The concept of multisensor data fusion dates back to the 1970s in the context of robotics and defense. Early work focused on establishing theoretical foundations and lexicons for combining data from multiple sources within permissible time frames. A widely cited definition by Hall and McMullen described data fusion as a ’hierarchical transformation of observed data from multiple sources into a form that enables decision making’ [27]. In practice, many initial fusion systems were deterministic or rule-based aimed at achieving specific fusion goals under hardware constraints of the time.

One pioneering framework was the Joint Directors of Laboratories (JDL) Data Fusion Model [3]. Developed in the military community in the 1980s, the JDL model defined a taxonomy of fusion across levels 0 to 4: from raw data alignment, to object refinement (state estimation), situation assessment, impact assessment, and process refinement. It emphasized combining sensor observations to estimate object identity and position, originally for surveillance/tracking applications. Steinberg et al. later revised the model in 1999 to refine these levels and generalize it to broader situations [28]. The layered approach of the JDL framework influenced many subsequent system designs, ensuring that each level of fusion produces outputs at increasing levels of abstraction. Although conceived for military sensing, the concepts of the JDL model are applicable to any multisensor system.

Around the same time, reference [29] proposed a simpler three-level architecture for sensor fusion. The lowest level dealt with raw signal fusion (often requiring training to learn correlations between sensors); the intermediate “evidence” level fused features or evidence using statistical methods (with spatial/temporal alignment as a preprocessing step); the highest “dynamics” level fused information in the context of system dynamics or models. Reference [29] introduced performance indicators such as the quality of fused information and robustness to uncertainties, which foreshadowed later work on fusion confidence and uncertainty estimation.

Another influential early framework was by Luo and Kay [30], who distinguished between multi-sensor integration (using multiple sensors to reach one decision) and multi-sensor fusion. They proposed hierarchical structure with distributed fusion centers, highlighting that fusion could occur at different hierarchy levels of a system. The data collected at the sensor level is integrated at the fusion centers, where the actual fusion is done. After processing all sensors, domain-specific high-level information of interest is obtained. The fusion process is supported by relevant databases and libraries. In the process of fusion, raw signals from individual sensors are abstracted to symbolic information. This idea of performing some fusion locally (sensor node level) and some globally (central level) is reflected in today’s edge vs. cloud fusion split in IoT systems.

Harris et al described another example of hierarchical fusion called the waterfall model [31]. The hierarchical levels are similar in essence as the earlier models with an emphasis on the processing functions of the lower levels. Sensors pre-processing is done at at level 1 while feature extraction and pattern processing in level 2. It is followed by situation assessment and decision making being done at level 3. Conceptually, the processed signal from level 1 are converted to fetaures in level 2 that leads to state description and querying in Level 3 of the model.

Bedworth and Brien [32] described a hybrid framework called the Omnibus model. This process model was inspired by conceptual OODA (Observe, Orient, Decide and Act) cycle called Boyd loop and the waterfall model. Various tasks in data fusion and its functional objectives are realized in different modules. Three levels of data fusion, that is, data, feature and decision level have been defined. Separate modules implement various level-wise tasks and meet their functional objectives.

Several foundational frameworks emerged during the 1970s–2000s that established the theoretical and architectural basis for multisensor data fusion. Table 4 summarizes these early models, highlighting their central ideas and lasting contributions. The JDL model remains one of the most influential, while subsequent frameworks such as the three-level architecture, waterfall model, and Omnibus model introduced alternative perspectives emphasizing hierarchy, distributed processing, and hybrid design. Collectively, these models shaped the evolution of modern sensor fusion approaches.

A common feature of all models discussed till now is the hierarchical transformation of data. This higher-level integration of locally processed sensor data at intermediate or final nodes can be suitably applied to modern autonomous systems also. Interestingly, although these models were proposed more than a decade apart, they embody the same fundamental principle of hierarchical fusion, where both the level of cognizance about the system and the granularity of information progressively increase across successive layers. To systematically map the actual fusion models deployed in autonomous systems, we first developed a unified architecture that aligns the levels proposed by different frameworks and examined the extent of consensus among them. The rationale for this integration lies in the observation that, despite differences in terminology and chronology, all major fusion frameworks embody a common principle of hierarchical refinement: data is progressively transformed from raw sensor measurements into higher-level situational understanding and decision support. By aligning these frameworks, we expose the underlying consensus of established fusion frameworks. The unified model draws upon the Joint Directors of Laboratories (JDL) data fusion model, which has been briefly introduced earlier, it is elaborated here to provide the rationale for harmonizing different models. The JDL levels can be summarized as follows: 1.

Level 0: Sub-Object Data Assessment (Source Preprocessing)—Deals with raw sensor data (signals, features, pixels, etc.), encompassing tasks such as noise filtering, feature extraction, registration, and alignment. Example: Cleaning raw radar or camera data and synchronizing sensing rates before applying detection algorithms.

By aligning the JDL model with other popular frameworks, an integrated multisensor fusion model has been shown in Fig. 1.

Traditional models established the architectural blueprints and terminology for sensor fusion. These can also be related to modern autonomous systems, as raw data from multiple sensors of autonomous systems must undergo several processing stages to become actionable information. In the next sections, we will discuss the historical trajectory of sensor fusion over the past four decades, examining how each decade introduced new sensors, fusion strategies, and application domains that collectively shaped the current state-of-the-art.

The 1990s marked a decisive transition in sensor fusion research, moving from conceptual discussions to increasingly concrete algorithmic implementations. Researchers explored a broad range of strategies to address uncertainty, adaptability, and system integration, with varying degrees of success. Key categories included probabilistic inference methods, neural network–based approaches, rule-based and evidence-theoretic reasoning, feature-level integration, and modular hybrid architectures. Collectively, these approaches laid important foundations, though they were frequently constrained by computational power and often tailored to narrow application contexts.

Probabilistic Fusion Methods: Probabilistic inference emerged as a rigorous framework for handling uncertainty. Cox et al. (1992) applied Bayesian inference to stereo vision, showing that probabilistic depth estimation could outperform deterministic triangulation by integrating evidence from stereo pairs [33]. Larkin (1998) used Bayesian networks to classify acoustic signals, explicitly capturing dependencies among features [34]. Shahbazian et al. (1993) introduced multi-hypothesis tracking, allowing radar and sonar to jointly maintain competing target hypotheses [35]. While these approaches formalized uncertainty propagation, their computational cost scaled poorly with the number of sensors, preventing real-time deployment in dynamic environments.

Neural Network-Based Methods: The growing availability of computational resources encouraged early use of neural networks for adaptive fusion. Perlovsky and McManus (1991) presented a maximum-likelihood neural network that adaptively classified sensor inputs, blending statistical estimation with learning [36]. Davis and Stentz (1995) demonstrated neural networks for autonomous outdoor navigation, where fused vision and range inputs were mapped to obstacle recognition in unstructured terrains [37]. Similarly, Cao and Hall (1998) applied neural networks to autonomous guided vehicles (AGVs) for vision-based navigation [38]. These methods demonstrated adaptability and the ability to capture nonlinear inter-sensor dependencies, but were shallow by modern standards, trained on limited data, and lacked interpretability, restricting their robustness in diverse conditions.

Rule-Based and Evidence-Theoretic Approaches: Rule-driven frameworks also played an important role. McKee (1993) proposed a taxonomy of “what can be fused” for vision systems, providing systematic guidance for constructing integration pipelines [40]. Blasch and Hong (1999) implemented Dempster–Shafer evidence theory in a “belief filtering” mechanism for target tracking [43]. This enabled reasoning under partial or conflicting evidence without requiring strict prior probabilities. Rule-based systems were transparent and interpretable, but generalization was limited, and belief combination rules were difficult to tune when ambiguity was high.

Feature-Level Fusion: Several works moved beyond raw data integration to focus on fusing intermediate representations. Tang and Lee (1992) proposed geometric feature relation graphs to preserve spatial consistency across multi-camera vision sensors [39]. Peli et al. (1999) introduced unified feature-level fusion before classification, improving recognition accuracy in multisensor vision systems [44]. These approaches demonstrated the utility of fusing more compact representations rather than raw data, reducing computational demands. However, they relied on precise calibration and were sensitive to occlusion, noise, and sensor failures.

Application-Specific Architectures: The decade also produced domain-tailored modular architectures. Kam, Zhu, and Kalata (1997) developed one of the earliest multi-sensor frameworks for mobile robots, integrating sonar, vision, and dead-reckoning [42]. Alag, Goebel, and Agogino (1995) proposed a supervisory fusion framework for Intelligent Vehicle Highway Systems (IVHS), emphasizing fault detection, redundancy, and supervisory control [41]. Beyond robotics, Mandenius et al. (1997) applied fusion in industrial bioprocessing, combining chemical and process sensors for real-time monitoring [45]. These architectures highlighted the feasibility of embedding fusion into control pipelines, but remained tightly coupled to specific sensor suites, limiting scalability and cross-domain applicability.

Overall, the 1990s advanced sensor fusion by formalizing uncertainty modeling, exploring adaptive neural approaches, and embedding fusion into practical systems. As summarized in Table 5, most methods remained constrained by high computational demands, narrow scope, and lack of real-time generalizability. Yet, they established enduring design principles—probabilistic rigor, adaptive learning, interpretable reasoning, and modular integration—that continue to influence sensor fusion research today.

By the end of the decade, sensor fusion had matured from theoretical constructs to operational prototypes in vision, robotics, defense, and industrial applications. However, most systems were specialized, computationally demanding, and lacked generalizable frameworks. The subsequent decade (2000–2010) witnessed increasing convergence and the emergence of machine learning as a unifying tool for sensor fusion across domains.

Building upon the foundational probabilistic, neural, and rule-based strategies of the 1990s, sensor fusion research in the 2000s advanced toward greater methodological rigor and broader applicability. Several developments characterized this decade: the extension of probabilistic filters for nonlinear and non-Gaussian systems, the integration of Bayesian and evidence-theoretic reasoning, the rise of distributed consensus schemes for multi-agent settings, and the early adoption of machine learning techniques to learn fusion mappings from data rather than relying solely on handcrafted rules. Collectively, these advances established many of the algorithmic templates that would later be scaled and generalized in the deep learning era.

Probabilistic Filtering and Bayesian Extensions: Kalman filter variants dominated this period, particularly in navigation and tracking tasks. The Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) were widely adopted for fusing inertial measurement units (IMUs) with GPS data, enabling more reliable state estimation under nonlinear dynamics [46]. Coué et al. (2002) applied Bayesian programming to automotive state estimation, demonstrating how Bayes filters could flexibly combine sonar, odometry, and other modalities under uncertainty [47]. Particle filters emerged as an important alternative, addressing limitations of Kalman-based methods by accommodating non-Gaussian noise and multimodal posterior distributions. These probabilistic methods significantly improved robustness in early autonomous robots and vehicles, although they remained computationally demanding for high-dimensional state spaces.

Rule-Based Evolution and Evidence-Theoretic Integration: Rule-based fusion was refined into more mathematically grounded frameworks. Koks and Challa (2003) proposed combining Bayesian methods with Dempster–Shafer (D–S) evidence theory, providing a hybrid reasoning scheme capable of fusing probabilistic estimates with uncertain or incomplete evidence [48]. This integration allowed richer representations of belief states but incurred high computational cost as sensor sets scaled. Distributed consensus algorithms also gained prominence in this era. Xiao et al. (2005) introduced a consensus-based scheme that enabled sensor networks or multi-robot systems to achieve agreement on global state estimates despite each node holding only partial information [49]. These approaches were particularly important for wireless sensor networks (WSNs), where centralized fusion was often infeasible.

Machine Learning for Fusion Mappings: A key innovation was the move toward learning fusion rules directly from data. Faceli et al. (2004) proposed a hybrid intelligent framework combining neural networks, fuzzy inference, and decision trees, allowing the system to adaptively determine fusion weights and mappings [50]. While deep learning was not yet viable, simpler models such as multilayer perceptrons and fuzzy systems demonstrated the feasibility of training adaptive fusion models. These early machine learning–driven systems reduced reliance on handcrafted rules, though their learning capacity was limited by data availability and computational constraints.

Decision-Level and Classifier Fusion: Decision-level fusion became increasingly popular for classification tasks. Instead of integrating raw signals, systems combined outputs of independent classifiers trained on individual sensor modalities. For instance, in wearable human activity recognition (HAR), classifiers based on accelerometers and gyroscopes could be fused via voting or weighted averaging to yield more robust predictions. This ensemble approach improved resilience to sensor failures and noise. While Chavez-Garcia and Aycard (2015) [51] formally studied multisensor decision fusion slightly after 2010, their work synthesized principles already established in the late 2000s, particularly in intelligent vehicle perception.

Application-Specific Advances: The decade also saw growing application diversity. Choi et al. (2011) applied hierarchical fusion of RFID and odometry for indoor robot localization, building on techniques developed in the late 2000s [52]. Lu and Michaels (2009) fused ultrasonic sensor data for structural health monitoring under varying conditions, addressing robustness challenges in safety-critical applications [53]. In agriculture, Huang et al. (2007) integrated multiple sensing modalities for precision farming, reflecting the growing role of sensor fusion in environmental and industrial domains [54]. Each domain imposed distinct requirements—low power consumption for wearable devices, high accuracy for aircraft navigation, or resilience to noise and environmental variation for outdoor robotics—driving tailored fusion solutions.

The representative works summarized in Table 6 illustrate how this decade broadened the methodological toolkit. Unlike the 1990s, where most systems were rigidly rule-based or narrowly probabilistic, the 2000s emphasized flexibility through probabilistic generalizations, distributed consensus, and adaptive machine learning. While these advances greatly expanded the scope of sensor fusion, limitations remained, particularly in computational scalability, dependence on expert tuning, and restricted ability to automatically learn complex feature hierarchies. These constraints would soon motivate the adoption of deep learning approaches in the following decade.

In summary, the early 2000s represented an era of methodological consolidation and gradual transition from handcrafted fusion rules toward data-driven adaptability. Probabilistic frameworks were extended to handle nonlinearities and non-Gaussian noise, distributed consensus schemes emerged for networked systems, and hybrid AI methods showcased the potential of learned fusion. While the computational and data limitations of the period constrained progress, this decade equipped the field with versatile building blocks—Kalman filter variants, Bayesian/evidence hybrids, consensus protocols, and ensemble learning—that directly informed the deep learning–driven breakthroughs of the 2010s. These advances thus represent the logical evolution of the 1990s prototypes into more flexible, scalable, and domain-diverse fusion frameworks.

Building upon the probabilistic, rule-based, and early machine learning approaches of the 2000s, the period from 2011 to 2020 marked a decisive transformation in sensor fusion. This shift was driven by three converging factors: the availability of large-scale multimodal datasets, rapid advances in deep learning, and growing deployment of autonomous systems in safety-critical contexts. Fusion models moved from handcrafted pipelines and shallow learners toward end-to-end trainable architectures capable of learning cross-modal representations directly from data. Research during this decade spanned autonomous vehicles, UAVs, precision agriculture, infrastructure monitoring, and wearable human activity recognition, demonstrating both methodological diversity and domain-specific innovation.

Deep Learning–Based Fusion Architectures: One of the most transformative advances was the adoption of deep neural networks for multi-sensor fusion. In autonomous driving, vision and LiDAR fusion evolved from late fusion of independent detections to early and mid-level feature fusion within deep networks. Approaches such as PointNet++ and multimodal convolutional fusion architectures enabled learned feature representations across modalities, significantly improving detection and localization accuracy [60,61]. Unlike handcrafted pipelines, these models could discover optimal cross-modal mappings, albeit at the cost of requiring large annotated datasets and high computational resources.

Distributed and Cooperative Fusion: Another major development was the emergence of cooperative and distributed fusion frameworks, especially for connected autonomous vehicles and IoT-driven systems. Cooperative perception (V2X) allowed vehicles to exchange sensor data, extending situational awareness beyond line-of-sight occlusions. Liu et al. (2023) [62] reviewed this paradigm, which was conceptually established in the late 2010s through simulation-based studies. These works emphasized the need for synchronization, low-latency communication, and consensus protocols, anticipating real-world multi-agent fusion systems.

Domain Diversification and Application-Specific Fusion: Fusion research extended into healthcare, agriculture, and infrastructure. In wearable HAR, Banos et al. (2012) combined accelerometer, gyroscope, and contextual sensors to mitigate noise sensitivity and improve recognition reliability [63]. In precision agriculture, Maimaitijiang et al. (2020) fused UAV imagery, satellite data, and ground-based sensors using machine learning for crop monitoring, enabling multiscale environmental insights [64]. Infrastructure monitoring adopted sensor fusion of accelerometers, strain gauges, and vibration sensors to detect anomalies in bridges and civil structures. These application-specific systems demonstrated that the core fusion principles—robust uncertainty handling, redundancy, and adaptive learning—could generalize across domains.

Reliability, Explainability, and Adversarial Concerns: By the late 2010s, researchers recognized that fusion systems in safety-critical domains required not only empirical accuracy but also transparency and robustness. Explainable AI (XAI) techniques were explored to interpret multimodal fusion decisions, particularly in healthcare and autonomous driving. Simultaneously, adversarial studies revealed vulnerabilities, such as perturbations or physical artifacts that could mislead fused perception systems. This highlighted the need for redundancy-driven architectures, formal verification of fusion pipelines, and design of fail-operational strategies for safety-critical deployments.

Representative studies from this decade are summarized in Table 7, which highlights the methodologies, fusion granularity, application domains, and technical contributions.

To complement these representative studies, Table 8 presents a focused technical comparison of widely used classical approaches—Extended Kalman Filters (EKF), Particle Filters (PF), Dempster–Shafer (D–S) theory, and Bayesian inference—under challenging conditions such as non-Gaussian noise and conflicting sensor evidence. This highlights the continued relevance of classical filters alongside modern learning-based approaches.

In summary, the 2011–2020 decade marked the transition from handcrafted, model-driven fusion toward data-driven and learned fusion paradigms. Deep learning architectures enabled joint feature representations across heterogeneous modalities, cooperative fusion expanded to multi-agent systems, and application domains diversified beyond vehicles and robots to healthcare, agriculture, and infrastructure. Despite these advances, classical methods such as Kalman filtering and Bayesian inference remained essential, particularly in constrained environments or where formal guarantees were required. The coexistence of classical and AI-driven approaches underscores the versatility of sensor fusion, while ongoing challenges in scalability, robustness, and verifiability continue to motivate research in the current decade.

Extending the deep learning–driven breakthroughs of the 2010s, sensor fusion research from 2021 onward has accelerated toward tackling real-world deployment challenges. Models are no longer expected to perform well only in controlled datasets but must generalize across environments, sensor suites, and tasks with minimal reconfiguration. This decade has also been marked by the emergence of transformer-based architectures, context-aware dynamic fusion, and practical demonstrations of cooperative perception in multi-agent systems. Fusion has become increasingly pervasive, appearing in domains as varied as autonomous firefighting robots, UAV-based wildlife monitoring, intelligent transportation infrastructure, and healthcare wearables.

Generalizability and Cross-Domain Adaptation: A central focus of this era is improving the robustness and scalability of fusion models. Systems trained on one platform (e.g., a specific vehicle type or city) are increasingly adapted to new conditions with minimal retraining, using transfer learning, domain adaptation, and synthetic-to-real approaches. High-fidelity simulators are leveraged to generate rare or safety-critical scenarios, with adaptation methods ensuring real-world applicability. Physics-informed neural networks emerged as a hybrid approach, embedding sensor physics into learning pipelines to reduce data requirements and enforce physical consistency.

Transformer-Based and Attention Mechanisms: Transformers and attention-based architectures became central to fusion pipelines. Chitta et al. (2022) proposed TransFuser, a transformer-based model that jointly encodes LiDAR and camera streams for autonomous driving [70]. These architectures enable multi-task and cross-modal learning, allowing a single network to perform detection, segmentation, and tracking simultaneously. However, they remain computationally heavy and require large training datasets. HydraFusion (Malawade et al., 2022) extended this by incorporating attention-driven context selection, dynamically weighting sensors depending on environmental conditions [71]. Such adaptive mechanisms improve resilience but increase training complexity.

Edge–Cloud Hybrid Fusion Architectures: The push toward deployment in connected and resource-constrained environments led to hybrid strategies. Edge devices handle low-latency, safety-critical decisions (e.g., obstacle avoidance), while cloud or roadside servers manage computationally intensive tasks such as global route planning or cooperative perception. This split addresses both responsiveness and scalability, though it introduces latency-management and bandwidth-allocation challenges.

Self-Calibration and Fault Tolerance: Autonomous systems now integrate self-diagnostic routines to detect and respond to sensor degradation (e.g., blocked LiDARs, degraded cameras). Multi-sensor redundancy allows systems to isolate and exclude faulty sensors or re-calibrate them dynamically. Tommingas et al. (2025) demonstrated fusion of UWB and GNSS with ML-based uncertainty modeling, highlighting the need for adaptable frameworks capable of self-healing in diverse environments [72].

In the domain of robust navigation, a GNSS/IMU/VO fusion framework with multipath inflation factor has been proposed to explicitly mitigate the challenges of urban multipath interference. By leveraging real-time IMU and VO inputs, the system dynamically adjusts GNSS weighting and adaptively updates VO velocity variance within a robust extended Kalman filter. Field tests in dense urban areas demonstrated 63.4% and 56.1% improvements in horizontal and 3D positioning accuracy, respectively, over conventional fusion schemes [73]. This work highlights the importance of incorporating environment-aware weighting models for next-generation positioning, navigation, and timing (PNT) systems. Beyond terrestrial navigation, recent work has demonstrated the value of multi-sensor association for high-precision space target localization. By fusing visible light and infrared detection with laser ranging under a Gaussian mixture TPHD framework, this approach achieves great accuracy, outperforming binary star angular-only methods [74]. This highlights how sensor fusion enables unprecedented precision in space situational awareness and orbital tracking.

Diversified Applications: Healthcare, smart cities, and environmental monitoring benefited significantly. Rashid et al. (2023) developed SELF-CARE, a wearable fusion framework for stress detection, combining multiple biosignals with context identification [75]. Hasanujjaman et al. (2023) fused autonomous vehicle and CCTV camera data for smart traffic management [76]. In addition, advances in embedded ultra-precision sensing have expanded the scope of sensor fusion to metrology and industrial domains. A recent study introduced a fiber microprobe interference-based displacement measurement system capable of measuring ranges up to 700 mm with subnanometer accuracy. Unlike conventional interferometers, this approach enables compact, embedded measurements in confined spaces, supporting applications in high-end equipment manufacturing and biomedical robotics [77]. Aguilar-Lazcano et al. (2023) surveyed sensor fusion in wildlife monitoring, highlighting challenges of sparse data and field deployment [78]. These illustrate how the principles of redundancy, adaptability, and interpretability are increasingly tailored to domain-specific constraints. In intelligent transportation and scene understanding, multi-modal remote perception learning frameworks have been introduced to integrate object detection with contextual scene semantics. For example, a Deep Fused Network (DFN) combines multi-object detection and semantic analysis, yielding improvement on SUN-RGB-D and on NYU-Dv2 compared to existing approaches [79]. These results underline the growing role of context-aware multimodal fusion for complex environments in autonomous driving and robotics. Industrial monitoring and predictive maintenance are also benefiting from self-supervised representation learning. A recently proposed multihead attention self-supervised (MAS) model learns robust features from multidimensional industrial sensor data using contrastive augmentation strategies. Applied to a real-world water circulation system, MAS improved anomaly detection performance without reliance on large labeled datasets [80]. Such approaches demonstrate the promise of representation learning in industrial sensor fusion for fault detection and equipment health monitoring.

Representative works from this period are summarized in Table 9, capturing the methodologies, AI/ML integration, fusion granularity, application domains, and limitations.

Key Trends and Challenges: The works in Table 9 reflect several defining directions. Transformer-based models and attention mechanisms (e.g., TransFuser, HydraFusion) dominate high-performance fusion pipelines but remain resource-intensive. Context-aware frameworks demonstrate adaptability but raise challenges in calibration and scalability. Application diversification is notable—ranging from autonomous driving to stress detection and wildlife monitoring—yet many domains suffer from data scarcity and lack of standardized benchmarks. Cooperative perception moved from conceptual discussions to initial real-world demonstrations, though interoperability and evaluation metrics remain unresolved.

Another important theme is hybridization: combining learning-based adaptability with model-driven rigor. Physics-informed neural networks, domain adaptation, and simulation-based training address limitations of purely data-driven methods. Similarly, hybrid edge–cloud fusion architectures balance real-time responsiveness with global situational analysis, though at the cost of latency management and secure communication. Finally, fault tolerance and self-calibration have become indispensable, marking a shift toward self-healing, resilient fusion pipelines capable of long-term deployment.

In summary, the 2021–2025 period marks the consolidation of deep learning and transformer-based architectures, the practical emergence of cooperative fusion, and the diversification of sensor fusion into new domains. The emphasis has shifted from achieving accuracy in benchmark datasets to ensuring robustness, scalability, and adaptability in highly dynamic real-world conditions. These trends set the stage for future research on verifiable, resource-efficient, and generalizable sensor fusion frameworks.

To address this limitation, explainable AI (XAI) has become central to sensor fusion research. By exposing how models approximate higher-level reasoning, XAI can bridge the trust gap between opaque neural fusion and stakeholder accountability. In autonomous driving, trustworthy deployment hinges on transparent fusion pipelines with interpretable decision-making at multiple abstraction levels [11]. Similarly, in medical contexts, opaque multi-modal fusion undermines clinical reliability; interpretable frameworks are increasingly recognized as prerequisites for adoption [12]. The absence of hierarchical, explainable fusion is therefore both a technical and socio-ethical barrier. Recent surveys [10] underscore that progress remains incremental, and much work is needed before interpretable and certifiable fusion frameworks can be reliably deployed in safety-critical environments.

Several research priorities emerge for bridging this gap:

Unified evaluation frameworks and context-aware benchmarks are needed to standardize interpretability metrics in autonomous domains.

Beyond deep learning, neuromorphic sensor fusion offers a promising path toward energy-efficient and inherently interpretable models. Ceolini et al. (2020) introduced one of the first multimodal neuromorphic benchmarks, integrating event-based vision (DVS) with electromyography (EMG) signals [85]. Using delta modulation, continuous EMG signals were converted into spike trains compatible with spiking neural networks (SNNs), while DVS provided native event-driven input. Fusion was achieved via late concatenation in the penultimate layer, followed by retraining of the output classifier across neuromorphic hardware platforms (Intel Loihi, ODIN+MorphIC). The released dataset comprised 15,750 samples from 21 subjects performing five static hand gestures, making it a pioneering benchmark for multimodal neuromorphic fusion. Results showed accuracy comparable to GPU baselines, while achieving energy-delay product (EDP) improvements of up to 600× on MorphIC and 30–50× on Loihi. Although inference latency increased modestly, the efficiency gains highlight the feasibility of neuromorphic fusion for embedded, low-power autonomous platforms.

This study opens several technical directions for neuromorphic fusion. First, encoding fidelity remains an open challenge: spike conversion from continuous signals risks discarding fine-grained information, motivating adaptive or learned encoding schemes co-designed with SNN architectures. Second, hardware–algorithm co-design is critical: current neuromorphic platforms face constraints such as limited neuron counts, fixed precision, and inefficient crossbar operations. Progress will require sparsity-aware SNN topologies and novel hardware primitives capable of handling dense multimodal streams. Third, standardized benchmarks are urgently needed. While Ceolini’s dataset is valuable, broader benchmarks reflecting dynamic driving, UAV navigation, or healthcare monitoring are necessary for systematic evaluation across modalities and platforms.

Explainability and Security in Neuromorphic Fusion: Neuromorphic systems also offer opportunities for explainability and robustness. The temporal and event-driven nature of SNNs makes causal reasoning more tractable, as spike timing and event sequences can be directly linked to decision outcomes. Developing XAI tailored for neuromorphic pipelines could deliver transparent reasoning for safety-critical systems such as AV perception or prosthetic control. Security is equally pressing: while neuromorphic fusion may resist conventional adversarial perturbations, it introduces new vulnerabilities such as spoofed event streams, requiring adversary-aware design and validation.

Generalization Across Domains: A key long-term challenge is extending neuromorphic fusion beyond static benchmarks to dynamic, heterogeneous domains. Late-fusion architectures demonstrated for DVS+EMG can be generalized to LiDAR, radar, inertial, and acoustic signals, supporting low-latency, always-on fusion in energy-constrained platforms. Hybrid pipelines—where neuromorphic encoders perform low-level, energy-efficient fusion before passing to deep learning or symbolic reasoning modules—could combine efficiency with semantic richness. Such hybridization points toward a future in which neuromorphic front-ends complement AI-driven back-ends, delivering scalable, interpretable, and trustworthy sensor fusion for autonomous systems.

In summary, future research must simultaneously advance the explainability of classical deep learning fusion systems and explore emerging paradigms such as neuromorphic computing. Together, these trajectories aim to reconcile the empirical success of modern AI with the interpretability, efficiency, and trustworthiness demanded by safety-critical autonomous deployments.