yoofiannan/drone-simready-scenes

SimReady drone scenes

Five seeded OpenUSD environments built from NVIDIA SimReady Central assets plus procedural outdoor geometry.

These cover five common commercial drone applications, not a statistically ranked global top five:

Scenario	Environment	Mission data
filming	Backlot, loop road, moving forklift and quadruped subjects	24-second subject animations, orbit/chase/reveal/top-down shot positions
inspection	Industrial plant, tanks, stack, pipe rack, OEM equipment	Inspection viewpoints and synthetic corrosion, cracks, paint damage and hotspot metadata
mapping	Uneven construction site, crane, stockpiles, excavation	Survey grid, oblique orbit, GCPs/checkpoints, camera/GSD assumptions, DEM
agriculture	Crop field, farmyard, refill station, power lines	Scouting/spraying grids, crop health/weed labels, synthetic stress raster, DEM
delivery	Depot, suburban blocks, launch pads, yards and utilities	Parcels, pickup pads, drop zones and cruise-altitude suggestions

Getting this dataset

The dataset is self-contained: generator code, 20 pre-built scenes and the 84 referenced SimReady asset packages (about 3.5 GB). Download it with the Hugging Face CLI, then run the quick start below from the downloaded folder. Skip the fetch step; the assets are already in place.

hf download yoofiannan/drone-simready-scenes --repo-type dataset --local-dir drone-simready-scenes
cd drone-simready-scenes

The asset packages are unmodified mirrors of their upstream Hugging Face datasets, pinned to the commits recorded in assets/index.json. Credits and source links are in ATTRIBUTION.md.

Quick start

Run from this project directory. Install uv and Python 3.12 or newer first.

uv sync --locked
# Already downloaded in this workspace. Needed on a fresh copy:
uv run python -m dronesim fetch

uv run python -m dronesim generate --scenario all --seeds 1 --validate
uv run python -m dronesim validate
uv run python -m unittest discover -s tests -v

The asset catalog currently contains 84 packages, with about 3.74 GB of local assets including download metadata, below the selected 30 GB budget. Downloads require network access. Generation, previews and tests run offline after setup. fetch --dry-run resolves revisions but does not estimate download size. The downloader does not enforce a hard byte cap if you expand the catalog.

Scenes for seeds 0, 1, 2 and 3 are included under scenes/. Open, for example, scenes/inspection/seed_0001/scene.usda.

# Generate a batch, or one scenario with explicit lighting and reduced clutter
uv run python -m dronesim generate --scenario all --count 10 --start 10 --validate
uv run python -m dronesim generate --scenario agriculture --seeds 42 --density 0.5 --lighting overcast --validate

# Omit the placeholder vehicle when integrating your own drone
uv run python -m dronesim generate --scenario delivery --seeds 42 --no-drone --validate

# Regenerate layout maps or inspect available asset pools
uv run python -m dronesim preview
uv run python -m dronesim list

--seeds takes explicit nonnegative integers. --count/--start selects a consecutive batch instead. --scenario accepts multiple names or all. --out PATH changes the output root. Reusing a scenario and seed overwrites that output, so use a separate output root for experiments. --no-preview skips layout images. Use positive, moderate --density values; placement capacity limits clutter counts and crop counts grow with density.

The same seed, options, asset index, dependency versions and output location reproduce the same files. Lighting has an independent random stream, so changing a lighting override does not reshuffle the layout. Relative asset references change when the output location changes; reproducibility is not promised across dependency or asset revisions. Preserve uv.lock and assets/index.json.

Files and coordinates

Each scenes/<scenario>/seed_XXXX/ folder contains:

• scene.usda: composed scene with relative references into the shared assets/ tree.
• mission.json: task descriptions and ground truth.
• layout.png: 2D layout map, not a rendered view of the USD scene.
• dem.npy: float32 terrain elevations for mapping/agriculture.
• stress_map.npy: float32 synthetic stress values for agriculture.

World coordinates are meters, Z up, kilograms, in a local frame, not latitude/longitude or a georeferenced map. Absolute waypoint heights are world Z unless a field explicitly says AGL. Raster arrays use [row_y, column_x], with Y increasing from the origin. DEM samples lie on grid vertices. Stress-map samples start at its origin at the stated spacing.

Common mission.json fields:

Field	Meaning
scenario, seed, density, physics_variant	Generation configuration
bounds	Nominal XY working region
home, drone_spawn	World XYZ positions
drone_proxy_prim	/World/Drone, or null with --no-drone
lighting	Sun/sky preset and sampled parameters
objects	Asset/procedural class, center, size, yaw in degrees and dynamic flag
assets_used	IDs into assets/index.json
warnings	Placement/physics compatibility notes
terrain	DEM file, origin, resolution and array shape, when applicable

Some aggregate procedural objects use <procedural:class> in objects[].path; these are bookkeeping records, not USD prim paths. Asset records have real paths under /World/Objects. Semantics use UsdSemantics.LabelsAPI; older simulators may need a translation to their own labeling API.

Scenario-specific fields:

• Filming: subjects[].trajectory rows use trajectory_fields (t seconds, XYZ center, heading degrees). shots[].camera rows are [t, x, y, z]. Subject transforms are time-sampled in USD at 30 fps. Shots are position recipes and guide curves, not automatically animated camera prims.
• Inspection: inspection_targets[].viewpoints contains position/look_at. defects contains class, target, position, normal, size, severity and RGB visibility. Hotspots are metadata only, not a thermal sensor model. Box-target defects approximate an asset's bounding surface, not its precise mesh surface.
• Mapping: survey contains world-XYZ waypoints, overlap/trigger/GSD assumptions and oblique_orbit. gcps contains local surveyed marker positions. Stockpile volumes and nominal GSD are approximations, not measurement-accuracy guarantees.
• Agriculture: crop contains prototype counts and instancer path. stress_patches, weed_patches and stress_map describe synthetic labels, not calibrated multispectral measurements. Crop instances are visual-only to keep scenes lightweight.
• Delivery: deliveries contains parcel prims, pickup/drop positions, drop radius and cruise height. A package prim or measured mass may be null. Drawn routes are illustrative, not collision-free flight plans.

Opening in a simulator

Copy both scenes/ and assets/, retaining their relative layout. Copying a lone USDA file is not sufficient. If generating elsewhere with --out, regenerate on the destination machine or preserve its relative asset paths.

1. On a supported Linux/Windows GPU machine, open scene.usda in Isaac Sim or another OpenUSD-capable application.
2. Check that the materials and asset references resolve. Procedural materials use UsdPreviewSurface; some library assets use NVIDIA MDL and require an MDL-capable renderer or material conversion.
3. Replace /World/Drone, or generate with --no-drone and spawn your vehicle at mission.json's drone_spawn.
4. Add your flight dynamics, rotor forces, controller/autopilot bridge and sensor configuration. The proxy is a 1.5 kg rigid body with a camera, not a flying vehicle implementation. It falls under gravity without a controller.
5. Translate mission data into your planner/controller. Check full swept flight paths against buildings, cranes, wires, moving subjects and terrain before flying.
6. Test contacts, joints, startup settling and time-sampled animation in the target physics engine.

There is one scene-level gravity/physics scene. Referenced asset PhysicsScenes are deactivated. World-space reset transforms on articulated links are made placement-relative. Visual-only OEM assets receive coarse box colliders. Some original asset physics/joint schemas are engine-specific.

--physics physx, --physics newton and --physics mujoco select recognized variants where the publisher supplies them. They do not convert USD into native MuJoCo/Newton/Gazebo formats or guarantee that all assets support the selected engine. Default generation preserves each asset's published variants.

Isaac Sim smoke test

tools/isaac_smoke_test.py is the first runtime check to do on a GPU machine. It cannot run on a Mac. Copy scenes/ and assets/ together, then run it with the Python inside your Isaac Sim install (or the .venv-isaac from the SimReady submission guide):

<isaac-python> tools/isaac_smoke_test.py scenes/inspection/seed_0001/scene.usda --seconds 20
<isaac-python> tools/isaac_smoke_test.py scenes/delivery/seed_0001/scene.usda --gui   # watch it

It opens the stage, lists unresolved references and composition errors, confirms a single active PhysicsScene, plays physics for the requested time, and flags any dynamic rigid body that fell more than 2 m or moved more than 0.5 m (--fall, --drift). The drone placeholder falls by design and is reported as a warning, not a failure. It then captures smoke_test.png from /World/Drone/Camera and writes smoke_test.json beside the scene. Exit code 0 means clean, 1 means failures, 2 means the stage did not open. It has not been run yet; the script was written against Isaac Sim 4.5–6.x APIs and the screenshot step is the most likely part to need a small adjustment on an older release.

Validation and limits

The static validator checks USD composition/dependencies, stage units, one active PhysicsScene, collider/semantic presence, asset parts escaping their placements, AABB overlaps among referenced objects, approximate ground contact, selected waypoint heights/bounds, viewpoints inside asset bounds, spawn intersections with asset bounds and filming animation samples.

Seeds 0–3 passed these checks for all five scenarios, 20/20 scenes. The integration tests regenerate all five seed-1 scenes in fresh Python processes and compare hashes of USD, JSON, PNG and NumPy files. A separate test checks actual UDIM tile resolution. USD's dependency scan can still print warnings for literal <UDIM> paths even when the numbered tiles exist.

These are baseline environments, not certified collision-free missions or fully validated physics simulations. Static checks do not cover every procedural overlap, every animated frame, visibility/occlusion, swept routes, flight dynamics, texture appearance or engine startup behavior. Bounding-box support checks are approximate. Procedural buildings and vegetation are deliberately simple. No Isaac Sim runtime, GPU rendering, autopilot or hardware-in-the-loop validation was performed on this Mac.

Assets and attribution

Sources: NVIDIA, ForgeCAD, ALLSIDES, Vertiv and deeproboticslab packages indexed by SimReady Central. All 84 downloaded package manifests declare CC BY 4.0. See ATTRIBUTION.md for per-package pinned source links and licenses; preserve source notices and check embedded material/code licenses before redistributing.

The generator references assets rather than modifying their downloaded files. Scene-level adaptations include coordinate/unit normalization, transform overrides, variant selection, collider proxies, physics overrides and semantic labels. Terrain, buildings, trees, crops, markings, hazards, task metadata and the drone proxy are procedural additions.

Project layout

catalog/assets.yaml selects assets and role tags. dronesim/fetch.py downloads packages and records revisions; probe.py measures source units/bounds/physics. usdkit.py authors USD. scenarios/ contains the five builders. generate.py, validate.py and preview.py handle output and checks.