Langhui Tech has deeply cultivated 3D spatial intelligence, building 18 high-quality multimodal datasets across four major domains: 3D scenes, point clouds, spatial semantics, and digital twins. Covering the full pipeline from photo-based modeling, UAV aerial photography, laser point clouds, to NeRF/3DGS reconstruction, we provide a solid training foundation from perception to generation for world models and spatial intelligence.
From Sora to Genie 2, from 3DGS to V-JEPA, world models are becoming the key path to AGI, and spatial understanding capabilities are ushering in a qualitative leap
OpenAI's Sora 2 takes video generation to new heights, capable of simulating physical world laws and generating coherent videos up to 1 minute long. Regarded as a milestone of "visual world models," it can understand gravity, collision, lighting, and other physical common sense, taking a key step toward AGI.
Google DeepMind releases Genie 2, which can generate explorable 3D virtual worlds from a single image, supporting keyboard-controlled character movement, jumping, and interaction. The world model can infer playable 3D environments lasting up to 1 minute from a single frame.
3DGS technology comprehensively surpasses NeRF in novel view synthesis, with 100x faster training speed and real-time rendering. Multiple SIGGRAPH 2025 papers apply 3DGS to dynamic scenes, large-scale outdoor reconstruction, and editable 3D content generation.
Yann LeCun's team's V-JEPA 2 is based on a joint-embedding predictive architecture, learning physical world laws from video through self-supervised learning, understanding object persistence and motion patterns without any annotation. Regarded as a key path to AGI.
Breakthroughs in world models for robotics, with models like NVIDIA GR00T enabling robots to learn manipulation skills by "imagining" future scenarios. Trained on millions of 3D scene data, robots can perform tasks zero-shot in unseen environments.
NVIDIA Omniverse has become the standard platform for industrial digital twins, supporting city-level and factory-level real-time simulation. In 2026, it integrates generative AI capabilities to automatically generate twin scenes, predict equipment failures, and optimize production processes.
Deeply cultivating four major domains: 3D scenes, point clouds, spatial semantics, and digital twins, building full-pipeline spatial data infrastructure from multi-source collection to intelligent annotation
Multi-view photo-based modeling, UAV oblique photography, and NeRF/3DGS training pair data, covering multi-scale 3D scenes from objects to buildings to cities.
LiDAR point clouds and mmWave radar data, including semantic annotations and instance segmentation, supporting high-precision spatial perception and mapping.
High-precision semantic data for multiple scenarios including parks, campuses, factories, ancient buildings, and cities, with instance-level and part-level fine annotations.
Digital twin data foundation for cities, factories, and parks, supporting real-time simulation, predictive maintenance, and collaborative decision-making.
Integrating UAVs, LiDAR, multi-view cameras, mobile mapping vehicles, and other multi-source collection devices, covering indoor and outdoor full scenes
Self-developed NeRF/3DGS/MVS reconstruction algorithms, millimeter-level precision, supporting large-scale real-time scene reconstruction and rendering
AI-driven large-scale semantic annotation platform, supporting automatic segmentation, instance recognition, and semantic understanding, with 10x efficiency improvement
Strictly following data security regulations, all spatial data is collected with authorization, providing compliant delivery solutions
Focusing on four major domains: 3D scenes, point clouds, spatial semantics, and digital twins, building the world model training data foundation
Multi-view photo-based 3D model reconstruction data, covering objects, buildings, and scenes, supporting photo-level 3D reconstruction and NeRF training.
UAV aerial oblique photography and orthophoto data, supporting large-scale urban and terrain 3D reconstruction.
High-quality paired data from images to 3D, including camera poses, depth maps, and novel view synthesis supervision signals, supporting NeRF/3DGS training.
3D model part-level segmentation and assembly data, supporting editable 3D generation and structured reconstruction.
Text-3D model alignment data, including Prompt, material, and geometric description annotations, supporting text-to-3D content AI training.
LiDAR point cloud data with semantic annotations, supporting high-precision spatial perception and mapping.
In-vehicle LiDAR collected outdoor large-scene point clouds, with 3D bounding boxes and object tracking annotations, supporting autonomous driving perception.
Indoor scene LiDAR scan point clouds, with instance segmentation annotations for rooms, furniture, equipment, etc., supporting indoor navigation and reconstruction.
4D mmWave radar point cloud data, complementing LiDAR's perception capabilities in adverse weather, supporting all-weather autonomous driving.
Multi-scenario high-precision semantic datasets, with instance and part-level annotations, covering buildings, roads, vegetation, facilities, and more.
Semantic annotation data for 3D models of ancient buildings and artifacts, with fine annotations for components, ornaments, and materials, supporting digital preservation.
Large-scale urban aerial semantic segmentation data, covering buildings, roads, vehicles, pedestrians, and other urban elements.
3D point cloud instance segmentation annotation data, supporting fine-grained 3D understanding at object and part levels, enabling robotic manipulation.
2D/3D panoptic segmentation annotation data, fusing semantic and instance segmentation, supporting comprehensive scene understanding.
Smart factory 3D twin data, including equipment models, production line layouts, and operational status, supporting the industrial metaverse.
City-level 3D digital twin data foundation, including buildings, roads, pipeline networks, traffic, and other urban elements, supporting smart city applications.
3D twin data for industrial parks, campuses, and scenic areas, including buildings, greenery, facilities, and dynamic crowd flow information.
3D twin data for power, oil & gas, new energy, and other energy facilities, supporting equipment operation and maintenance and fault prediction.
Showing the JSON annotation structure of 4 data types: 3D scenes, point clouds, spatial semantics, and digital twins
Six typical AI application scenarios based on Langhui World Model/Spatial datasets
3D reconstruction based on multi-view images or point cloud data, building high-precision digital models, supporting digitalization in cultural heritage, industry, architecture, and other fields.
Robot autonomous navigation, obstacle avoidance, and manipulation based on spatial semantic understanding, supporting industrial, service, logistics, and other multi-scenario robot applications.
Building digital twin models for cities, factories, and parks, supporting real-time simulation, predictive maintenance, and collaborative decision-making, empowering the industrial metaverse.
Generating 3D models from text or images, accelerating game, film, and metaverse content creation, with AI driving a content production revolution.
AR/VR content overlay and interaction based on spatial perception, supporting immersive applications in education, training, tourism, and more.
Training autonomous driving perception models based on point cloud and multi-view data, supporting high-level autonomous driving environment understanding and decision-making.
Frequently asked questions about World Model/Spatial datasets
Langhui World Model/Spatial datasets cover four major domains: 3D scenes, point cloud data, spatial semantics, and digital twins, totaling 18 core datasets. Includes photo modeling, UAV aerial photography, NeRF/3DGS training pairs, laser point clouds, autonomous driving point clouds, indoor point clouds, mmWave radar, park/campus/factory/ancient building/city high-precision semantic, 3D instance segmentation, panoptic segmentation, factory/city/park/energy twins, etc. Totaling 7.25M+ multi-view photos, 1TB+ point cloud data, and 2000km² aerial coverage.
Suitable for six typical scenarios: 3D reconstruction, embodied intelligence robots, digital twins, 3D content generation, AR/VR interaction, and autonomous driving. Supports training of world models, NeRF/3DGS, spatial understanding, embodied intelligence, and other AI large models, providing spatial understanding capability training data for Artificial General Intelligence (AGI), and has served multiple research institutions and enterprises.
Collection precision reaches millimeter-level (laser point clouds), centimeter-level (UAV aerial photography), and 4K/8K ultra-high-definition (photo modeling). Coverage includes: indoor and outdoor full scenes, city-level large-scale (2000km² aerial coverage), multi-scale (object/building/park/city). Collection devices cover multi-source sensors including UAVs, LiDAR, multi-view cameras, mobile mapping vehicles, and 4D mmWave radar.
We adopt a "expert review + AI assistance" dual quality control system. Our self-developed AI-assisted annotation platform performs automatic segmentation, instance recognition, and semantic understanding, with 10x efficiency improvement. A professional team conducts multiple rounds of review for annotation quality. Semantic annotation coverage reaches 92%, reconstruction precision reaches millimeter-level, and data update frequency is quarterly, ensuring data timeliness and accuracy.
Supports image formats (JPEG/PNG/TIFF/RAW), 3D model formats (OBJ/PLY/GLTF/FBX/USD), point cloud formats (LAS/LAZ/PCD/PLY), annotation formats (JSON/XML/COCO/KITTI/NuScenes), twin data formats (USD/Industry Foundation Classes), and various other delivery methods. Supports customized delivery and API integration.
Contact Langhui data experts to get the complete World Model/Spatial dataset catalog, trial access, and customized solutions