Spatial Foundry

Spatial learning requires correspondence between visual observations and the underlying scene structure. Explicit geometry, object identities and camera calibration provide supervision beyond image appearance alone.

Floor plans, point clouds and editable meshes provide complementary representations of scene structure. Calibrated perspective and panoramic RGB are associated with metric depth, object IDs and 3D bounding boxes. Video frames are paired with camera trajectories and per-frame poses.

Residential layout estimation, metric 3D reconstruction, instance-level scene understanding, embodied navigation and camera-conditioned world modeling.

Scene-level splits support evaluation across environments. Metric geometry and calibrated observations provide references for scale, spatial layout, object localization and cross-view consistency.

The floor-plan, point-cloud and mesh previews depict the same two-floor Blender scene. Floor plans are derived from horizontal mesh sections. Point colors are estimated from UV textures and material Base Color; procedural shaders are approximated. The mesh preview shows actual surface geometry. Floors are separated for visibility in the default 3D previews; stored metric coordinates are unchanged.

Limited environmental diversity can bias models toward familiar layouts and object configurations. Variation in scene structure and furnishings supports evaluation of generalization across residential environments.

Residential scene configurations are paired with images, depth, camera calibration, object annotations and 3D geometry. Camera trajectories and corresponding videos preserve the relationship between local observations and global scene structure. Scene counts include furnishing variants of the same layout.

Cross-scene generalization for depth estimation, reconstruction and object perception; scene retrieval; and camera-conditioned generation across diverse residential layouts.

Partition by underlying layout or scene family to control leakage between furnishing variants. Compare performance across room configurations, object arrangements and camera viewpoints.

The 32-second demonstration combines a 36-scene gallery with first- and second-floor RGB views rendered from the featured Blender scene. The interactive preview uses 400,000 spatially sampled points from the supplied PLY, retaining its stored RGB and original coordinates. Floor separation and ceiling cutaways affect display only. The demonstration presents reference data rather than learned model outputs.

Geometric learning depends on consistent camera models and metric scale. Misalignment between appearance, depth and camera calibration introduces ambiguity into both supervision and evaluation.

Perspective RGB and depth are rendered directly from the source scene, with camera intrinsics and extrinsics retained. Equirectangular panoramas and explicit 3D geometry provide complementary representations of the same environment.

Perspective and panoramic depth estimation, RGB-D fusion, multiview reconstruction, novel view synthesis and geometric evaluation of neural scene representations.

Paired RGB-D observations and calibrated views support depth-error measurement, metric-scale assessment and cross-view reprojection tests.

Perspective RGB and depth are rendered directly and unprojected using camera calibration. The final whole-house cloud is independently sampled from scene meshes and approximately colored by RGB projection; it is not reconstructed from a single panorama.

Instance-level scene understanding requires object identity to remain consistent across viewpoints. Explicit 3D bounds and dimensions provide geometric references for localization and spatial relationships.

Scene meshes, RGB observations and point clouds share object IDs and metric 3D bounds. Projected annotations establish correspondence between visible image regions and the underlying scene instances.

3D object detection, visual grounding, cross-view instance association and structured scene prediction from video.

Evaluate localization, dimensions and instance correspondence against scene annotations. Semantic and orientation conventions are documented with the annotation fields.

The first- and second-floor inspection sequences share object IDs, metric dimensions and world-axis-aligned bounding boxes between RGB and point-cloud views. Displayed names use curated annotations; fine-grained semantic labels have not all been independently audited.

Textual motion instructions do not fully specify a camera trajectory. Explicit pose sequences define position, orientation and timing, enabling reproducible evaluation of camera control and spatial consistency.

Video frames are associated with camera poses, intrinsics and timestamps. Routes through furnished 3D scenes support controlled changes in viewing direction, height and speed while preserving the underlying environment.

Camera-conditioned video generation, visual navigation, controllable scene exploration and spatial-memory evaluation through revisitation.

Condition models on reference observations and target camera poses. Evaluate trajectory adherence, motion continuity and consistency across rooms, floors and revisited locations.

The reference sequence contains 2,923 frames at 24 fps and follows a continuous two-floor camera route. Traversal speed varies across room exploration, stair traversal and low-information transitions. The sequence is rendered in Blender; no learned video model is evaluated.

Multi-agent world modeling requires consistent state across observers. A scene update initiated by one participant should be reflected in other viewpoints and retained in subsequent observations.

Independent participant cameras are linked to shared object identities, transforms and edit events. Synchronized observations provide references for distinguishing viewpoint changes from changes in scene state.

Collaborative exploration, multi-agent spatial reasoning, shared-state prediction and consistency evaluation for online world models.

Compare participant observations against the same state transitions. Assess event synchronization, cross-view agreement and persistence of changes over time.

H1 and ANYmal D represent participants in a local two-client TCP prototype. The video is pre-rendered from synchronized scene states. It does not evaluate robot dynamics or a publicly deployed generative world model.

Spatially consistent editing requires changes to persist beyond a single image. Updated object geometry must remain coherent with occlusion, illumination and subsequent camera viewpoints.

Editable object meshes, materials and transforms support explicit scene-state changes. Before-and-after observations are linked to the corresponding edits and camera poses.

Interactive scene editing, instruction-guided 3D manipulation, action-conditioned world models and counterfactual scene synthesis.

Evaluate whether an edit matches the target transformation, preserves unrelated scene content and remains consistent in subsequent views.

Object translation, removal, insertion and rotation are applied to the Blender scene and rendered along a continuous camera path. The video illustrates data for real-time editing workflows; it is pre-rendered rather than a live editor capture.