city/industrial IoT; non-residential pathway; duplicates. Publications published before 2020 are mentioned as background materials but not included in the 113 systematically reviewed articles (see Figure 4).

3.3. Study Selection Process

Phase 1—Identification: The database search yielded 3422 records; supplementary searches added 28, totaling 3450. Phase 2—Screening: Deduplication removed 532 duplicates (2918 unique). Title/abstract screening by two independent reviewers (pilot calibration: Cohen’s κ = 0.84) excluded 2420 records, advancing 498. Phase 3—Eligibility: Two reviewers independently assessed 498 full texts; 385 were excluded, and 113 were included. Categorized exclusion reasons: wrong topic or domain (n = 148, 38%); insufficient AI/ML component—purely hardware or protocol-only studies (n = 89, 23%); wrong publication type—editorials, theses, extended abstracts (n = 52, 14%); insufficient methodological quality (low JBI score; n = 41, 11%); duplicate or superseded by a more comprehensive work already included (n = 31, 8%); and not in English (n = 24, 6%). Exclusion reasons were independently coded by both reviewers; inter-rater agreement on the exclusion category was κ = 0.79 (see Figure 5).

Figure 5. Schematic of the systematic search strategy.

3.4. Quality Assessment

Methodological quality was assessed using a modified JBI Critical Appraisal Checklist adapted for the heterogeneous study designs in this corpus (empirical, system design, review/survey, and analytical/conceptual papers). Two reviewers independently assessed each study; disagreements were resolved through discussion and, where consensus could not be reached, by a third-reviewer tiebreak. The inter-rater agreement for quality scoring was Cohen’s κ = 0.81 (substantial agreement). Each study was scored against eight criteria (maximum 16 points): (C1) clear research objective or question; (C2) adequate description of participants/systems/data; (C3) valid and reliable measurement or data collection method; (C4) appropriate analytical or experimental method for the study type; (C5) control of confounds or alternative explanations; (C6) findings supported by the data presented; (C7) ethical or reproducibility considerations addressed; (C8) conclusions consistent with findings. For non-empirical studies (conceptual, system-design, review papers), criteria C3, C4, and C5 were adapted: C3 assessed whether the methodology or design rationale was transparently described; C4 assessed whether the evaluation or validation approach was appropriate for the study type; and C5 assessed whether limitations and scope boundaries were acknowledged. Scores were mapped to quality bands: High, ≥12/16; Moderate, 8–11/16; Low, <8/16. Low-quality studies were retained but flagged; their evidence was weighted conservatively and not used as primary support for conclusions. The quality distribution is reported in Table 4, and individual scores are provided in Appendix C.

Table 4. Quality assessment summary—distribution of included studies across quality ratings by study type.

Study Type	High (n)	Moderate (n)	Low (n)
Empirical (53)	31	19	3
System Design (32)	14	15	3
Review/Survey (17)	12	5	0
Analytical (11)	5	5	1
Total (113)	62 (55%)	44 (39%)	7 (6%)

3.5. Data Extraction and Synthesis

A systematic extraction form was used: bibliographic features, type of study, technology of interest, field of application, methodology, essential conclusions, and limitations. Inductive coding in NVivo (v14) was used to perform thematic synthesis following the methodology proposed by Thomas and Harden (2008) [111]. The coding was conducted in three steps: (1) line-by-line open coding of the results of each included study, generating initial descriptive codes (e.g., on-device inference latency, occupant privacy concern); (2) grouping codes into descriptive themes based on semantic similarity (e.g., edge inference, TinyML deployment, and MCU power constraints coalesced into the TinyML tier); and (3) organizing analytical themes into the structure of Section 4, Section 5 and Section 6.

The two reviewers independently coded 20% of the studies; inter-rater reliability for the theme assignment was κ = 0.76 (substantial agreement). Disagreements were resolved through discussion, and the codebook was refined over three iterative review cycles. One reviewer performed data extraction, and a second reviewer verified all 113 studies; discrepancies were resolved before synthesis. The 17 included review/survey articles were treated as sources of contextual triangulation rather than primary units of evidence: their findings were used to support or contrast trends identified in primary studies but were not counted as empirical evidence. This approach avoids the duplication of evidence where review articles overlap with primary sources. Where a claim was supported solely by review literature, it was explicitly identified as such. A hierarchy of evidence weighting was applied across study types: (1) multi-site empirical field deployments with end-users; (2) single-site empirical studies; (3) laboratory/simulation studies; (4) system design papers with partial evaluation; (5) review and survey papers (contextual); and (6) conceptual and theoretical papers. This hierarchy informed confidence levels in Table 5. The complete data extraction table (n = 113) is provided in Appendix B. Descriptive summary: publication years 2020 (14), 2021 (22), 2022 (19), 2023 (24), 2024 (21), 2025 (13); study types: empirical (47%), system design (28%), review (15%), conceptual (10%); domains: general (32%), health/AAL (24%), energy (22%), security (14%), other (8%). These study types were not treated as equivalent in the synthesis. Formal evidence weighting hierarchy was applied throughout: (1) multi-site empirical field deployments with real end-users carried the highest evidential weight; (2) single-site empirical studies; (3) laboratory and simulation studies; (4) system design papers with partial evaluation; (5) review and survey papers used for contextual triangulation only, not as primary evidence units; and (6) conceptual and theoretical papers. Claims in Section 4, Section 5 and Section 6 are grounded in this hierarchy, and readers can verify the evidence type supporting any specific claim by consulting the study design column in Appendix B, where each of the 113 included studies is classified by design type and evidence maturity label.

Table 5. Evidence maturity assessment of key technologies and paradigms.

Technology/Paradigm	Evidence Maturity (M1–M4)	Deployment Status	Studies (n)
HEMS/Energy Optimization	M1 (Commercial/Validated)	Commercial products	~25
HAR/Fall Detection	M2 (Empirically Validated)	Pilot deployments	~22
Edge AI/TinyML	M2 (Empirically Validated)	Lab + limited field	~15
Matter/Thread	M2 (Empirically Validated)	Early commercial	~12
Digital Twins (Energy)	M2 (Empirically Validated)	Simulation studies	~10
Federated Learning	M3 (Proof-of-Concept)	Lab benchmarks	~8
LLM Orchestration	M3 (Proof-of-Concept)	Proof-of-concept	~7
Affective Computing	M3 (Proof-of-Concept)	Conceptual/lab	~5
Swarm Robotics	M4 (Conceptual/Early-Stage)	Simulation only	~4

4. The Technological Pillars of Modern Intelligent Homes (RQ1)

Building on the systematic methodology described in Section 3, the following sections present the synthesized findings, beginning with the core technological pillars that underpin modern intelligent home systems.

The progression from connected to intelligent homes is driven by three interdependent pillars: perception, intelligence, and connectivity. It is important to clarify that these pillars are not generic IoT architectural layers—they represent specific technological instantiations that reached simultaneous practical maturity only in the post-2020 period. The first pillar is not sensing in the broad IoT sense, but specifically the emergence of privacy-preserving modalities—chiefly mmWave radar and Wi-Fi CSI—that enable rich, room-level behavioral inference without cameras, a capability that became deployable in domestic settings only after 2020 [61,62,63]. The second pillar is not distributed intelligence generically, but the crystallization of a three-tier computing continuum—cloud for large-model training and federated aggregation, edge gateways for sub-100 ms local inference, and TinyML on bare microcontrollers for milliwatt-level always-on sensing—as a coherent, co-designed architecture rather than isolated deployment choices [16,17,32]. The third pillar is not connectivity in general, but the 2022 ratification of the Matter protocol as the first cross-vendor, IP-native unification layer for smart home devices, an event that structurally changes the interoperability landscape in ways that all pre-2022 reviews could not assess [18,19]. Critically, it is the simultaneous convergence of these three specific developments—not any one of them in isolation—that creates the qualitatively new design space within which the emerging paradigms of Section 5 (LLM-based cognitive orchestration, Digital Twins, hyper-personalization), and the socio-technical challenges of Section 6 become both technically possible and socially urgent.

4.1. Advanced Sensing and Perception

▪

Visual and RF-Based Sensing

Visual and RF-based sensing constitute two primary modalities for perception in intelligent homes, each characterized by different trade-offs between accuracy and privacy. Camera-based systems provide high-fidelity human activity recognition (HAR) but introduce significant privacy concerns [23,49,95,112,113]. Conversely, RF-based techniques, such as mmWave radar, enable trajectory tracking, fine-grained motion detection, and vital sign monitoring (e.g., heart rate and respiration) without capturing identifiable visual information [61,62,114,115]. IR thermal imaging is used to offer occupancy detection based on body heat with less intrusion [84,116]. Wi-Fi Channel State Information (CSI) sensing utilizes the infrastructure that is already in place to monitor the presence and activity based on signal perturbations as a device-free, infrastructure-reusing technology [62,63]. Figure 6 illustrates the trade-off between information fidelity and privacy intrusiveness across these sensing modalities, demonstrating how privacy-preserving alternatives such as mmWave and Wi-Fi CSI can achieve comparable behavioral inference to camera-based systems while substantially reducing privacy risk.

Figure 6. Balancing information fidelity and user privacy in intelligent home sensing.

▪

Acoustic, Environmental, and Power-Based Sensing

Low-cost VOC, PM2.5, and CO₂ environmental sensors can be used to achieve environmental sensor fusion and thus dynamic Indoor Air Quality (IAQ) management [49,64]. Appliance disaggregation has been significantly enhanced using transformer-based NILM models [117].

▪

Multi-Modal Sensor Fusion

Fusion unites heterogeneous data streams on data, feature, and decision levels into single context models [83,84,98,118]. Evidence maturity varies: camera-based HAR is well-validated; mmWave shows promising but limited deployment evidence; multi-modal frameworks remain predominantly proof-of-concept.

4.2. The Intelligence Engine: A Hierarchical Computing Continuum

▪

Cloud Tier

The cloud tier is essential for large-scale model training, federated learning aggregation [32,33,119], and hosting billion-parameter LLMs [12,13]. However, sole reliance on cloud processing introduces significant constraints, including latency, privacy exposure, reliability risks, and bandwidth limitations [16,58] (see Figure 7).

Figure 7. The hierarchical intelligence engine of the modern intelligent home.

▪

Edge Tier

The edge tier delivers low-latency local decision-making, directly supporting time-critical applications such as fall detection and voice response [16,114,120,121]. It also enhances privacy by processing data locally rather than transmitting it to the cloud [16,55,58,122] and maintains operational continuity during internet outages [16,123].

▪

TinyML Tier

The TinyML tier deploys ultra-lightweight machine learning models directly on microcontrollers, enabling milliwatt-level always-on sensing and on-sensor data preprocessing [17,124,125,126,127]. Empirical evaluations of TinyML in real domestic settings remain limited; most included studies report laboratory benchmarks rather than field deployments.

4.3. The Connectivity Fabric: Matter and Thread

The Matter protocol offers cross-vendor IP-based interoperability [18,19], and Thread is the low-power mesh-based transport with self-healing topology [18,100,128]. The Wi-Fi is the high-bandwidth data backbone of data-intensive applications like video streaming. Bridges provide pre-Matter compatibility (e.g., Z-Wave, Zigbee), thereby preserving existing consumer investments. Figure 8 summarizes this unified connectivity fabric, showing how Matter, Thread, Wi-Fi, BLE, and LPWAN protocols occupy complementary roles across different device types, data rates, and range requirements within a single intelligent home architecture. The Bluetooth Low Energy (BLE) is a critical component of commissioning (onboarding) new Matter devices [18,104]. In specialty applications, Low-Power Wide-Area Network (LPWAN) systems like LoRaWAN have been proposed as a highly long-range, low-bandwidth capture data of large estates where Wi-Fi and Thread cannot reach [22,129]. Shortcomings consist of partial coverage of device categories, slow uptake, the so-called Matter-Plus vendor lock-in phenomenon, where vendors only give innovative functionality to proprietary systems, and an open semantic interoperability gap [18,19].

Figure 8. The unified connectivity fabric of the intelligent home.

5. Emerging Trends and Paradigms (RQ2)

Having established the foundational technological pillars of intelligent homes, we now examine the emerging paradigms that build upon these capabilities and are reshaping human–home interaction.

5.1. Generative AI and LLM-Based Cognitive Orchestration

LLMs enable three qualitatively new capabilities for home intelligence.

First is the semantic interpretation of open-ended intentions. Rather than parsing rigid commands, LLMs understand abstract natural-language directives. For example, a request such as “I want to watch a movie in a comfortable place—it is a bit too hot in here” can elicit coordinated actions across air conditioning, lighting, and media playback systems without requiring the user to specify each device individually [12,13,130].

Second is reasoning about time and context. LLMs use real-time sensor data and historical patterns, such as the time of day, recent device activations, and noise logs, to come up with smart answers. The system uses the location, time, and activation history to give a contextual answer when someone asks, “What was that noise?” This is a big step up from following commands to understanding conversations.

Third is multi-step task planning and orchestration. The LLM acts as a cognitive orchestrator, generating unique action plans across multiple devices in response to high-level requests. A command such as “Get everything ready before my parents come this weekend” can trigger calendar inspection, thermostat adjustment to age-appropriate temperatures, and the reconfiguration of security settings [12,13,131]. LLMs also enable a shift from user-pulling—user-initiated commands—to system-pushing—proactive suggestions based on continuously updated models of user preferences and habits.

It is important to note that the evidence base for these capabilities is limited. The hallucination risk, inference latency, and real-world deployment evidence remain unestablished. LLM-based home orchestration is rated largely conceptual (M4) or early proof-of-concept (M3) in this review. No included study reported a longitudinal field implementation involving end-users in a real home. The assertions in this section are therefore research directions and design proposals, not proven domestic implementations (see Figure 9).

Figure 9. The LLM as the central cognitive orchestrator in the intelligent home. Downward arrows indicate the flow of inputs into the orchestrator and the propagation of outputs from it; horizontal arrows between processing stages denote their sequential execution order.

5.2. Hyper-Personalization and Behavioral Learning

Environmental personalization (climate zones, lighting) is the most mature application [6,30,70,117,132,133]. Health personalization extends to sleep optimization and nutrition [20,134]. Multi-occupant profiling requires preference arbitration via multi-agent systems [67,135,136]. Affective computing, inferring emotional states for responsive routines, is classified as early-stage/conceptual (M4) in this review: no included study provided field-validated evidence of reliable affective state inference in uncontrolled domestic settings [10,71,87] (see Figure 10).

Figure 10. The continuous learning loop for hyper-personalization in intelligent homes. Downward arrows indicate the sequential flow across pipeline stages; the horizontal arrow denotes internal processing within the user modeling stage; the return path from Stage 4 back to Stage 1 represents the continuous learning cycle through which user feedback iteratively refines the personalization policy.

5.3. Digital Twins for Home

Core DT applications include MPC for HVAC optimization [132,137,138], integration with renewable energy systems [34,138,139], and predictive maintenance through sensor-to-model comparison [34,105,140]. Advanced applications extend to safety simulation [34,35,141] and AAL activity recognition [20,53]. Energy optimization has the strongest empirical support; safety and health applications remain largely conceptual (see Figure 11).

Figure 11. The architecture and function of a home Digital Twin. Downward arrows indicate the flow of real-time sensor data from the physical home to the data bridge layer; upward arrows represent optimized control commands returned to the physical environment; bidirectional arrows between the data bridge and digital twin denote two-way synchronization of IoT data and control signals; dashed arrows indicate internal communication between IoT protocol endpoints and the edge gateway.

5.4. Intelligence for Health and Well-Being

The AAL applications mmWave-based fall detection [114], gait monitoring [20,116], ADL tracking [20,53] make it possible to use aging-in-place [46]. Passive physiological measurements of heart rate, respiration, and quality of sleep, which qualify as clinical-grade remote patient monitoring (RPM), hold promise to enable constant health monitoring without the wearable burden [62,142]. The wellness apps include optimizing sleep and individualized fitness advice [20,116,134,143,144]. This is a field that has comparatively good empirical data on fall detection and ADL monitoring. Figure 12 maps the layered architecture of the intelligent home as a health and well-being guardian, illustrating how passive sensing, physiological monitoring, and wellness applications combine to support aging-in-place across varying levels of clinical acuity.

Figure 12. The intelligent home as a health and well-being guardian. Upward arrows indicate the sequential flow of health data across the four processing layers, from raw multi-modal data acquisition (Layer 1) through AI analytics (Layer 2) and personalized intervention (Layer 3) to final goal achievement (Layer 4); arrow colors correspond to each respective layer. The curved arrow on the right denotes the continuous feedback loop through which outputs from upper layers iteratively refine processing in lower layers.

5.5. Sustainability and Prosumer Energy Intelligence

Machine learning in HEMS (AI) is used to predict future loads and do strategic shifting. The home is an element in a virtual power plant because the administration of demand responds automatically through grid-interactive functionality. Extensions include the ability to connect electric vehicles to the home (V2H) and the grid (V2G), active management of the building envelope, and precise water management [23,24,30,139,145,146,147]. Energy management is one of the most advanced fields in terms of research (see Figure 13).

Figure 13. The intelligent home as a grid-interactive prosumer. Solid blue arrows indicate the flow of input data into the AI decision core and the propagation of control commands to resource management modules; yellow bidirectional arrows denote two-way energy exchange between home resources and controllable loads; bidirectional arrows in the energy flow panel represent grid import and export interactions between the smart home and the electrical grid; and solid arrows toward the sustainability outcomes panel represent the measurable environmental and economic impacts of the energy management strategy.

5.6. Swarm Robotics and Multi-Agent Systems

Multi-agent coordination to distribute tasks among heterogeneous teams of robots is necessary, as well as collision-free multi-agent path planning (MAPP) to operate in shared dynamic settings and extensive integration of the intelligence infrastructure of the home [48,106,148]. The Digital Twin has two purposes: as a testing field and control center, which presents a spatial simulation of the house to all robots and assists in coordinating their actions. The connection fabric creates the effect of near-real-time inter-robot communication. The situational awareness of the swarm is complemented by the shared perception of fixed home sensors (e.g., security cameras). This paradigm is the most hypothetical, and the literature reviewed is mostly in the form of conceptual frameworks and simulations. Evidence rating: the paradigm is M4 (Conceptual/Early-Stage) in all studies that were included in this paradigm. The literature reviewed is mostly conceptual frameworks and simulations of multi-robots; no study in it introduced a domestic field implementation of a heterogeneous robotic swarm with actual end-users. The assertions in this part are therefore research directions and architectural conjectures and not proven applications. Figure 14 depicts the conceptual architecture of a domestic robotic swarm, illustrating how heterogeneous robot teams, shared home sensor perception, Digital Twin coordination, and the Matter/Thread connectivity fabric are envisioned to work together, while making clear that this integration remains at the simulation and conceptual stage with no field-validated domestic implementations reported in the included corpus.

Figure 14. The The domestic robotic swarm: a multi-agent system for physical automation. The downward arrow denotes AI-generated control commands dispatched from the central coordination layer to the robotic execution fleet; the upward arrow represents real-time sensor feedback and state updates returned to the intelligence layer to enable continuous coordination and adaptive task allocation.

6. Challenges and Open Issues (RQ3)

The following challenges were identified inductively through a thematic synthesis of the 113 included studies and the broader cited literature; they are presented as evidence-identified unresolved issues rather than independently asserted propositions, and readers are directed to the evidence type labels in Appendix B to assess the evidentiary basis of any specific claim.

6.1. The Privacy–Security–Trust Trilemma

Hyper-personalization requires continuous multi-modal data collection, which creates significant privacy pressure [36,58,149,150]. Federated learning and edge processing offer partial mitigation [16,32,91,118], but genuine data sovereignty remains unachieved [16,151,152].

The expanding device ecosystem creates large attack surfaces. Inexpensive, high-volume sensors represent the most frequent entry points for attackers [31,58,153]. Physical adversarial attacks—using specially crafted signals to deceive camera, LiDAR, and microphone systems—are an emerging safety-critical threat [47,106].

The black-box problem compounds these security concerns. Deep learning and LLM decision-making processes lack transparency. When an LLM assistant refuses to act without explanation, it causes user frustration and erodes trust. Developing XAI systems capable of providing real-time, intelligible explanations of LLM reasoning remains an open research problem [36,37,154,155].

6.2. Interoperability and Fragmentation

Three interoperability issues remain despite the Matter standard. It should be noted at the outset that the evidence base for these limitations is partially architectural—derived from peer-reviewed analysis of Matter’s design and early deployment literature rather than large-scale empirical field studies, of which only three field-oriented studies were identified in the included corpus. The analytical claims that follow should therefore be read as evidence-informed architectural assessments rather than empirically established findings.

First, the legacy device bridge burden introduces additional error risks and architectural inefficiencies when connecting pre-Matter devices to the new standard. Second, the Matter-Plus vendor lock-in dilemma arises because Matter defines a shared grammar of communication but does not mandate a shared vocabulary of device capabilities, leaving vendors free to restrict advanced functionality to proprietary systems. Third, the semantic interoperability gap means that different brand devices can interpret the same high-level commands, like “create a relaxing setting,” in different ways if they do not all use the same ontological model. This disconnect is especially bad when LLMs are involved because these models need a lot of semantic knowledge to make coherent action plans for multiple devices [18,19].

6.3. Ethical and Societal Implications

Ethical and socio-technical concerns were explicitly raised in at least nine included studies and across multiple domains of the corpus [25,36,58,103,149,150,151,156,157], as well as in the broader cited literature on responsible AI and smart home governance. Algorithmic bias—whereby demographic groups underrepresented in training data are subjected to systematically higher error rates—carries the risk of so-called digital redlining, where access to premium health and safety features becomes contingent on adherence to majority behavioral norms [36]. This concern was raised primarily in security, HAR, and health-monitoring studies within the corpus, where performance disparities across age, gender, and mobility profiles were documented or flagged as unaddressed limitations. Coercive control through smart home technology represents a distinct and underexplored ethical dimension: offenders can exploit cameras, sensors, remote locking systems, and environmental controls as instruments of domestic abuse [58,149,156]. This problem was noted in a few studies included in the collection and the wider literature cited on it; it is comparatively underrepresented in the corpus because it has been under-published rather than not causing real-world harm, as it is discussed in the limitations in Section 9. Lastly, the digital divide, a phenomenon fueled by the high price of both hardware and subscriptions and broadband connection, implies that intelligent home functions will continue to be available to only wealthy residents, further separating developed and developing areas [25,151,158]. It should be noted that the empirical measurement of these socio-technical harms within the included corpus remains limited; the claims in this section are therefore presented as identified concerns requiring dedicated empirical investigation rather than established prevalence findings, consistent with the evidence maturity boundaries discussed in Section 7.2.

6.4. Reliability and Robustness

Cascading single points of failure. Single points of failure in critical functions may be cascaded—one central connection point may be critical to automation, security monitoring, and health monitoring, and a single false negative of a fall detection algorithm may trigger an unattended medical emergency [6,56]. ML models that are trained on curated data are brittle in the presence of the long tail of real-world domestic experiences, e.g., pets playing with sensors or odd visitor behavior outside of training data [6,56].

6.5. Human Factors and Usability

Calibration failures occur when user expectations of AI capabilities do not match what the system can actually deliver, leading to frustration and eventual abandonment [6,25,36,37]. A key driver of this mismatch is the black-box nature of LLM reasoning—users who cannot construct an accurate mental model of how the system makes decisions will progressively lose trust in its autonomy. The universal inclusivity imperative requires universal design principles to be designed at the very beginning: the existing smartphone-centric controls and the complicated voice recognition systems tend to disenfranchise older adults and people with disabilities instead of empowering them [20,21,157]. Installation and administration complexity—even with the Matter standard—is too complex to be used by an average consumer; there is no out-of-the-box experience yet to make the technology easily accessible to mass markets [7,18,19,25,158].

7. Discussion

7.1. Synthesis Across Research Questions

An intelligent home is not determined by a particular technology but by the interaction of sensing, intelligence, and connectivity (RQ1). This convergence allows paradigm shifts between the idea of command parsing and cognitive orchestration, scheduled and predictive maintenance, passive consumption, and active participation as prosumers (RQ2). There is a high gap in maturity between proven areas (energy, HAR, fall detection) and the hypothetical frontiers (LLM orchestration, affective computing, swarm robotics). Challenges (RQ3) are socio-technical issues that demand interdisciplinary solutions. This hybrid form of the review (evidence synthesis plus a prospective roadmap) is premeditated and explicitly under control: each of the claims in Section 4, Section 5 and Section 6 can be directly linked to the studies included in the review, and all the claims in Section 7 and Section 8 are based on the evidence base and are then projected to research directions. This difference is what explains why review papers were considered in the frame of contextual triangulation and not primary evidence (Section 3.5) and why the evidence maturity label (M1–M4) is used everywhere to indicate the level of strength behind every assertion.

7.2. Evidence Maturity Assessment

Evidence maturity labels in Table 4 follow explicit, pre-defined criteria applied by two independent coders (inter-rater agreement κ = 0.78): (M1) Commercial/Validated—technology has been independently replicated across ≥3 peer-reviewed field deployments with real end-users; (M2) Empirically Validated (Lab)—evaluated in controlled or simulated laboratory settings with adequate experimental rigor, but lacking multi-site or longitudinal field evidence; (M3) Proof-of-Concept—working prototype demonstrated, but evaluation limited to a single study, small sample, or developer-controlled environment; (M4) Conceptual/Early-Stage—theoretical framework or preliminary simulation only, with no working system evaluated against real-world data. Where studies reported conflicting maturity assessments for the same technology, the more conservative (lower maturity) label was assigned. Disagreements between coders were resolved through discussion; unresolved cases defaulted to the lower maturity level.

Maturity coding procedure: the unit of analysis was the technology/paradigm cluster (e.g., “mmWave-based HAR”, “LLM home orchestration”) rather than the individual study. For each cluster, the two coders independently reviewed all studies mapped to that cluster during thematic coding, completed a structured maturity coding form recording: (a) the highest evidence type available (field deployment/lab/simulation/conceptual); (b) number of independent replications; (c) whether end-users in real homes were involved; and (d) quality band of the supporting studies. The maturity label was then assigned following the M1–M4 criteria: a cluster could reach M1 only if ≥3 independently published field deployments with real end-users were identified among the included studies; M2 if well-controlled lab evidence was available but field evidence was absent or limited to a single site; M3 if a working prototype had been demonstrated but not rigorously evaluated; and M4 if only theoretical frameworks, architectural proposals, or simulations were present. It is important to note that the M1–M4 criteria are operationalized against specific, verifiable conditions rather than subjective impressions: M1 is assignable only when ≥3 independently published field deployments with real end-users are identifiable among the included studies; M2 requires well-controlled laboratory evidence with adequate experimental rigor but absence of multi-site field evidence; M3 requires a working prototype demonstrated in at least one included study; and M4 applies when only theoretical frameworks, architectural proposals, or simulations are present. These are binary-checkable conditions applied against the extracted evidence, producing a structured classification instrument with documented inter-rater reliability (κ = 0.78) rather than a purely qualitative judgment. A representative sample of studies supporting each maturity assignment is included in Appendix B (flagged with maturity label in the extraction table).

7.3. Comparison with Prior Reviews

In contrast to reviews based on architecture [3,26,27], this synthesis examines the effect of the Matter standard and the semantic gap. This review includes generative AI disruption as opposed to ML surveys [10,13,29]. This work is different in that the subjects are described using a horizontal cross-domain approach compared to vertically siloed reviews [12,20,21,23,30].

7.4. Mapping to Identified Gaps

This section explicitly maps the five research gaps identified in Section 2.5 to the specific sections of this review that address them, demonstrating that the review’s structure was deliberately designed to close each identified gap.

Gap 1: the absence of a post-2020 horizontal synthesis integrating sensing, intelligence, and connectivity advances—is addressed in Section 4, which provides the first systematic evidence-based account of the three converging technological pillars as a co-evolved architecture rather than isolated developments.

Gap 2: the lack of integrated treatment of cross-cutting technologies reshaping multiple domains simultaneously is addressed across Section 4 and Section 5, which examine how federated learning, TinyML, LLMs, and Digital Twins each operate across sensing, energy, health, and security domains rather than within a single vertical silo.

Gap 3: the absence of reviews examining hybrid edge-cloud architectures rather than purely cloud-centric models is addressed specifically in Section 4.2, which characterizes the three-tier cloud-edge-TinyML computing continuum as the defining architectural shift of the post-2020 intelligent home.

Gap 4: the insufficient treatment of socio-technical challenges, including algorithmic bias, explainability, and human-agent collaboration, is addressed in Section 6, which systematically identifies and evidentially grounds five categories of unresolved challenges that extend beyond the classical security and privacy concerns of prior reviews.

Gap 5: the absence of any reassessment of the original Ambient Intelligence vision against the capabilities of modern enabling technologies is addressed in Section 7 and Section 8, which evaluate the degree to which the AmI vision has been realized, identify where the most critical gaps between vision and reality remain, and propose a research roadmap to close them.

8. Future Research Directions (RQ4)

Drawing on the synthesis and identified gaps discussed in the previous section, we now outline key future research directions required to advance the intelligent home paradigm. This section presents a normative research roadmap that is derived from, but deliberately distinct from, the emerging paradigms described in Section 5. Where Section 5 provided an evidence-based descriptive account of what is currently emerging in the literature—classified by evidence maturity (M1–M4) and grounded in the 113 included studies—this section answers a different question: given what the evidence reveals about the current state of development, what specific research investments are most urgently needed to advance the intelligent home vision responsibly, equitably, and at scale? The five directions that follow are therefore not repetitions of Section 5’s findings but prescriptive extensions of them, each directly motivated by the maturity gaps, unresolved challenges, and socio-technical tensions identified in Section 4, Section 5 and Section 6. Readers are encouraged to read these directions in conjunction with the evidence maturity labels in Table 4, which indicate where the research need is most acute.

8.1. Explainable and Transparent AI

A primary research need is the development of natural language explanation systems grounded in causal reasoning and powered by LLMs. For example, a system should be able to explain the following: “I unlocked your front door because I recognized your face, no other people were around the entrance, and you arrived at your usual time of day” [36,37,159,160]. Such explanations should be adaptive and contextualized—varying in depth and technical detail according to the significance of the action, the user’s technical background, and their current context. Additionally, user-controlled model auditing tools should enable occupants to query what data the AI collects and how that data informs its decisions, thereby building justified confidence in the systems installed in their homes.

8.2. Privacy-by-Design and Novel Sensing

The studies in this direction must have a number of complementary objectives. To begin with, privacy-sensitive sensing modalities such as mmWave radar and structural vibration sensing as an alternative to airborne sound capture need to be created and tested in a domestic environment. Second, better TinyML models must apply the principle that the raw data never leaves the device, and thus, anonymity-preserving feature extraction is achieved fully on-sensor. Third, a generation of synthetic data based on existing data should be considered to allow the strong training of models without revealing actual user data. Fourth, differentiating privacy guarantees must be coupled with massively scalable federated learning systems with the ability to cope with the extreme heterogeneity of devices within each household [16,32,33,123,151].

8.3. Lifelong and Compositional Learning

A key research direction is the development of self-learning algorithms that adapt continuously to changes in occupants’ lives. These algorithms must update user preference models without suffering the catastrophic forgetting of previously learned behaviors [6,19,27,161].

Meta-learning—or learning how to learn—offers a promising pathway. Such systems can adopt new user routines on a few-shot basis, using only a small number of examples [6,24,162,163,164]. Complementing this, compositional learning enables systems to recombine previously learned knowledge in new contexts. Together, these approaches provide a new avenue toward generalizable domestic intelligence.

Human–AI model calibration is equally important. Intuitive interactive feedback loops—in which the system proposes a behavioral hypothesis and the user offers granular corrections—enable effective co-training of the intelligent home over time [6,165,166].

8.4. Standardized Security and Ethical Auditing

Security certification for intelligent home architectures should follow standardized audit procedures analogous to appliance safety certification, including benchmark suites and automated penetration testing tools tailored to intelligent home systems [36,103,106]. Bias detection and fairness toolkits should provide customized practical instruments—with diverse and representative datasets—that identify discriminative outcomes and propose targeted remedies. Finally, Ethical Impact Assessments (EIAs) should be established as mandatory formal documents within the product development lifecycle, requiring developers to identify and report all potential privacy intrusions, discriminatory outcomes, and social harms before a product is introduced to the market.

8.5. Human–Agent Teaming

Future intelligent home systems should support mixed-initiative interaction—neither fully automated nor fully compliant with user commands. When a conflict arises, the agent should enter a transparent negotiation process with the user. Achieving this requires advances in dialogue systems and multi-objective optimization [6,74,78,94,98]. Effective proactive assistance depends on accurate inference of the user state. Multi-modal analysis—combining behavioral, vocal, and physiological signals—can be used to predict cognitive load, emotional state, and user intent [83,84,167,168].

Physical assistance represents the longer-term frontier. Mobile robotic manipulators operating in shared human environments must support dynamic space navigation and safe physical human–robot interaction (pHRI). They should also be capable of learning complex multi-step tasks from human demonstrations [169,170]. Evidence rating: human-agent teaming is rated M3–M4 across the included corpus. Mixed-initiative interaction and intent modeling have been demonstrated at the prototype level (M3) in controlled laboratory settings. Embodied AI and safe pHRI in real domestic environments remain at the conceptual or simulation stage (M4). Claims regarding long-horizon teaming capabilities represent proposed research directions rather than documented deployments.

9. Limitations

While the proposed research directions highlight promising pathways forward, it is important to acknowledge the limitations of the present review, which are discussed in this section. This review was conducted with rigorous adherence to PRISMA 2020 guidelines; nonetheless, several limitations must be acknowledged. These are organized below under two categories: study-level limitations arising from the scope and design decisions of this review, and threats to the validity of the review process itself.

9.1. Study-Level Limitations

▪

The English-language restriction may exclude relevant non-English research, introducing language bias. Although English dominates the indexed engineering and computer science literature, important contributions from non-English-speaking research communities—particularly in Asia and Europe—may be underrepresented.

▪

The exclusion of gray literature, patents, and industry white papers means that commercially deployed innovations may be underrepresented, particularly for rapidly evolving technologies such as Matter and LLM-based orchestration, where industry practice often leads academic publication.

▪

The exclusion of purely hardware-focused or protocol-only studies without an AI/ML component (n = 89, 23% of excluded full-text records) represents a principled scope boundary that nonetheless limits coverage of sensing infrastructure and connectivity hardware developments that are foundational to intelligent homes but whose contributions are independent of AI integration. Readers interested in the hardware substrate—including antenna design, radar hardware characterization, and low-power MCU architectures—are directed to dedicated hardware-focused reviews in the cited literature.

▪

Additional databases, such as Google Scholar, used as a primary source, could have identified further relevant studies not indexed in the four databases searched.

▪

The modified JBI quality assessment was applied pragmatically across heterogeneous study designs. Although the adaptation is documented, it introduces a degree of subjective judgment that purpose-built single-design checklists would avoid.

▪

Thematic synthesis inherently involves subjective judgment in theme development. This was mitigated through multi-reviewer independent coding, a structured codebook, and three iterative codebook refinement cycles, but residual subjectivity cannot be fully eliminated.

▪

Significant advances may have occurred between the search cutoff date of 12 June 2025 and the date of publication. The LLM orchestration and Matter/Thread fields in particular are evolving at a pace that may render some findings partially outdated by the time of publication.

▪

The breadth of horizontal coverage across five technological dimensions necessarily limits the depth of treatment within any single domain. Readers seeking in-depth coverage of a specific technology are directed to the domain-specific reviews cited in Table 2.

▪

As a PRISMA-compliant systematic literature review, this work is, by methodological design, a synthesis and appraisal instrument rather than a theory-building one. The development of formal taxonomies, predictive models, or deductive theoretical frameworks derived from this evidence base constitutes a natural direction for future primary research. The M1–M4 evidence maturity instrument introduced here is offered as an analytical foundation to support such work.

9.2. Threats to Review Validity

The following threats to the validity of the review process were identified and, where possible, mitigated:

• Selection bias: The tripartite search query may over-represent synthesis-oriented work relative to primary experimental studies. This was mitigated by supplementary backward and forward citation searching, manual venue searching, and the sensitivity analysis reported in Section 3.1, which confirmed that omitting the scope pillar did not meaningfully increase the yield of primary experimental studies.
• Database coverage bias: IEEE Xplore and ACM Digital Library dominate engineering literature. Socio-technical and HCI-oriented studies may therefore be underrepresented relative to their actual volume in the field.
• Keyword drift: The rapidly evolving terminology of this field—including emerging terms such as ‘agentic AI’ and ‘matter interoperability’—may not have been fully captured by search terms fixed at query design time. This is acknowledged as an inherent limitation of any time-bounded systematic review in a fast-moving domain.
• Publication bias: Positive results and novel architectures are more likely to be published than null results, deployment failures, and usability problems. The latter are therefore likely underrepresented in the corpus, which may lead to an overly optimistic picture of technological readiness.
• Rapid field evolution: The intelligent home field is advancing faster than the typical review-to-publication cycle. Readers should interpret the findings of this review as a structured snapshot of the 2020–2025 literature rather than a definitive account of the current state of technology.

10. Conclusions

This systematic review involved the synthesis of 113 studies (2020–2025) to give the complete picture of the intelligent home at an important evolution point. Three converging pillars that support the change of connected to intelligent domestic environments include multi-modal privacy-preserving sensing, hierarchical cloud-edge-TinyML intelligence engine, and the standardized unified Matter/Thread connectivity standard. These pillars are making possible the new paradigms of LLM-based orchestration, hyper-personalization, simulating the Digital Twin, proactive health monitoring, prosumer energy management, and multi-agent robotics. The home ceases to be a reactive tool and becomes a collaborative partner.

But the evidence maturity evaluation indicates a diverse range of commercially available solutions to conceptual models. To achieve the AmI vision, it is necessary to fix the privacy–security–trust trilemma, bias in algorithms, unstructured problems, and human–agent collaboration calibration. The suggested roadmap, which includes explainable AI, privacy-by-design, lifelong learning, ethical auditing, and human-agent teaming, serves as a guided way to go. Finally, the effectiveness will be quantified through the fair, clear, and compassionate assistance of intelligent homes to a wide variety of residents, and it will take multidisciplinary cooperation, which will last over time.

Appendix D.1. IEEE Xplore Digital Library

Search fields: All Metadata

(“smart home” OR “intelligent home” OR “home automation” OR “ambient intelligence” OR “connected home” OR “domestic IoT”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “Internet of Things” OR “edge comput*” OR “TinyML” OR “federated learning”) AND (“trend*” OR “framework*” OR “architect*” OR “review” OR “survey” OR “paradigm*”)

Filters: Year: 2020–2025; Content Type: Conferences, Journals. Results: 987 records.

Appendix D.2. ACM Digital Library

Search fields: Abstract, Title, Keywords

[[Abstract: “smart home”] OR [Abstract: “intelligent home”] OR [Abstract: “home automation”] OR [Abstract: “ambient intelligence”]] AND [[Abstract: “artificial intelligence”] OR [Abstract: “machine learning”] OR [Abstract: “deep learning”] OR [Abstract: “internet of things”] OR [Abstract: “edge computing”]] AND [[Abstract: trend*] OR [Abstract: framework*] OR [Abstract: review]]

Filters: Published: 01/2020–06/2025; Publication Type: Research Article, Review Article, Conference Paper. Results: 643 records.

Appendix D.3. Scopus

Search fields: TITLE-ABS-KEY

TITLE-ABS-KEY((“smart home” OR “intelligent home” OR “home automation” OR “ambient intelligence”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “Internet of Things” OR “edge comput*” OR “TinyML”) AND (“trend*” OR “framework*” OR “architect*” OR “review” OR “survey”)) AND PUBYEAR > 2019 AND PUBYEAR < 2026

Filters: Document Type: Article, Conference Paper, Review; Language: English. Results: 1124 records.

Appendix D.4. Web of Science Core Collection

Search fields: Topic (TS)

TS=(“smart home” OR “intelligent home” OR “home automation” OR “ambient intelligence”) AND TS=(“artificial intelligence” OR “machine learning” OR “deep learning” OR “Internet of Things” OR “edge comput*”) AND TS=(“trend*” OR “framework*” OR “architect*” OR “review” OR “survey”)

Filters: Timespan: 2020–2025; Document Types: Article, Proceedings Paper, Review Article; Languages: English. Results: 668 records.

Table A5. Summary of database search yields and record counts after deduplication. Records were retrieved from four primary electronic databases (IEEE Xplore, ACM Digital Library, Scopus, and Web of Science) and supplemented by manual searches of reference lists and gray literature. Deduplication was performed using reference management software, resulting in 2918 unique records carried forward to the title and abstract screening stage. All searches were conducted between 9 and 12 June 2025.

Database	Records Retrieved	After Deduplication
IEEE Xplore	987	987
ACM Digital Library	643	548
Scopus	1124	849
Web of Science	668	534
Supplementary Searches	28	28
Total	3450	2918 (unique)

Table A5. Summary of database search yields and record counts after deduplication. Records were retrieved from four primary electronic databases (IEEE Xplore, ACM Digital Library, Scopus, and Web of Science) and supplemented by manual searches of reference lists and gray literature. Deduplication was performed using reference management software, resulting in 2918 unique records carried forward to the title and abstract screening stage. All searches were conducted between 9 and 12 June 2025.

Database	Records Retrieved	After Deduplication
IEEE Xplore	987	987
ACM Digital Library	643	548
Scopus	1124	849
Web of Science	668	534
Supplementary Searches	28	28
Total	3450	2918 (unique)