From Open-World Games to Geospatial Data, City Models and Digital Twins

Original illustration of a fictional coastal city changing from a rendered scene into a color-coded 3D city model.
Concept illustration of a fictional coastal city, generated for Homann Software.

One crossing, two sets of questions

Imagine driving through a virtual coastal city. The road dips under a bridge, towers slide past the windscreen, and their shadows cross the carriageway. In a planner’s 3D view, the same scene could look remarkably similar. The planner needs to know how tall a particular building is, which road actually runs over the bridge, and who measured the geometry and when.

Rockstar and Take-Two confirmed that Grand Theft Auto VI launched on 19 November 2026 for PlayStation 5 and Xbox Series X|S. Rockstar sets the story in Vice City and across the fictional state of Leonida. Its official setting guide also features the Leonida Keys and Grassrivers. Release status checked on 20 November 2026. Rockstar launch notice, Rockstar game page, Only in Leonida, Take-Two launch release.

Vice City gives us a familiar way into the design problem. The engineering examples below come from published documentation for other systems. The bridge, pipeline and lab are our own geospatial examples.

A game city is designed to feel coherent in motion. The player expects routes to be traversable, collisions to make sense and a busy street to react to events. A 3D GIS supports questions about spatial entities that can be measured and traced back to evidence. A game can have rich object semantics, navigation and simulation, and a GIS can be immersive and interactive. The difference lies in what each system has to substantiate.

What 3D GIS adds to the scene

A geographic information system (GIS) captures, manages, analyses and displays information linked to a location. A building record can carry a stable identifier, a geometry, a use code, links to its parts and the survey or dataset that supplied it. A defensible height result needs a definition, such as ridge above local ground or roof elevation in a given vertical datum, as well as the source, observation date and uncertainty. A processing date answers a different question from the date on which the building was observed. PostGIS spatial data model, OGC CityGML conceptual model.

  • A 2D map can show building footprints and road centre lines in plan. It can store heights and network relationships as attributes.
  • A 2.5D elevation surface assigns one height to each horizontal position in that surface. A bare-earth terrain grid is a typical example. At the crossing, the road and bridge require separate surfaces or another representation to capture both heights.
  • A 3D representation can hold stacked surfaces, overhangs or distinct solids at the same horizontal location. A closed building shell and two separated surfaces under a bridge encode different information even though both are three-dimensional.

A terrain model describes the ground. A digital surface model can also include the tops of buildings and vegetation. A building model represents individual structures and may distinguish wall and roof surfaces. A textured mesh records a set of surfaces with imagery. One mesh can encode several heights at a location, but the triangles may span an entire block with no building boundaries or IDs. A sharp photograph can paint a doorway on a simplified wall. Conversely, a plain survey-derived surface can have a well-characterised position. Reviewing positional error, missing features and observation dates requires evidence about the data and how it was made. USGS elevation terminology, CityGML 3.0 Users Guide.

A plan overlay, elevation and disconnected road graph for the same grade-separated crossing.
Comparison model. The bridge and lower road overlap in plan. Their surfaces have a vertical gap and their graph edges have no shared node at the crossing.

How large virtual worlds fit into memory

Consider what a viewer needs while approaching a block. Spatial partitioning divides the world into cells or tiles so the client can fetch a small part of it. A spatial index answers a different question, such as which building footprints intersect a corridor. An index can shortlist objects using their bounding boxes, then an exact predicate can test the candidates. A renderer may use a hierarchy of tile bounds to decide which branches to visit. The query store and the renderer can each choose an index that suits their workload. PostGIS spatial indexes, Cesium’s 3D Tiles specification repository.

Streaming brings relevant assets in and evicts them within a memory budget. A networked viewer also has to budget bandwidth and request latency. Prioritising tiles near the camera or along its movement can help, and a coarser parent tile can keep the view usable while finer content loads. Frustum culling skips objects outside the camera’s viewing volume. Occlusion can skip objects blocked by nearer surfaces when the renderer has enough evidence. Offline, geometry simplification reduces triangles and may combine many objects into a coarser proxy. Epic’s World Partition and hierarchical level of detail documentation gives one example using cells and proxy meshes. Epic World Partition, Epic hierarchical level of detail.

Rendering, collision, navigation and analysis have different tolerances. The façade a player sees can include each window recess. The collision surface can be a flat wall. A navigation mesh can encode walkable space for a specific agent radius and step height. A city inventory can preserve the building as an identified object even when the renderer merges its roof with neighbouring structures in a distant tile. Each representation can serve its task as long as the conversion and omissions are understood. Epic navigation system, OGC 3D Tiles.

Coordinate precision becomes visible at scale. A 32-bit float with a value near six million metres has increments of roughly half a metre. Centimetre-scale motion can be lost when arithmetic is performed directly on those values. An application can preserve Earth-fixed positions at higher precision, then express local geometry relative to a nearby origin for rendering. A game can also use local origins in a wholly fictional world. Conversions need a clear definition of where the origin is, which way each axis points, and what a unit means. Cesium georeferencing example.

Why crossing lines can describe two different routes

In our example, the bridge runs east–west and the road under it runs north–south. A point in the middle is inside both footprints. In elevation, the bridge underside is 6 m above the modeled road surface. In the transport graph, the two edges have different endpoint IDs. There is no junction at the overlap.

Snapping every pair of crossing lines into a shared node would create a turn that no vehicle could make. A height check can help detect a bad join, but topology explicitly records connectivity. Ramps can connect two elevations some distance away. Two routes at the same height can still be separated by a barrier or a turn restriction. A network needs connectivity and the constraints for its intended users. ASAM OpenDRIVE road linkage.

A walkable area for a game character answers questions about that character’s movement rules. A traffic model needs directed links, permitted turns, vehicle restrictions and often demand or signal data. An accessibility assessment may need gradients, surfaces and kerb information. A rescue route can depend on verified width, clearance, access permissions and current closures. The appropriate question decides which geometry, attributes and topology must survive the pipeline.

Where CityGML, 3D Tiles, glTF and GeoJSON fit

CityGML 3.0 defines a conceptual model for urban objects, their geometries and relationships. GML/XML is one standard encoding. CityGML 3.0 has LoD0 through LoD3. It dropped the former LoD4 convention for interiors and allows applicable indoor and outdoor features to be represented at different LoDs. A building could have a LoD2 outer shell and LoD0 floor-plan geometry for its interior if the dataset includes those features. OGC CityGML standards and encodings, CityGML 3.0 conceptual model.

3D Tiles 1.1 organises large geospatial scenes for streaming and rendering. It defines tile hierarchies and bounds and supports structured metadata, including metadata associated with features in glTF content. A tileset can expose feature IDs and selected properties without carrying every relation or history record from its upstream model. The data producer decides what to include and whether IDs remain stable across tiling runs. OGC 3D Tiles 1.1 listing, 3D Tiles specification repository.

glTF 2.0 is a runtime asset format for meshes, materials, textures, nodes and other scene data. A tile can reference a glTF asset. glTF core gives you the asset and local transforms. The enclosing tileset or application provides the geospatial placement and the mapping back to source objects. Khronos glTF, 3D Tiles specification repository.

GeoJSON is useful for exchanging features such as footprints, points and lines with properties and optional IDs. RFC 7946 uses WGS 84 longitude and latitude in that order, with optional ellipsoidal height in metres. A GeoJSON feature can carry route attributes in its properties, with any graph interpretation defined by the application. Its standard geometry types do not include a volumetric solid. UTM eastings and northings require transformation before they can be written as RFC 7946 positions. IETF RFC 7946.

The word detail answers three different questions. CityGML LoD characterises an object’s geometric representation. Rendering LoD decides which geometry to draw at runtime. In 3D Tiles, a client can project a tile’s geometric error into a screen-space error in pixels and refine when that error exceeds its threshold. Data quality describes how well a dataset represents reality, including accuracy, completeness and freshness. A low pixel error tells you how close an approximation looks on screen. It says little about whether the source roof was measured correctly. Research on building LoDs likewise found meaningful variation in what a level label represents. 3D Tiles geometric error and refinement, Biljecki, Ledoux and Stoter on building LoDs.

Georeferenced scenes in a game engine

Cesium for Unreal is a documented example of streaming georeferenced 3D Tiles into Unreal Engine. It is an independent technology comparison with its own toolchain. The example has no GTA VI integration. A georeference connects a position on the WGS 84 globe to the engine’s local coordinate system. Document both frames, the axis mapping, the handedness and the metre-to-engine-unit scale before comparing their coordinates. A simple check is to place points due east and above a known origin, then confirm their direction and displacement in the target scene. Cesium for Unreal, Adding datasets, Placing objects on the globe.

An ellipsoidal height measures from a reference ellipsoid. German DHHN2016 normal heights use an official vertical reference. For compatible ETRS89 inputs, Germany’s GCG2016 quasigeoid gives the height anomaly to subtract from ellipsoidal height to obtain DHHN2016 normal height. Treating an official normal height as a WGS 84 ellipsoidal value can shift an entire layer vertically. Validate the frame, realisation and units, and compare against known control points in the correct height reference. The lab below uses an ellipsoidal origin and skips conversion to official German heights. BKG height transformation, PROJ topocentric conversion.

For example, import CityGML building objects into a feature store, derive glTF geometry for 3D Tiles, and expose selected footprints as GeoJSON. Keep the source identifier, relevant attributes, provenance and relationship keys together through conversion. When one building becomes several mesh parts, each part can preserve a reference to the same building ID. When a tile combines geometry from multiple features, feature-level IDs can link the selected surface back to its authoritative object. Test that relationship after retiling, especially across tile boundaries and between coarse and fine levels.

A real data candidate near Mainz

For a pilot, start with a few blocks in Mainz’s Altstadt. On 20 November 2026 the state’s bulk file server returned the start of LoD2_32_446_5538_2_RP.gml on request. Its envelope covers a 2 km square from E 446,000 m and N 5,538,000 m. The inspected file uses CityGML 1.0 namespaces and the AdV LoD2 building profile. The accompanying delivery note lists 5,723 buildings, delivery on 23 April 2026 and an indicated data vintage from 11 September 2025 through 11 March 2026. These values describe that delivery. They are distinct from the statewide catalogue’s 1 April 2026 status and 1 km tiling description. Official tile directory, Mainz tile delivery note, Official LoD2 metadata.

The file declares ETRS89 / UTM zone 32N (EPSG:25832) horizontally and DHHN2016 normal heights (EPSG:7837) vertically. LVermGeo reports horizontal accuracy corresponding to the cadastre and typical height accuracy of around 1 m, with potentially larger deviations for some complex roofs. LoD2 gives a standardised roof representation. It gives no assurance of detailed façades or interiors. The sample header and delivery note establish format, reference and delivery metadata. The completeness and accuracy of individual features would need separate checks for a real project. LVermGeo open data, Official LoD2 metadata.

GeoShop’s generated extracts and bulk delivery are distinct channels. A shop maintenance notice can affect one route while a tile is available through bulk download. The product metadata specifies Data Licence Germany – Attribution – Version 2.0. For a redistributed extract, carry the provider’s required attribution and year of data acquisition, the licence and dataset links, and mark edits such as clipping or conversion. The browser lab uses its own synthetic crossing so each height and network relationship can be inspected directly. GeoShop notices, LoD2 delivery page, GovData licence terms.

When the city becomes a digital twin

In this article, a digital twin is a digital representation of an identified physical entity or process that is kept in sync at a stated frequency and fidelity for a defined purpose. A bridge twin might link inspection records and scheduled condition updates to a stable asset ID. To support closures, it might also ingest a timestamped operational feed, disclose its freshness and flag conflicts. Updates can arrive in batches or in near real time, depending on the decision. The operating process needs a data owner, stale-data rules, checks against the physical asset and a way to correct errors. Digital Twin Consortium definition.

Design questionOpen-world game3D GISDigital twin
Primary objectiveCoherent, responsive player experienceInspect and analyse features in a spatial referenceTrack an identified physical system for a defined decision
Identity and stateWorld entities and runtime or saved game stateFeature IDs, source records, dates and relationshipsAsset identity, observations and a governed update cycle
Example acceptance checkTraversal and collision behave as designedCoordinates, topology and quality meet query requirementsUpdates reconcile with observations within defined limits

These are primary objectives. Games, spatial systems and twins can all use meshes, semantic data and simulations. The physical link required of a twin is an additional obligation. A fictional game city can instead evolve according to authored rules and player actions.

Different simulations also need different evidence. Traffic needs a connected network plus behavioural inputs and comparison against observed flows. Shadows need the correct location, time, orientation and relevant obstructions. Energy estimates need construction, occupancy and weather assumptions with calibration against suitable data. Flood analysis requires terrain, treatment of bridges and culverts, rainfall or inflow, boundary conditions and a checked hydraulic model. Animating a water surface supplies a visual effect. A flood-depth prediction also needs modeled water movement and validation against observations or benchmarks. US Army Corps of Engineers HEC-RAS documentation.

A pipeline for decisions as well as display

Source data flows through validation into a versioned feature store, which feeds visualisation and analysis. Validated updates return to the store.
Comparison model. Streamed geometry and analysis both use the versioned feature store. New observations go through validation before they change a release.
  1. Ingest source assets and record provider, dates, reference systems, quality and licence. Keep enough of the source to reproduce the import.
  2. Validate and transform. Check schema, geometry, required attributes and relationship targets. Transform coordinates with named operations and controls. Quarantine invalid records and make omissions visible.
  3. Maintain the authoritative feature store. Track stable IDs, topology, validity periods and versions. Keep changes to public render releases traceable to the source revision.
  4. Build the viewing and analysis paths. Publish lightweight tiles with selected metadata for streaming. Send measurements and routing requests to the versioned analytical geometry and graph. A generalised tile can lose narrow gaps, sub-features, precise boundaries, links and even whole objects. Keep a release ID so a selection in the viewer can resolve against the same data revision in the query service.
  5. Update and reconcile. Validate changes, rebuild affected tiles, and publish a consistent snapshot. If an update fails, keep serving the last coherent release and disclose its age. Feed discovered errors back to the data owner.
Use case What the 3D view adds What makes the result defensible
City planning Compare a proposal’s massing and views from the street Surveyed reference data, scenario versioning and checks for missing obstructions
Infrastructure communication Show neighbours where a bridge deck sits relative to a road closure Approved drawings, current works information and dimensions from the maintained record
Training Let learners rehearse routes and recognise spatial hazards Scenario validation with practitioners and separate checks of field procedures

For related Learning Hub examples, see how a Minecraft extension keeps its placement rules testable and how a robot mission coordinator treats observed state and control as separate concerns.

Try the crossing in your browser

Rotate the model below, select an object and run the spatial query. All four features are synthetic. A nominal WGS 84 origin near Mainz supplies their coordinate frame.

For offline use or to inspect the source, download the 3D GIS lab (ZIP). Extract it and open index.html in a modern browser. Both versions use the browser’s Canvas API and plain JavaScript.

View a static screenshot and sample result
The browser lab displays an amber bridge, a purple lower road, an object inspector and the spatial query result.
Screenshot of the lab after running the query.
  1. Rotate the city and switch to Plan view to see where the amber bridge and violet road overlap.
  2. Click either object or select it from the inspector. Compare its stable ID, class, data date, source, local height range and graph nodes.
  3. Run the query at E 0 m, N 0 m. The returned footprints are BR-01 and RD-02. The modeled vertical gap is 6 m and their graph edges share no node.

The footprint query uses local east and north coordinates in metres. For the adjoining transformation exercise, the lab includes WGS 84 geographic to Earth-fixed conversion and Earth-fixed to local east, north and up conversion, tested against axis controls and the published PROJ example. The fixture geometry is already authored in local metres. A numerical match to the control establishes how the conversion behaves under those assumptions. Survey accuracy, actual clearance and safety require different evidence. PROJ topocentric reference and control.

If someone asks whether they can turn from the lower road onto the bridge, show them the graph and the edges behind it. If they ask whether a vehicle fits underneath, show them the measured surfaces, the vertical reference, the date and the tolerances. A city can look convincing in both cases. The evidence needed to answer each question is different.