From real data to a flyable world.
Elevation models, map data, rock, plants, water, a city, traffic, collision and a flight model. Each step with its numbers, its sources and what is still open.
- Bugaboos
- 24.4 km square at 1:1
- Terrain grid
- 2 m, 16-bit
- Physics
- fixed 500 Hz
- Target
- 40 fps at maximum quality
Contents
Every map starts as a table of heights.
Public airborne LiDAR terrain models are the ground truth. We resample them to a 2 m grid in a metric projection and export 16-bit heightmaps together with the exact vertical scale. One unit in the sim is one centimetre and nothing is exaggerated.
| Map | Heights from | Notes |
|---|---|---|
| Bugaboos | LidarBC 1 m DEM (2015 flight, five tiles, about 1.4 GB); NRCan MRDEM-30 for gaps and the horizon | UTM zone 11N, 24.4 km square, horizon rings out to 80 km. The DEM shows the glacier state of 2015 without snow; the snow is modelled on top. |
| Dolomites | Trentino LiDAR 0.5 m (16 % of the core area, CC-BY 4.0), South Tyrol LiDAR 2.5 m (73 %, CC0), Veneto LiDAR 5 m (11 %); Copernicus GLO-30 only for horizon and background | ETRS89 / UTM 32N. Overlapping sources are registered against each other first (Veneto: shifted 0.49 m E, 0.98 m N, 0.10 m up, 0.8 m residual). |
| Lisbon | DGT LiDAR terrain model 0.5 m / 2 m (flown 2024, ±10 cm); the surface model gives building heights | ETRS89 / PT-TM06 (EPSG:3763), the native frame of the DGT data. |
The pipeline (tools/terrain, Python and GDAL)
- Download and mosaic the tiles.
- Reproject into a metric CRS.
- Cut and resample to 2 m (the Bugaboos map is 12,193 × 12,193 samples).
- Check overlaps between sources; keep the measured layer untouched and label everything else as derived.
- Export a 16-bit heightmap and compute the exact Z scale for 1:1 in centimetres.
- Derive slope, height and aspect masks for the materials.
Heights say where the ground is. OpenStreetMap says what stands on it.
OpenStreetMap supplies land cover, water, roads, paths and building footprints. They are never taken as given: streams are snapped to the terrain, buildings are checked against LiDAR heights, and every class is combined with the terrain before it becomes a material or a placement.
- Dolomites, Sella tile: 2,602 OSM buildings, 1,920 of them rural.
- Lisbon: 53,000 buildings in the 10 km box, 3,006 in the high-resolution core; only 35 % carry a height tag, so heights come from LiDAR.
- Roads become the lane graph for traffic; water lines become stream beds; OSM data is ODbL and credited in the app.

Where the LiDAR is soft, we add detail and say so.
The South Tyrol 2.5 m data holds almost no relief below roughly 20 m on slopes: at 5, 12 and 25 m wavelengths it carries 3 %, 6 % and 30 % of the power of the 0.5 m Trentino data. The detail is not blurred, it is missing. Sharpening cannot bring it back (our first attempt doubled the error, 5.6 m to 11 m, and was dropped).
Instead the missing spectrum is synthesised per slope class, learned from the Trentino 0.5 m data, and applied only on open rock and scree between 28° and 42° at strength 0.6. Meadows, forest and tracks stay exactly as measured. On a hidden 1 km strip the slope median matches (22.7° vs 22.7° real, 99th percentile 78.2° vs 79.9°).

The Bugaboos map uses the same idea with learned detail, a summit correction and a physical erosion pass (stream power plus rockfall and talus slopes).
A heightmap cannot stand a wall upright.
Cliffs and overhangs need geometry. Walls steeper than about 55° are generated as Nanite rock meshes whose parameters are measured on the real 0.5 m walls: spacing of strata bands, joint directions and gully density. The closer you can fly, the more triangles the wall gets.
| Tier | Where | Cost |
|---|---|---|
| A, hot spots | Langkofel north face, Sella north walls: everything you fly close past | about 4.8 M triangles per km² |
| B, rock | all other walls above about 50°, same generator, coarser | 5 to 10 times fewer triangles |
| C, ground | landscape at 2 m, masks and instanced plants | about 2–3 MB per km² for the landscape |

Honest limit: no terrain model can measure an overhang. Overhangs and roofs in the mesh are plausible, not measured; we do not use photogrammetry or any 3D tiles.
Snow is modelled on top of a summer elevation model.
The Bugaboos are a fixed high-winter scene: glaciers and slopes are fully covered, crevasses are mostly bridged, and the granite spires hold snow only on ledges and bands. The rule "steeper than about 50° is rock" comes straight from the elevation model. The sun stands low, about 20° above the horizon at noon at 50.7° N.
- Glacier outlines come from the BC Freshwater Atlas; the rasterised mask matches the polygons with an overlap (IoU) of 0.98.
- Snow and ice use layered materials (snow with subsurface light and sparkle, glacier ice with blue transmission).
- Dolomites, spring: patches of old snow are calibrated against Sentinel-2 images.

Plants are placed by maps, not by hand.
Where a tree, a shrub, a boulder or a patch of grass may stand is decided by a stack of layers: soft biome weights, terrain (height, slope, aspect), sun exposure, hydrology, and canopy height measured by LiDAR (surface minus terrain). On top come per-species density curves, colonies and clusters, and a point pattern that avoids the look of a grid. Everything is deterministic, so a rebuild gives the same forest.


Grass is a separate system: it is generated on the GPU in tiers and culled by distance. It is the biggest single cost on a meadow, which is why the performance section below has a switch for it.
Needles, twigs, limbs: a spruce built bottom up.
The current trees are fine from far away and simple up close. We are building needle-level conifers with Unreal's Nanite foliage: each limb mesh bakes needles (three-sided pyramids, 1.6 cm, three triangles) into twigs, side twigs and a limb with a bark tube; a tree is a trunk plus 160 to 240 shared limb parts. Far away, the same tree becomes voxels.
- Spruce library: 8.9 M triangles, 235 MB on disk; each tree asset itself is about 80 KB of transforms.
- A 25.7 m tree is 240 parts and about 82 M virtual triangles.
- Status: lab prototype. GPU time and memory are not measured yet, wind is not done, and the crown still reads sparse near the trunk.


Water follows the terrain downhill, always.
OSM stream lines are snapped to the deepest line of the terrain, so they always run downhill. A bed 0.6 m deep is cut into the landscape, limited to 2 m below the natural channel because real streams pass under roads in pipes. Rivers are 6 to 15 m wide, creeks 1 to 3 m. Lakes follow the measured surface, lowered by at most 1.5 m (dams and reservoirs).
The water itself is a ribbon mesh with single-layer water and flow data in its UVs: the slice has 86,000 triangles for 26 streams and about 16 waterfalls. Next: a baked flow map and a looping normal flipbook per reach type, made from a real fluid simulation. A first test reached a median speed of 1.3 m/s at 14 cm depth; it is research, not in the build yet.

Footprints from OSM, heights from LiDAR, details from a generator.
In the Dolomites 175 huts are placed from OSM on levelled foundations, with their own collision. 720 more rural buildings (up to 150 m²) are prepared; 67 of the candidates were skipped as too steep, 14 failed a validation, and a few sit on roads, lakes or streams.
Lisbon's city is generated: footprint and height from OSM and LiDAR, then windows, balconies and shutters from a kit (about 14,900 kit instances in the fill set alone). 118 LiDAR-only buildings that OSM does not have were found by comparing the height model with the footprints.
The 25 de Abril bridge, built from its real numbers.
The bridge is a procedural model driven by one parameter set and fitted to the OSM outline (its supports match within 0.2 to 4.5 m). Road heights follow the DGT LiDAR profile: 71.4 m at the north anchorage, 75.4 m on the bents, 73.8 m at the south anchorage.
- Span sequence
- 2 × 99.64 – 483.42 – 1,012.88 – 483.42 – 98.64 m
- Between anchorages
- 2,277.64 m
- Tower height
- 190.47 m above the water
- Clearance
- 70 m for ships
- Deck
- 24.5 m wide, 10.7 m truss depth, six road lanes above, two rail tracks below
- Main cables
- 2 × 58.6 cm, 11,248 wires of 4.87 mm; second cable pair added in the 1999 retrofit

Ground audit
A first review found "places that are just the satellite photo with nothing on it". So the ground is audited: in the high-resolution core, every metre that has no content (building, road surface, tree, lawn, furniture, car, container, water) within 6 m counts as empty. Fills are added until the number drops. Empty ground went from 74.1 ha (16.5 %) to 14.7 ha and then to 8.4 ha (1.9 %): 108 generated buildings (1.44 ha, no overlaps, at least 1.66 m apart), and on plazas 315 benches, 315 bins, 224 lamps and 893 bollards.


Cars follow lanes, trains keep a timetable, all of it repeatable.
Traffic is its own small simulation, not an engine crowd system. It runs on a fixed 1/30 s step with a seed per lane, so a lane's cars do not depend on any other lane and a recording replays identically. Cars use the Intelligent Driver Model (acceleration 1.6 m/s², braking 2.2 m/s², headway 1.5 s, minimum gap 2 m) on lanes taken from the OSM roads; a new car enters with a random gap of 45 to 80 m.
- Five simple car models (278 to 322 triangles) with per-instance paint; a four-car train (26.4 m cars) every 120 s from alternating sides, 22 m/s on the bridge.
- Rendering: full models within 150 m, a 12-triangle box impostor beyond. The whole simulation costs about 30 µs per step with 1,700 cars.
- Collision: 64 invisible boxes follow the nearest vehicles within 90 m. Impact speed is measured relative to the vehicle, so a car at 17.1 m/s hits a hovering drone like a 17 m/s crash.
People and aircraft on real routes, with elevation profiles.
Ski tourers, helicopters, bush planes and eagles follow polylines. Routes are least-cost paths on an 8 m terrain model (Dijkstra): slope costs grow up to 30°, 30 to 35° is a boot-pack or skinning exception, above 35° is impassable, glacier and open snow are cheaper. They are proposals, not surveyed tracks. The system itself is deterministic (1/30 s step, seeded).
| Tour from the hut | Length | Climb | Top |
|---|---|---|---|
| Bugaboo–Snowpatch col | 1.78 km | 521 m | 2,748 m |
| Pigeon–Howser col | 3.52 km | 814 m | 2,962 m |
| Vowell Glacier plateau | 4.47 km | 819 m | 2,789 m |
| Anniversary Peak | 3.94 km | 1,054 m | 2,937 m |
- Six ski tours, five heli pick-up spots on flat glacier (slope under 8°, no cliff within 100 m) and two loops of 25.8 km and 36.0 km, two bush plane routes at 55 m/s, circles for five eagles and two ravens.
- A moving agent can crash the drone: a helicopter at 30 m/s into a hovering drone counts as a 30 m/s impact.
- Status: proven in a lab scene; the models await review and nothing is in the public beta yet.


Everything solid has to stop the drone.
After a flight through Lisbon, cars, guard rails and more turned out to have no collision. So collision is audited like data: for up to 20 random instances per mesh group, a 15 cm sphere (the drone's own collision shape) is swept at the object from 3 m away, from eight directions at six heights. A group passes only if the query that the drone itself uses is blocked.
| Group | Before | After |
|---|---|---|
| Bollards (13.8 k) | 1 / 20 | 20 / 20 |
| Benches | 17 / 40 | 40 / 40 |
| Scrub bushes (28 k) | 0 % | 100 % |
| Trees and palms in the core | 1,457 / 3,088 (background trees 0 %) | 1,671 / 1,872, the rest is grass |
| Guard rails, railings, parapets | not detected | 20 / 20 (10,047 boxes) |
| Bridge hanger ropes (6 cm) | none | 20 / 20 (1,224 boxes) |
Cost of all fixes: 5,948 components and 108,718 instances, +213 MB in the editor. Moving traffic passes 4 of 4 crash tests.
Own integration, fixed 500 Hz step, deterministic.
Chaos physics is not used for flying: it is too imprecise for FPV and depends on the frame rate. The flight model integrates a six-degree-of-freedom body itself at a fixed 500 Hz; the engine is used only for collision sweeps. A recorded stick log replays identically.
| Part | Value | Basis |
|---|---|---|
| Mass, thrust | 0.377 kg, thrust-to-weight 3.3 (12.2 N), hover at about 30 % | DJI spec; reviewer and DVR data |
| Drag | 0.00733 / 0.00827 / 0.0117 N/(m/s)² (× air density ratio) | fitted to a DVR top speed of 32.9 m/s |
| Top speed | 31.8 m/s physical; governor 27 m/s (FCC) or 19 m/s (EU) | DJI; EU dynamics estimated |
| Manual rates | 220 / 450 / 0.56 (roll), 180 / 450 / 0.56 (pitch), 60 / 320 / 0.56 (yaw) | factory values read from the DJI app |
| Air | ISA density from altitude and temperature; thrust ∝ (ρ/ρ₀)^0.87 | physics; exponent from analysis |
| Ground effect, propwash | up to × 1.35 close to the ground; up to 15 % thrust loss in propwash | literature; estimated |
| Wind | 5 m/s from 240° by default, 30 % gusts, a lee field in the Bugaboos | estimated |
| Battery | voltage sags under load; thermal model (20 °C start, 600 s time constant) | estimated |
The last column matters: we mark what is DJI data, what is measured and what is estimated. The least certain parts are the motor and rate time constants, braking in manual mode, climb rate and voltage sag; they will be checked against real Avata 2 flights.
Every content step is measured before it is kept.
The target is 40 fps (25 ms per frame) at maximum quality on an M4 Max. Each content step is flown on a script through fixed camera views; if it costs frames it is reworked or cut. Measured on our test machine with the packaged build 0.2.2 at 1080p:
| Map | Frame rate | Memory |
|---|---|---|
| Bugaboos | 66 fps | 13.7 GB |
| Lisbon (beta) | 50 fps | 10.4 GB |
| Dolomites | 33 fps | 17.8 GB |
The Dolomites are below the target and the breakdown shows why. GPU milliseconds saved when a feature is switched off (Dolomites slice, 1440p internal):
| Feature | GPU ms saved | Where it hurts |
|---|---|---|
| Volumetric clouds | 5 – 11 | valleys and forest from afar |
| Landscape grass | up to 12.6 | mountain pass, meadows |
| Trees from afar | 7 – 10.5 | forest from 300 m |
| Lumen global illumination | 3 – 8 | inside forest |
| Shrubs | 9 | aerial views only |
Cliffs, scree blocks, water and fog are nearly free. Ideas in the queue: clouds as a cached sky (−5 to −11 ms) and grass that follows altitude (−5 to −12 ms).
What can be a number is a script.
- Flight model: automated tests for top speeds (8, 16, 27 m/s), climb and sink (6 m/s) and rotation rate (667°/s), run on every change.
- Determinism: traffic and ambient life have headless self-tests that print seeds, schedules, bounds, level of detail and cost.
- Collision: the audit above, plus crash tests that place a car, train, helicopter or plane in front of the drone.
- Performance: scripted flights through fixed camera views at 1080p, run on a quiet machine because run-to-run noise is about ±1 fps.
- Packaged build: the game is tested as the beta pilots get it, not only in the editor. The last run: 10 of 10 tests green.
How the world looks is judged by eye, with pictures side by side. How it behaves is judged by numbers. If you find something that feels off, tell us.