Aerial Omniverse Digital Twin (AODT)

• AODT is an aerial digital twin environment that replicates airborne platforms, scene geometry, and radio propagation for system-level validation and mobility simulation.
• It integrates heterogeneous data streams and multi-layer architectural foundations to support precise planning, resource optimization, and real-time control.
• AODT employs custom conversion pipelines and standardized interfaces to reconcile diverse 3D formats and RF parameters, enabling actionable insights across robotics, communications, and AR applications.

Aerial Omniverse Digital Twin (AODT) denotes an aerial digital-twin environment in which airborne platforms, users, scene geometry, radio propagation, mobility, and operator interaction are represented in a synchronized virtual system. In the literature surveyed here, the term appears explicitly in the context of airborne base-station deployment, where AODT is used as the system-level, high-scale validation and mobility digital twin that complements Sionna’s differentiable link-level optimization (Belgiovine et al., 17 Aug 2025). Closely related work on 6G digital twin networks, aerial corridor communications, aerial-ground robotics, and augmented-reality network twins shows the broader technical substrate from which AODT emerges: physically accurate 3D modeling, two-way data exchange, site-specific channel or semantic scene representations, and human-in-the-loop planning and control (Lin et al., 2022).

1. Definition and scope

AODT belongs to the wider class of digital twin networks (DTNs), defined as “a digital replica of the full life cycle of a physical network.” In that DTN formulation, the twin uses data and models to create a physically accurate network simulation platform, maintains up-to-date network status, can predict future network state, includes interfaces to both the physical network and network applications/users, supports two-way communication between physical and virtual systems, and enables real-time interactive mapping and closed-loop decisions (Lin et al., 2022). AODT specializes this logic to airborne or aerially mediated scenarios.

In the explicit AODT usage presently available, the platform is not treated as a monolithic optimizer. Instead, it is assigned a distinct systems role within a multi-digital-twin pipeline: Sionna performs gradient-based optimization, AODT handles validation and mobility simulation, and a Shared Data Layer enables bidirectional communication through standardized exchange of 3D models, ABS configurations, UE trajectories, and simulation results (Belgiovine et al., 17 Aug 2025). This operational partition is central to the current meaning of AODT in the airborne-base-station literature.

Related implementations indicate that the same functional pattern can be instantiated without the term “AODT” being used explicitly. A Unity-based twin for the MorphoGear aerial-ground robot acts as a scene replica, control panel, and path-planning interface (Karaf et al., 23 Apr 2025), while an augmented-reality twin of an optical network fuses a network model, a 3D operating environment, and immersive operator interaction (Chen et al., 2023). This suggests that AODT is best understood as a systems category defined by synchronization, scene fidelity, and closed-loop aerial operation, rather than by a single canonical software stack.

2. Architectural foundations

The most explicit architectural precursor for AODT is the three-layer reference architecture for 6G DTNs: 6G Physical Network Layer, 6G Twin Layer, and 6G Network Application Layer (Lin et al., 2022). Within the twin layer, three domains are distinguished. The data domain collects, stores, serves, and manages data from the physical system; the model domain contains basic models for network elements and topology plus functional models for planning, traffic analysis, fault detection, emulation, and prediction; and the management domain handles lifecycle and security functions including model creation, update, monitoring, authentication, authorization, encryption, and integrity protection (Lin et al., 2022). For aerial systems, this decomposition maps directly onto airborne telemetry ingestion, scene and channel modeling, and operational governance.

Omniverse is positioned in this lineage as a “scalable, multi-GPU, real-time reference platform for building and operating metaverse applications.” In DTN practice it contributes a high-level programming abstraction, large-scale modeling, AI/ML training and inference, APIs, and visualization (Lin et al., 2022). The significance for AODT is not merely rendering. Omniverse is treated as a runtime substrate for real-time interaction, AI-accelerated workflows, GPU-backed simulation, and real-world digital-twin deployment.

The architectural details of recent airborne work make these abstractions concrete. In the multi-DT airborne-base-station framework, the software bridge had to resolve a 3D scene format mismatch between Mitsuba3 in Sionna and OpenUSD in AODT via a custom Blender conversion pipeline, while also adding JSON-based imports for radio-device configurations and synchronizing center frequency $f_c = 3.5$ GHz, antenna patterns, material properties, ray interaction types and limits, and power normalization/SIR computation (Belgiovine et al., 17 Aug 2025). Such details show that AODT interoperability is, at present, an engineering problem of scene conversion, RF-parameter alignment, and cross-platform state consistency.

A second architectural axis concerns interfaces. DTN work distinguishes network-bound, application-bound, and intra/inter-DTN interfaces, with timing classes spanning real-time (~ms), near-real-time (~second), and non-real-time (~minute) (Lin et al., 2022). This interface taxonomy is especially relevant to AODT because airborne applications combine hard timing constraints for control and mobility with slower loops for planning, operator intent, what-if analysis, and inter-twin coordination.

3. Scene construction, sensing, and twin-state representation

AODT-relevant systems do not rely on a single representation of the world. One line of work constructs a semantic scene representation rather than a dense geometric reconstruction. In MorphoNavi, a physical MorphoGear robot uses a single monocular RGB camera to detect objects, estimate their positions, and transmit a structured scene into Unity as a live visualization and planning environment (Karaf et al., 23 Apr 2025). Detected entities include desks, office chairs, suitcases, compressor, office cabinet, boxes, sports mats, and a Unitree Go1 robotic dog. The twin preserves approximate size and meaning even when exact shape fidelity is lost.

The mapping pipeline in that system combines open-vocabulary object detection with zero-shot segmentation and monocular depth estimation. The evaluated detectors are DINO-X, Grounding DINO 1.6 Pro, Grounding DINO 1.5 Pro, OWLv2, and OWL-ViT, with Grounding DINO 1.5 Pro selected as the best-performing model (Karaf et al., 23 Apr 2025). Distance estimation fuses a geometry-based monocular estimate and a learned depth estimate from Depth Anything v2 plus Segment Anything v2. The explicit equation is reported as

$d = f \frac{h_m}{h_{px},$

and the final distance is computed as a weighted average of 80% geometric estimate and 20% depth-based estimate; if object dimensions are unknown, the system uses only the depth estimate (Karaf et al., 23 Apr 2025). This is an AODT-relevant example of a twin whose state is semantic, object-oriented, and bandwidth-conscious rather than point-cloud-centric.

A second line of work constructs a radio-channel-centric twin. In the 5G aerial-corridor framework, the channel twin (CT) is defined as the radio-channel component of a broader digital-twin environment and is treated as the source of physics-consistent link states and high-fidelity CSI (Tarafder et al., 6 Jul 2025). The site is reconstructed from OpenStreetMap into Blender, ITU material profiles are assigned, actual BS coordinates are taken from OpenCelliD, and the scene is ray-traced in NVIDIA Sionna RT to generate a channel tensor

$\mathbf{H} \in \mathbb{C}^{M \times L \times N},$

where $h_{i,j,k}$ is the gain from BS $j$, antenna $k$, to UAV $i$ (Tarafder et al., 6 Jul 2025). In this representation, the twin state is not primarily semantic object placement but site-specific propagation, blockage, multipath, and beam-alignment structure.

A third representation couples infrastructure state and physical operating environment. The augmented-reality network twin models network topology, 3D models of network equipment, and a 3D map of the surrounding lab environment captured with Microsoft HoloLens 2, together with 3D CAD models rendered as virtual holograms (Chen et al., 2023). Although not aerial, it demonstrates that the digital twin may include both cyber topology and environmental spatial context, a design pattern directly transferable to aerial command, maintenance, or corridor operations.

Taken together, these systems show that AODT-compatible scene construction can be semantic, geometric, RF-physical, or mixed. A plausible implication is that an operational AODT should be viewed as a layered representation problem rather than a single-model problem.

4. Planning, optimization, and control support

AODT is not merely descriptive; it is used to support planning and decision-making. In MorphoNavi, the twin feeds an Aerial-Ground A* planner operating on a 3D grid with semantics: red cubes are occupied or impassable, yellow cubes are dangerous nodes adjacent to occupied nodes with increased cost, and blue nodes are free nodes (Karaf et al., 23 Apr 2025). Motion costs differ by layer: the ground layer has no cost because walking is considered safe, the first layer has high cost because takeoff and landing are unsafe, and higher layers have normal cost though flight is less preferred than walking indoors. The resulting path is decomposed into mode-specific sub-missions—ground motion, takeoff, flight, landing, and further ground motion—because different segments require switching ROS nodes and packages; a mission manager performs the switching and can initialize nodes or reboot hardware if needed (Karaf et al., 23 Apr 2025).

In the 5G aerial-corridor setting, the twin supports cross-layer resource optimization rather than robot navigation. The system studies a multi-cell aerial corridor network with $M$ drone receivers and $L$ terrestrial base stations, each BS using a uniform planar array (UPA) and a beamforming codebook

$$\mathcal{C}_l=\{w_{1,l}, w_{2,l}, \cdots, w_{N,l}\},$$

with one-to-one drone–beam–BS association constraints (Tarafder et al., 6 Jul 2025). The DT/CT-driven pipeline is explicitly: real-world site data $d = f \frac{h_m}{h_{px},$0 3D scene reconstruction $d = f \frac{h_m}{h_{px},$1 ray tracing and channel extraction $d = f \frac{h_m}{h_{px},$2 CT-generated CSI $d = f \frac{h_m}{h_{px},$3 beam optimization and association optimization $d = f \frac{h_m}{h_{px},$4 throughput evaluation (Tarafder et al., 6 Jul 2025). The optimization is split into two stages: Stage 1 selects array-level beamforming weights to maximize antenna gain using dual annealing; Stage 2 solves the joint UAV–BS–beam association problem as a linear sum assignment problem using the Hungarian algorithm (Tarafder et al., 6 Jul 2025). The twin is therefore actionable because it supplies site-specific, ray-traced CSI to an optimizer.

In airborne-base-station deployment, the optimization problem is again two-stage but uses a different twin division of labor. Sionna optimizes location, then orientation and transmit power, while AODT performs large-scale validation and mobility-aware assessment (Belgiovine et al., 17 Aug 2025). The location loss is

$d = f \frac{h_m}{h_{px},$5

with terms for coverage, attraction to AOIs, repulsion, and collision avoidance (Belgiovine et al., 17 Aug 2025). Orientation and power are then optimized through a differentiable coverage-map and SIR-map pipeline, with objectives including a max-min fairness loss and AOI-focused SIR optimization (Belgiovine et al., 17 Aug 2025). Importantly, the paper states that AODT is not the optimizer; it is the validator and mobility engine (Belgiovine et al., 17 Aug 2025).

A recurring boundary condition is that not every paper invoking digital twins for UAV-related operation provides control logic. The Internet of Senses overview associated with immersive UAV teleoperation discusses AI, semantic-aware communication, zero-touch networks, privacy/security, and low latency and network requirements, but does not provide a policy engine, reachability analysis, constraint checker, predictive conflict resolver, or autonomous command correction logic for UAV command approval, denial, or correction (Sehad et al., 2023). This distinction matters because it separates architectural motivation from implemented AODT control mechanisms.

5. Operator interfaces, immersive interaction, and resilience workflows

Human interaction is a persistent theme across AODT-related systems, but the interface modalities vary sharply. In the Unity-based aerial-ground twin, there is bidirectional communication between robot and Unity; commands are sent from Unity to the robot via ROS-TCP-Connector v0.7.0, state is sent from robot to Unity for visualization, and Unity loads scene data “on the fly” from JSON in Streaming Assets without recompilation (Karaf et al., 23 Apr 2025). The twin thereby functions simultaneously as a visual replica and a mission planning interface.

A more immersive operator paradigm appears in the augmented-reality network twin. The architecture combines on-site AR headset(s), a remote edge server with NVIDIA RTX A6000 GPU, the physical network, an SDN controller, a network element database / network connectivity graph, and ML models for fault localization and object detection (Chen et al., 2023). The AR headset shows navigation arrows and destination flags, displays detected cards with labels, bounding boxes, and confidence scores, colors cards by alarm status, supports synchronized 3D model manipulation for remote collaboration, and enables voice and video communication (Chen et al., 2023). The system operated over 86 km of single-mode fiber, with webcam images at 5 frames per second, a measured maximum AR bitrate of 330 Mb/s, and < 35 ms total round-trip latency for the identification result (Chen et al., 2023). While terrestrial, this is a direct precedent for immersive, low-latency, multi-user digital-twin operation.

The most expansive interface vision comes from the Internet of Senses framing. That work identifies VR/AR/head-mounted devices, haptic/tactile interfaces, holographic communication, multi-sensory data, and all-sense communication as ingredients of immersive teleoperation, with 6G network substrate and technologies such as joint communication and sensing, visible light sensing, reconfigurable intelligent surfaces (RIS), networked sensing, beyond-mmWave communication, AI, body area networks, wireless energy transfer, and unmanned aerial vehicles as enablers (Sehad et al., 2023). At the same time, it is explicitly a proposal/overview paper and does not define a specific UAV digital-twin architecture, synchronization protocol, control-plane separation, delay-compensation scheme, or field trial (Sehad et al., 2023). In AODT terms, it supplies a conceptual horizon rather than an implemented operator stack.

AODT’s role in resilience is clearest in the multi-DT airborne-base-station study. There, AODT is used to detect real-time coverage problems affecting a critical UE by monitoring received power over time and applying a threshold-based detector parameterized by $d = f \frac{h_m}{h_{px},$6, $d = f \frac{h_m}{h_{px},$7, and $d = f \frac{h_m}{h_{px},$8 (Belgiovine et al., 17 Aug 2025). Once a sustained drop is detected, the relevant trajectory segment is sent back into the optimization loop, and the ABS is assigned a three-phase mobility plan: reaction phase, stationary phase, and return phase (Belgiovine et al., 17 Aug 2025). This is a rare example in which AODT is embedded in a bidirectional resilience workflow rather than serving only as a visualization or validation endpoint.

6. Implementations, empirical evidence, and limitations

The strongest implemented AODT-adjacent robotics result is the MorphoNavi search-and-rescue demonstration. The system uses Unity 2022.3.9 and Python 3.10 at the operator workstation, while the robot stack includes ROS2 Iron, mavros, Ardupilot v4.4.1, Ubuntu Server 22.04, OrangePi 5b, OrangeCube, STM32-based limb controller, and an ELP-USBFHD05H 2MP camera, with VICON used for ground-truth position evaluation (Karaf et al., 23 Apr 2025). In a room of 6 × 10 × 4 m and a planning area of 5 × 8 × 3 m, with obstacles including tables, boxes, chairs, desks, cabinet, and compressor, the robot successfully located a hidden Unitree Go1 robotic dog behind obstacles while the operator monitored the process (Karaf et al., 23 Apr 2025). Reported metrics were ratio of detected objects to total target objects, accuracy of object detection / pose estimation, and calculation time. Grounding DINO 1.5 Pro achieved a 97.4% detection ratio and 7.34 s calculation time, and the overall system reported 97.4% object detection rate, 13.6 cm average position estimation error, and average processing time of about 7.3 s/image (Karaf et al., 23 Apr 2025). The limitations are explicit: occlusion sensitivity, degradation when object shapes or orientations are irregular or unknown, and the fact that the pipeline is not real-time yet (Karaf et al., 23 Apr 2025).

The strongest aerial-communications optimization evidence comes from the site-specific channel-twin framework. At 100 m altitude, the proposed HF-CT-driven method reported throughput gains of 10.9% over LF-CT, 13.7% over the 3GPP statistical CT, 67.2% over random, and 80.7% over closest-BS (Tarafder et al., 6 Jul 2025). Average throughput for the proposed method was 324.66 Mbps at 75 m, 187.35 Mbps at 100 m, and 107.90 Mbps at 130 m, with the performance gap narrowing as altitude increased (Tarafder et al., 6 Jul 2025). Runtime grew with user density, reaching roughly 393 s at 40 UAVs for the HF-CT and LF-CT optimization schemes (Tarafder et al., 6 Jul 2025). This establishes that a channel-centric digital twin can materially improve aerial resource allocation, but also that fidelity and optimization cost are tightly coupled.

The most explicit AODT evaluation appears in the multi-DT airborne-base-station study. AODT validation used AODT version 1.1.1, the Tokyo PLATEAU 3D city map, 50 UEs, 10 ABSs, a 60 s simulation duration, 1 s time steps, 50 UEs moving within each AOI during validation, and 500K rays per ABS (Belgiovine et al., 17 Aug 2025). AODT was configured to match Sionna as closely as possible with $d = f \frac{h_m}{h_{px},$9 GHz, ITU concrete material, TR 38.901 antenna pattern for ABSs, and half-wave dipole for UEs (Belgiovine et al., 17 Aug 2025). The validation showed SIR gains of roughly 3 to 20 dB in AOIs 1, 2, and 4, with slight reductions in AOIs 0 and 3, which already had high SIR (Belgiovine et al., 17 Aug 2025). The same study also documents practical constraints: no native procedural import/export for radio-device placement, the need for custom BS mobility, retention of only the strongest 500 channel taps per antenna pair, and single-AOI-at-a-time validation workarounds (Belgiovine et al., 17 Aug 2025).

A broader Omniverse context is provided by the 6G DTN example HeavyRF, which integrates Omniverse, HeavyDB, real-time RF propagation simulation, and geospatial/customer data, and performs ray-traced RF simulation at 10 cm resolution, far finer than conventional 4G planning tools using approximately 10 m resolution (Lin et al., 2022). Although not aerial, it demonstrates that Omniverse-based DTNs can be built as operational systems for planning, simulation, and visualization at substantial scale.

A persistent misconception is that all UAV or immersive-teleoperation papers titled around digital twins present implemented aerial twin systems. The Internet of Senses paper frequently associated with this theme is, in fact, a visionary framing document: it contributes high-level conceptual motivation only, provides no UAV-specific digital twin architecture, no command validation logic, no math/models/algorithms for UAV control, and no empirical evidence for a realized aerial digital-twin teleoperation system (Sehad et al., 2023). The current literature therefore supports AODT strongly as a systems direction and, in specific cases, as a concrete validation environment, but the maturity of implementations remains uneven across robotics, communications, and immersive-operation use cases.