BIM2RDT: Digital Twin Conversion

Updated 12 July 2026

BIM2RDT is a conversion framework that transforms static Building Information Models (often IFC-based) into dynamic, semantically rich digital representations.
It integrates diverse methodologies such as unified network graphs, rule-based decision systems, robot-ready site digital twins, and scan-to-BIM pipelines to enhance computability and interoperability.
The approach streamlines verbose BIM data into structured formats like relational tables, knowledge graphs, or operational digital twins, enabling efficient reasoning, real-time monitoring, and archival preservation.

BIM2RDT denotes a class of transformations in which Building Information Modeling data are converted into downstream representations that are more directly usable for reasoning, learning, real-time operation, or preservation. In the cited literature, the acronym is explicit as “Building Information Models to Robot-Ready Site Digital Twins,” but closely related work also uses the term conceptually for pipelines such as “BIM/IFC → Unified Multi-dimensional Network → Relational/Topological/Design-feature Tables or Data Structures,” BIM-to-rule/decision representations, BIM-to-real-time data platforms, and BIM-to-metaverse or archival representations (Akhavian et al., 25 Sep 2025, Han et al., 8 May 2025, Arias et al., 2022, Brazauskas et al., 2021, Ibara et al., 7 Aug 2025, Borghoff et al., 2024). This suggests that BIM2RDT is best understood as an umbrella concept for converting static, heterogeneous BIM into structured representations that preserve semantics while improving computability, interoperability, or operational relevance.

1. Terminology, scope, and recurring interpretations

A recurrent source of confusion is that BIM2RDT is not used uniformly. In some work it is the name of a specific framework; in others it is an interpretive label applied to a transformation that the paper implements under a different name. Across these usages, the common denominator is a conversion from BIM, typically IFC-based, into a representation that is easier to query, align with reality, stream in real time, or preserve.

Interpretation	Core transformation	Representative source
Robot-ready site digital twin	Static BIM + IoT + robot visual-spatial data → dynamic, safety-first DT	(Akhavian et al., 25 Sep 2025)
Unified network representation	BIM/IFC → multi-dimensional graph → relational/topological/design-feature tables	(Han et al., 8 May 2025)
Rule/decision representation	BIM model → geometric and semantic facts → CLP/CASP program → compliance queries	(Arias et al., 2022)
Real-time building platform	BIM/BMS/sensor data → event-driven spatio-temporal platform	(Brazauskas et al., 2021)
Metaverse runtime	IFC → glTF → Unity → Cluster	(Ibara et al., 7 Aug 2025)
Requirements representation	BTS text → structured concepts, properties, and relations linked to BIM	(Nahri et al., 19 Aug 2025)
Archival representation information	BIM file → AIP with representation information and context information	(Borghoff et al., 2024)

Another common misconception is to treat BIM2RDT as a purely geometric conversion. Several of the cited works explicitly reject that reduction. The target representation may be graph-structured, logic-executable, sensor-linked, safety-aware, or archival, and the retained information commonly includes semantic, spatial, topological, contextual, and temporal layers rather than geometry alone.

2. IFC-to-structured relational and topological representations

One influential formulation treats BIM2RDT as a transformation from IFC into a unified network-based representation that embeds “semantic-spatial-topological” design features (Han et al., 8 May 2025). The motivation is that IFC is hierarchical, verbose, and non-intuitive for automated analysis, while tensor or feature-vector approaches tend to ignore relative position and connectivity, and existing graph approaches remain task-specific.

In that formulation, each BIM component becomes a node, and relationships become typed edges. The graph is written as $G=(V,E)$ , where nodes correspond to components such as IfcWallStandardCase, IfcBeam, IfcPipeSegment, IfcDoor, IfcWindow, floors, fittings, or equipment, and edges encode host, connection, touch-floor, or local spatial relations. Semantic node features are extracted from basic information, geometry, materials, and IfcPropertySet. Edge features encode topological type or, for spatial edges, a richer descriptor including spatial relationship category, angle between positioning lines, shortest distance vector endpoints, signed shortest distance, and angle with the horizontal plane.

A central extension of IFC usage is the addition or simplification of relations not directly convenient in standard IFC traversal. Doors and windows are directly associated with walls by bypassing IfcOpeningElement; IFCRELCONNECTSPATHELEMENTS is used to retrieve connection edges; a computed “touch floor” relation is introduced from bounding-box positional relationships; and a local neighborhood within a radius of 0.5 m is used to define spatial edges. Five spatial categories are distinguished: different surface, interfaces non-parallel, interfaces parallel, point-to-line, and point-to-point. The representation is made compact by suppressing a spatial edge whenever the same pair of nodes is already linked by a host, connection, or touch-floor edge.

This graph is explicitly presented as serializable into RDT-like relational schemas. The paper gives a Component table with identifiers, type, FamilyName, geometry representation, material and properties, and a Relationship table with source and target components, relation type, and spatial descriptors such as distance, angle, angle to horizontal, spatial category, and vector endpoints. In this sense, BIM2RDT is not merely graph construction; it is a staged conversion in which IFC becomes a machine-readable intermediate network that can then be flattened into relational tables, knowledge graphs, or other structured reasoning substrates.

3. Robot-ready and safety-first site digital twins

A second, explicit meaning of BIM2RDT is “Building Information Models to Robot-Ready Site Digital Twins,” defined as an agentic AI framework that transforms static BIM into a dynamic digital twin for autonomous field operation under safety constraints (Akhavian et al., 25 Sep 2025). Here BIM serves as the a priori map, while the runtime twin is updated by fusing three streams: geometric and semantic BIM data, IoT activity and safety data, and robot-collected visual-spatial data.

The architecture is organized around several agents. A semantic perception agent uses YOLOE open-vocabulary object detection with LLM-generated prompts derived from BIM families. A geometric registration and mapping agent uses Shi-Tomasi corner detection, depth projection, and ICP-based alignment against BIM geometry. A safety agent processes wearable Hand-Arm Vibration data and maps safety events into IFC entities such as IfcEvent and IfcTask. A planning or path-optimization agent is conceptually linked to prior BIM+MHA*+NLP work and uses the updated twin to recompute safe missions.

The key registration method is Semantic-Gravity ICP (SG-ICP). Its objective augments the standard ICP data term with a gravity bias term so that the solution favors alignment between a model upright vector and measured gravity. The novel step is that the gravity weight is object-adaptive: an LLM infers a family-specific upright bias $\gamma_f \in [0,1]$ from BIM semantics, and the effective weight is written as $\lambda_{\mathrm{new}}=\gamma_f \cdot \gamma_{\mathrm{initial}}$ . This yields what the paper calls a semantic gravity field, in which “Concrete Pillar” or “Steel I-beam” receive a strong upright expectation, while a “Traffic Cone Orange 70cm” is allowed more tilt.

Experiments compare SG-ICP with standard ICP in four scenarios. In scenarios with occluded features, RMSE reductions of 64.3%–88.3% are reported, and the method is described as avoiding convergence to physically implausible orientations. Scenario 4 illustrates an important nuance: standard ICP can obtain better Hausdorff metrics by converging to an upside-down tripod, while SG-ICP preserves a semantically correct upright tripod. The paper therefore frames the gain not simply as lower geometric residuals, but as improved alignment under physical plausibility constraints.

The same framework integrates real-time HAV monitoring in compliance with ISO 5349-1. Frequency-weighted RMS acceleration and daily exposure $A(8)$ are computed from wearable sensors; when the Exposure Action Value of $2.5\, \text{m/s}^2$ is reached, the system instantiates IfcTask, creates warnings, and records the event in the twin. In this formulation, BIM2RDT is explicitly cyber-physical: the target representation is a robot-usable, semantically aligned, and safety-annotated site twin rather than a static model or a purely analytic graph.

4. Rule, requirement, and decision representations

A logic-oriented interpretation treats BIM2RDT as the conversion of BIM into executable rule and decision representations (Arias et al., 2022). The starting point is that IFC “only represents basic building objects and static information of their properties,” while many compliance questions require derived information such as evacuation routes, aggregate window areas, or uncovered beams. The proposed remedy is to translate BIM geometry and semantics into Prolog or s(CASP) facts and then express regulations as rules and constraints over those facts.

The representation uses predicates such as object(Label, Id, LowerPoint, UpperPoint, Extra) and geometric encodings such as box(...), poly_extrude(...), shape_union, shape_intersect, and shape_subtract. CLP(Q/R) handles dense numeric constraints, while s(CASP) adds stable model semantics, constructive negation, and default reasoning for vague concepts. Regulations such as minimum window width or boiler–ventilation compatibility become executable predicates, and conflicting design updates can be reconciled with priority rules through predicates such as valid_data/2 and canceled/2.

The framework supports both compliance checking and geometric inference. In the office benchmark, the first uncovered beam is found in 0.104 seconds, and the full model is processed in approximately 48 seconds, with 511 beams detected as not covered by more than 1300 structural objects. The same formalism is also used to encode ambiguous notions such as “small room,” allowing multiple stable models when evidence is incomplete or discretionary.

A closely related textual branch of BIM2RDT converts unstructured French Building Technical Specification documents into structured requirements linked to BIM concepts (Nahri et al., 19 Aug 2025). The paper formalizes a requirement as

$\mathcal{R}q = (\mathcal{C}, \mathcal{R}, \mathcal{P}),$

where $\mathcal{C}$ is the concept, $\mathcal{P}$ is the set of properties, and $\mathcal{R}$ is the set of relations. The corpus contains 1,505 pages, approximately 652k words, approximately 19,726 sentences, and 61 different authors; 233 raw requirements are manually annotated with Doccano using POBIM, the CEN Product Dictionary, and the Model BIM Dictionary. Transformer-based NER models, specifically CamemBERT and Fr_core_news_lg, achieve F1-scores over 90%, while Random Forest is reported as the best RE model with an F1 score above 80%.

Taken together, these two lines of work show that BIM2RDT can target executable logic or requirement graphs rather than geometry or runtime twins. In both cases, BIM-related information is recast into a representation where compliance, decision support, or verification becomes a query problem over structured facts, constraints, or triples.

5. Reality-data reconstruction and geometric digital twinning

A reverse-direction but closely related strand treats BIM2RDT as maintaining a consistent mapping between BIM and “reality data twin” representations derived from scans or point clouds. In the SRBIM framework, the core engine for such a pipeline is a unified semantic reconstruction architecture that converts 3D point cloud scenes into IFC BIM through three modules: semantic segmentation with PTv2, Mesh Segments Generation and Refinement, and BIM Reconstruction (Cheung et al., 2024).

The input scene is written as $S=(F_g,F_c)$ , where geometric and color features are jointly processed. PTv2 produces semantically enriched point sets, which are partitioned into semantic segments. For each segment, Poisson Surface Reconstruction yields an initial mesh; vertex densities are normalized, and vertices below the quantile threshold $\gamma_f \in [0,1]$ 0 are removed; Laplacian smoothing then produces refined mesh segments. IFC mapping proceeds by direct class-name matching where possible, with IfcBuildingElementProxy as fallback when no direct IFC class is available. The result is an IFC-compliant BIM that preserves project containment, geometry, and approximate surface appearance through average RGB assignment.

A road-oriented variant develops a scan-to-BIM framework for as-built roads that can also be read as scan-to-road-digital-twin geometry (Ding et al., 2024). It works from semantically labeled point cloud data and covers six asset types: Road Surface, Road Side (Slope), Road Lane (Marking), Road Sign, Road Light, and Guardrail. Sectional polygon contours are extracted and stored in a JSON dictionary structure, from which primitive geometric digital twins are created and then exported to IFC. On six real-world road segments with total length 1,200 meters, the reported average distance error is 1.46 centimeters and the processing speed is 6.29 meters per second. The paper also reports a large storage difference between JSON and IFC outputs, with approximately 48.1 MB versus approximately 236.71 MB for the six segments.

These works show that BIM2RDT need not be unidirectional from BIM to downstream abstractions. A plausible implication is that the term also covers bidirectional maintenance between BIM and scan-derived reality models: scans can reconstruct or update BIM, while BIM can serve as the semantic and interoperability layer for the resulting digital twin.

6. Real-time platforms, interoperability, and long-term persistence

A real-time systems interpretation appears in the Adaptive City Platform, which addresses what it explicitly identifies as the BIM→Real-Time Data gap (Brazauskas et al., 2021). The platform combines BIM, BMS, and IoT through an event-driven architecture based on MQTT, Eclipse Vert.x, PostgreSQL, a file-system time-series archive, and a Real-Time Monitor pushing updates over WebSockets. BIM geometry is normalized into a hierarchical “crate” ontology for buildings, floors, and rooms, with sensor metadata linked through parent_crate_id and consistent coordinate systems. The reported average latency is approximately 160 ms from sensor transmission to front-end visualization, and the underlying SmartCambridge architecture had previously handled approximately 1000 sensors sending data every 20 seconds without slowdown. Here the target representation is a spatio-temporal operational model rather than a design file.

A metaverse-oriented runtime pipeline appears in work on multi-format 3D data integration in Cluster (Ibara et al., 7 Aug 2025). For BIM specifically, the implemented path is IFC → glTF → Unity → Cluster, with automated material replacement, world-coordinate UV generation, and interaction assignment from BIM semantics such as door open/close behavior and chair seating behavior. In a smart office digital twin case, BIM geometry is linked to IoT data through Cumulocity API and a relay server; changes in physical sensors are reported as appearing virtually with approximately 0.3 s latency. This use of BIM2RDT foregrounds multi-device access, simultaneous multi-user participation, and collaborative decision-making.

A longer-horizon interpretation appears in archival research on representation information repositories for BIM (Borghoff et al., 2024). Rather than treating RDT as a runtime twin, this work addresses the transformation of BIM into durable technical documentation within the OAIS model. Representation information is defined as all information needed to interpret a bit sequence for a designated community, and the proposed repository organizes that information through 23 BIMcore content elements covering structural and semantic features, tools for file-format handling, and contexts. The AIP is conceptualized as a package containing Content Information, Preservation Description Information, Packaging Information, and Descriptive Information, with Content Information itself including the data object plus its representation information. This version of BIM2RDT emphasizes that long-term usability depends not only on preserving files but also on preserving the syntax, semantics, software dependencies, context information, and significant properties needed to make future interpretation possible.

Across these platform and preservation variants, interoperability is achieved by normalization into structured intermediate layers: JSON or JSON-LD envelopes, IFC-based entities, glTF/Unity assets, crate hierarchies, or OAIS-compliant information packages. The common pattern is the same as in the graph, rule, and robot-oriented literature: BIM2RDT is a conversion discipline for turning rich but static BIM into representations that are actionable in computation, operations, collaboration, or preservation.