3D Object Library
A library of high-fidelity, fully decomposable 3D objects spanning indoor and outdoor environments. Each object is individually editable and composable, making the library ideal for training 3D-aware foundation models and for use across a broad range of AI and visualisation tools.
PBR materials are available on demand for all objects, with a subset already PBR-ready for immediate delivery.
Supported delivery formats include mesh files (FBX, OBJ), code-format scripts (Houdini, Blender, SCAD, JSCAD, STP / STEP CAD, etc.), and Three.js — a web-native format well-suited for AI research pipelines.
Optional add-ons include physics and dynamics metadata, collision data, and NavMesh — available upon request.
3D Scene Library
Photorealistic scenes covering a wide range of indoor and outdoor environments, built in Unreal Engine 5 with full PBR materials and physically accurate lighting.
Indoor scenes are available as simple single-room environments or complex multi-room custom builds. Outdoor scenes range from street blocks, plazas, and small parks to full neighbourhoods, multi-block urban areas, and extended terrain such as farms, forests, and industrial zones.
Every 3D Scene Library delivery includes:
- RGB video frames
- Ground-truth depth maps (per-pixel, per frame — extracted directly from Unreal Engine 5)
- Ground-truth camera poses per frame (engine-rendered)
- Human-annotated trajectory data for navigation training (where applicable)
Optional add-on modalities:
- Normal maps (per-frame surface normals)
- Multi-view synchronous capture (multiple synchronised virtual cameras)
- 3D bounding box annotations with Semantic and Instance Labels
- Collision data and NavMesh for physics-enabled scenes
Supported delivery formats: Point Cloud (.ply / .las) · Mesh (FBX, OBJ, USD/USDZ) · NeRF · 3D Gaussian Splatting (3DGS) · Isaac Sim / USD (simulation-ready, for robot policy training in NVIDIA Isaac Sim).
Modular, decomposable scene assets are provided where applicable for object-permanence training.
3D Scene Library - Sub-library
Game-like Recording Library
High-resolution game-like recordings rendered from 3D scenes using Unreal Engine 5. These recordings extend the 3D Scene Library with real-time-style simulation and frame-level synchronization, adding control signal data essential for VLA and embodied AI training.
Every game-like recording includes all standard 3D Scene Library modalities, plus:
- Ground-truth metric depth (from Unreal Engine 5)
- Synchronised keyboard and mouse / gamepad input logging at every frame
- Single or multi-view synchronous capture from freely configurable virtual cameras
- Rigorous manual validation: Verified synchronisation and spatial alignment of camera poses, RGB frames, and depth maps, with multi-round checks for clipping, ghosting, and blur artefacts
Additional modalities available on request:
- Normal maps and HDR / IBL lighting data
- PBR material maps
- Unreal Engine 5 scene assets or rendered video sequences, depending on delivery scope
Customised trajectory clips are available with client-specified camera paths and variable clip lengths.
Egocentric Industrial Video Library
First-person perspective recordings of human operators performing tasks in real industrial environments, purpose-built for embodied AI and VLA (Vision-Language-Action) model training.
Footage is captured from the operator’s point of view, covering a range of industrial tasks and equipment interactions. Deliverables include raw video footage plus task-relevant annotations. Specific modalities and annotation scope are confirmed per engagement.
