A comprehensive spatial library purpose-built for AI training, spanning 3D objects, photorealistic scenes, game-like recordings, and egocentric industrial videos. Objects are available across mesh, code-format script, and Three.js delivery formats. Contact us for current library scope and availability.
3D scenes ship with the full modality set that world models need: RGB, ground-truth depth, ground-truth camera pose, and trajectory data. Optional add-ons include normal maps, multi-view synchronous capture, 3D bounding box annotations, and control signal logs.
Keyboard, mouse, and gamepad input logs per frame are available in the game-like recording sub-library of the 3D Scene Library — essential for VLA and robot imitation learning.
VoxelCore provides comprehensive coverage of indoor (simple and complex) and outdoor (simple and complex) environment types. Most synthetic data providers focus only on roads.
Scenes can be exported in NVIDIA Isaac Sim-compatible USD format with physics properties, articulation definitions, and sensor placement, ready for robot manipulation and navigation policy training.
Real-world captured scenes are deliverable as 3DGS or NeRF representations in addition to point cloud and mesh formats, supporting novel-view synthesis and embodied AI applications.
Game-like recordings offer photorealistic visual diversity that bootstraps world model pre-training faster than lower-fidelity synthetic datasets.
VoxelCore-owned and licence-clean. Data ships in standard formats compatible with common data loaders — integration takes hours, not weeks.
Diverse scene coverage and ground-truth sensor data for large-scale pre-training of general-purpose world simulators and video prediction models.
Control signal logs paired with visual data provide ground-truth action labels for Vision-Language-Action model training.
Long-tail outdoor scenes (adverse weather, complex geometry) to supplement training pipelines alongside real-world sensor data; ground-truth metric depth enables perception model validation.
Indoor and outdoor 3D environments with ground-truth trajectory, game-engine control signals, and Isaac Sim-compatible scene export to support robot navigation and manipulation policy learning.
High-fidelity 3D object assets (including Three.js, code-format scripts, and mesh files) and single / multi-view captures from 3D scenes provide rich training signal for 3D generation, Gaussian Splatting research, and NeRF reconstruction.
