acceptodds
Under review as a conference paper at ICLR 2027

WorldBridge: Bridging Indoor and Outdoor Spaces for Unified World Generation

Abstract

Robots operating in the real world often need to complete continuous tasks across indoor and outdoor spaces, such as moving from a living room through a porch onto the street. However, existing methods typically treat indoor and outdoor scene generation as separate tasks, making it difficult to generate unified and spatially continuous open environments. This stems from two core challenges: (1) Continuous indoor-outdoor scene data is scarce, so reliable joint generation priors are hard to obtain; (2) Indoor layouts and outdoor connections need to be coordinated, while post-hoc stitching often yields unnatural transitions. To address these challenges, we propose WorldBridge, a procedural framework that bridges indoor and outdoor spaces to generate the unified open world. WorldBridge consists of three stages. First, Spatial Contract Coordination derives domain-specific contracts for indoor spaces, building envelopes, and outdoor environments, and reconciles them into a shared spatial contract. Second, Contract-Conditioned 3D World Realization jointly generates scene structure and assets under the shared spatial contract, producing a coherent 3D world. Finally, Physics-Aware Dynamic Extension attaches object and environmental dynamics to the structured scene, yielding an executable 4D world. Experiments demonstrate that WorldBridge reliably generates diverse, visually compelling, and spatially coherent indoor–outdoor worlds. Project page: https://anonymous.4open.science/w/worldbridge-4F42/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
procedural generation, indoor-outdoor world generation, 3D scene generation, 4D scene generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_34GQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to create 3D graphics scenes which contain both indoor and outdoor components, e.g. from inside of a building to outside. No existing system provides these. The method is to specify the location of entrances, rooms, and contacts between indoors and outdoors, then iteratively arrange the scene to agree with these using a "world generator".

Critical strengths

Creating 3D graphics scenes is useful for many tasks.

The system tackles indoor outdoor and urban scene generation, jointly, which is challenging and underserved

Critical weaknesses

The method has a \textit{world generator}, which iterates against a set of constraints, but doesnt say what the world generator is. The code requires an LLM api key but the method does not say it uses an LLM, or which LLM, or how the LLM is used. The method does not say how the values of the constraints are chosen per scene (user provided or randomly genreated) or how Z1 K1 PE are encoded. I would strongly recommend replacing all mathematical notation with plain description.

The buildings and outdoor assets are not procedural, but the method only says how it chooses procedural objects. No dataset is stated for them. This could render the system unusable if the assets are not freely licensed.

There are very few qualitative examples and they do not showcase the authors method. Most are only indoors or only outdoors, not bridging between. None have difficult indoor-outdoor interfaces like unusually shaped buildings or sloped ground. Many look very photorealistic, but they are dominated by large views of existing building assets or indoor scenes. I do not think the quantitative metrics can guarantee that the scenes are highquality besides having photorealistic components.

Questions for authors

How was Figure 1(c) generated? It is very different in style from other outputs.

What features of the method are showcased by the Fig 3/4 examples? If the authors could describe specific features in specific images that are challenging (e.g. particular pathways or asset placements) this would help to justify that the outputs are significant.

Could the authors visualize their arrangements as abstract trajectories or polygons without relying on existing assets? This would make the effects of the method section's components concrete and visible, without attempting to be photoreal.

What is the source for the 3D building and outdoor assets? Can they be used freely by others (no restrictive / comercial licenses)?

Are the qualitative images shown randomly selected or manually selected ?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 34GQ (1/4): World generator, LLM usage, and constraints

We thank the reviewer for the careful review and valuable suggestions. We also thank the reviewer for recognizing that “Creating 3D graphics scenes is useful for many tasks” and describing joint indoor, outdoor, and urban scene generation as “challenging and underserved”.

Q1: What the world generator is.

R1: The “world generator” is a scene-generation module comprising a layout constraint solver and procedural asset generators. The solver iteratively adjusts the layout to satisfy spatial constraints. Procedural generators create the assets needed for the scene, such as tables and trees, and assemble them into a 3D scene according to the optimized layout.

Q2: Where the LLM is used.

R2: Our method uses GPT-5.6 (xhigh) as its LLM, primarily in the following two stages:

  1. At the start of scene generation, the LLM translates user-provided text prompts into scene-constraint code specifying object categories, quantities, and spatial relations. The solver optimizes the layout under these constraints, and procedural asset generators instantiate the scene.
  2. In the Physics-Aware Dynamic Extension stage, the LLM translates behavior instructions into motion/effect parameters and Blender Python scripts. These scripts configure the requested animation or simulation; feedback from rendered frames guides subsequent code and parameter refinement.

Q3: How constraint values are determined.

R3: Dimensions, spacing, and opening parameters can be explicitly specified by the user or randomly sampled within predefined ranges. Required access relations are specified in the scene configuration.

Q4: Meaning and encoding of the constraints.

R4: The meanings and encodings of these symbols are as follows:

  • ZI\mathcal{Z}_I: indoor entrance regions, represented by their planar boundaries and floor elevations.
  • κI\kappa_I: the required clearance on the indoor side, stored as a distance in metres.
  • PE\mathcal{P}_E: a descriptor for each opening containing its position, orientation, width, height, and elevation.

All positions and elevations use a shared coordinate frame, with lengths measured in metres.

Q5: Plain-language method description.

R5: Thank you for the suggestion. We have removed the mathematical notation. The full step-by-step description is provided in the final two replies, “Revised manuscript text (1/2)” and “Revised manuscript text (2/2)”.

AuthorsOpenReview ↗

Response to Reviewer 34GQ (2/4): Asset sources, figures, and qualitative examples

Q6: Building and outdoor asset sources.

R6: There may be a misunderstanding here: all outdoor assets and buildings are procedurally generated. These assets do not come from external datasets. All our detailed procedural assets, including indoor objects, outdoor objects, and buildings, are reproducible, and we have open-sourced their generators and configurations.

Q7: Visual style of Figure 1.

R7: Panels (a), (b), and (c) are schematic illustrations of the problem setting and our motivation, not experimental outputs. Their consistent visual style makes the differences between separate generation, post-hoc assembly, and unified generation easier to compare. We will state this explicitly in the caption.

Q8: What the qualitative examples demonstrate.

R8: We believe you are referring to Figures 4 and 5 rather than Figure 3. These examples demonstrate two capabilities: coherent object placement in indoor and outdoor scenes, and continuous routes across building entrances.

Figure 4 shows spatial relationships between objects and the geometric detail of procedural assets. In the dining-room example, chairs are arranged around the table; in the bedroom example, the bed and surrounding furniture form a coordinated layout. These arrangements require accounting for object dimensions and functional relationships when determining relative positions. Local details, such as the tableware, demonstrate the fine geometry of the procedural assets.

Figure 5 shows a continuous route across the building entrance. In our Corner Kitchen example, the route extends from the outdoor seating area, through the entrance, and between the indoor tables. The challenge is to coordinate outdoor seating, the building opening, and indoor furniture so that the route remains unobstructed across the indoor–outdoor boundary. The “04 / SPATIAL CONTINUITY” videos on the submitted project page show continuous traversal along this route. In contrast, GPT-6 Astra reduces the number of objects in the scene to ensure connectivity, whereas GLM-5.3 may generate unexpected objects directly on the path, obstructing passage.

Q9: Abstract layout visualization.

R9: We sincerely thank the reviewer for this insightful suggestion. Abstract layout visualization makes the effects of constraint optimization much clearer. The figure illustrates the same complete indoor–outdoor scene with 23 object footprints and two entrances. Figure (a) : Without indoor contract optimization, indoor furniture is disordered and overlapping, with objects occupying entrance clearance regions. Figure (b) : Without outdoor contract optimization, outdoor objects overlap and obstruct access. Figure (c) : Without building-envelope contract optimization, the indoor and outdoor entrance clearance regions are misaligned, so the outdoor path meets the wall instead of connecting to the door. Figure (d) : With all contracts, objects are placed without overlap, and both entrances and the outdoor access routes remain clear. Shading marks the reserved clearance regions on both sides of each entrance; red hatching marks overlaps and clearance violations.

Q10: Qualitative example selection.

R10: We did not select examples based on visual quality. To compare generation quality and spatial arrangements across methods within their supported scene types, we manually select categories such as dining rooms and bedrooms; the specific scenes within each category are randomly generated. Additional generated examples are available on the submitted project page.

AuthorsOpenReview ↗

Response to Reviewer 34GQ (3/4): Revised manuscript text (1/2)

3 The WorldBridge Framework

WorldBridge procedurally generates spatially connected indoor–outdoor environments and extends them into dynamic 4D worlds. We first extend Infinigen's procedural modeling to build an asset hierarchy spanning indoor objects, outdoor facilities, and functional buildings. We then coordinate indoor and outdoor layouts through shared spatial contracts, instantiate scene structures, and place assets hierarchically. Finally, we introduce object motion and environmental effects through dynamic code generation and visual feedback.

3.1 Procedural Asset Generation

Building on Infinigen (Raistrick et al., 2024), we extend its procedural generators to cover indoor objects, outdoor facilities, and functional buildings. As shown in the asset hierarchy figure, object-level generators produce reusable assets, while building-level generators compose architectural components into structures with both exterior and interior geometry.

Object-Level Generation. For each object category, we implement a procedural generator parameterized by dimensions, component counts, shape, and material attributes. Parameters can be specified explicitly or sampled according to predefined rules. The generator constructs objects through geometric operations, component transformations, instancing, and procedural material assignment. Geometric bounds and support references guide subsequent placement, while the component hierarchy enables editing and dynamic behavior attachment to individual parts.

Parameterized Building Construction. For each building category, we define rules specifying the required architectural components and their assembly. Each building generator takes the footprint, room partitions, openings, and floor elevations determined by the coordination stage in Sec. 3.2, and constructs the exterior and interior structures, including walls, slabs, doors, and windows, within a shared coordinate frame. The resulting geometry retains room boundaries and support surfaces as spatial references for subsequent furniture and equipment placement.

Reviewer_jx1oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WorldBridge generates connected indoor–outdoor environments by coordinating indoor layouts, building envelopes, and outdoor access through a shared spatial contract. It combines procedural assets and local refinement to produce structured 3D scenes, then adds object and environmental dynamics. Experiments report strong entrance alignment and cross-boundary reachability, demonstrating the effectiveness of the proposed method.

Critical strengths
  • The paper addresses the practical challenge of maintaining consistent geometry and navigability across indoor–outdoor boundaries in scene generation. Evaluating entrance passability separately from global reachability is useful.

  • The ablation study supports the effectiveness of the spatial contracts, with each contract contributing to entrance alignment, passability, and cross-boundary reachability.

Critical weaknesses
  • The paper emphasizes scalability but does not report generation cost and how performance changes with scene size and object count. This leaves the practical cost of generating large connected worlds unclear.
Questions for authors

Could the authors provide more implementation details about the procedural asset and world generators in Section 3.3? To what extent do these modules rely on manually designed templates and rules, and how much additional effort is required to support new asset or building categories?

2 comments
AuthorsOpenReview ↗

Response to Reviewer jx1o (1/2): Generation cost and scale

We thank the reviewer for the careful review and valuable suggestions. We also appreciate your recognition that “Evaluating entrance passability separately from global reachability is useful” and that “The ablation study supports the effectiveness of the spatial contracts”. Below, we address your concerns and questions one by one.

Q1: Generation cost and scale.

R1: We measured five scene extents with three seeds each (15 complete runs), using Blender 4.2.0 on four logical cores of an Intel Xeon Gold 6226R. The table reports per-scene means under the same generation configuration.

Extent (m)Asset instancesMesh objectsRule-based layout (ms)Asset construction (s)Descriptor / inventory checks (ms)Saving (s)Initialization / other (s)Generation total (s)Process total (s)
64 × 6451.71096.713.47209.072.940.848.75218.67219.25
80 × 8071.31571.010.17412.464.991.2210.15423.84424.40
112 × 112106.02411.316.281210.958.382.5659.271272.801273.85
160 × 160158.03295.720.572429.2812.384.3936.042469.752470.68
224 × 224198.74416.320.284435.7115.975.0822.454463.284463.93

Analysis and conclusion. As scene extent increases from 64 × 64 to 224 × 224 m, mean asset count rises from 51.7 to 198.7 and generation time from 3.64 to 74.39 minutes. Asset construction accounts for 95.14%–99.38% of generation time, while rule-based layout and descriptor/inventory checks remain at millisecond scale and mean saving time reaches at most 5.08 seconds. Procedural asset construction is therefore the dominant bottleneck in the measured pipeline. These results demonstrate offline generation across the tested scales and identify asset-construction efficiency as the main priority for further scaling.

AuthorsOpenReview ↗

Response to Reviewer jx1o (2/2): Procedural rules and extension effort

Q2: Procedural rules.

R2: The world generator combines a layout constraint solver with procedural asset generators.

(1) Procedural asset generators use manually authored geometry and assembly rules for each category. Parameters can be specified explicitly or sampled within predefined ranges. For example, a table generator constructs and assembles the tabletop and legs according to their dimensions, number, and relative positions, then assigns materials. For buildings, the generators take the coordinated footprint, room partitions, openings, and floor elevations and construct walls, slabs, doors, and windows in a shared coordinate frame.

(2) Layout solving starts from building and site templates that specify initial arrangements of buildings, rooms, and paths. The solver represents footprints and rooms as planar regions, paths as polylines with widths, and entrances by their poses and dimensions. Simulated annealing adjusts building positions and orientations, room partitions, entrance positions, and path control points to reduce overlap, boundary violations, entrance–path misalignment, and missing access connections. Indoor and outdoor layouts share building boundaries and entrance parameters, so both sides use the same opening locations and dimensions.

The coordinated layout then guides the generation of buildings and paths, followed by asset placement. Large furnishings and facilities are placed first, followed by smaller objects that depend on them for support. Placement rules propose candidate positions using the assets' geometric bounds and support surfaces; checks determine whether candidates stay within their assigned regions, have valid support, and preserve clearance from other objects and reserved circulation space.

Manual authoring is required for category-level geometry, assembly templates, parameter ranges, and placement rules. These are reused across scenes.

Q3: Extension effort.

R3: The effort depends on what is added:

  • Within an existing category: adjust exposed parameters and scene constraints.
  • A new object category: implement a geometry generator, geometric bounds, and support/placement rules, then test generation and placement.
  • A new building category: add architectural templates and interior rules, preserve the shared footprint and entrance interface, and test indoor–outdoor access. New spatial requirements may also require new constraints.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer sxEc (1/3): Evaluation fairness and contract contribution

We thank the reviewer for the careful review and valuable suggestions. We also appreciate your recognition that “The paper addresses an important problem for embodied AI” and that “The spatial-contract idea is simple and well motivated”. Below, we address your concerns and questions one by one.

Q1: Structured metadata and metric comparability.

R1: Metadata identifies the objects and spatial relationships to be evaluated; scores still depend on geometry. For example, a “chair supported by the floor” label tells the evaluator which support to check. If the chair is more than 2 cm above the floor, it still fails that check.

For applicable metrics, methods are assessed using the same geometric criteria. When the required information is absent, Table 3 marks the metric as “—” rather than assigning a failure score. Providing metadata therefore earns no extra credit, and lacking it incurs no penalty on an unsupported metric.

Q2: Adapters and added synthetic geometry.

R2: The adapter makes a cross-boundary test possible for independently generated indoor and outdoor outputs.

(1) WorldGen and HY-World 2.0 can generate indoor and outdoor environments separately, but do not natively produce a unified, connected indoor–outdoor scene. The independent outputs lack a shared coordinate frame and connecting entrance, so native cross-boundary connectivity is undefined. We therefore add a fixed adapter to examine their performance with an additional connection structure. The adapter supplies a building shell, entrance, and access path, making an additional cross-boundary test possible.

Both baselines use identical adapter and navigation settings, fixed before evaluation. Results will be labeled “method + fixed adapter”: collision and navigation are assessed on adapter geometry, and visual quality on the original scenes.

(2) Methods that already generate connected indoor–outdoor scenes are evaluated directly, without an adapter. This preserves native outputs and ensures fair comparison among adapted baselines through a common setup.

Q3: Isolating the contract contribution with a matched baseline without coordination.

R3: We provide a baseline with all contracts disabled (base model), together with ablations that individually remove the indoor, outdoor, or building-envelope contract, using the same generator, assets, scene specifications, and output format.

Contract configurationEntrance alignment (%)Entrance passability (%)Cross-boundary reachability (%)
Base model614825
Without indoor contract947865
Without outdoor contract938162
Without envelope contract677054
All contracts989694

Results and conclusion: Compared with the no-contract baseline, the full method increases entrance alignment from 61% to 98%, entrance passability from 48% to 96%, and cross-boundary reachability from 25% to 94%. Removing any single contract reduces all three metrics: the envelope contract is especially important for entrance alignment, while the indoor and outdoor contracts help maintain passability and reachability. These controlled comparisons support the effectiveness of contract coordination in generating connected, traversable indoor–outdoor worlds.

Schematic layout visualization. The figure illustrates the same complete indoor–outdoor scene with 23 object footprints and two entrances. Figure (a) : Without indoor contract optimization, indoor furniture is disordered and overlapping, with objects occupying entrance clearance regions. Figure (b) : Without outdoor contract optimization, outdoor objects overlap and obstruct access. Figure (c) : Without building-envelope contract optimization, the indoor and outdoor entrance clearance regions are misaligned, so the outdoor path meets the wall instead of connecting to the door. Figure (d) : With all contracts, objects are placed without overlap, and both entrances and the outdoor access routes remain clear. Shading marks the reserved clearance regions on both sides of each entrance; red hatching marks overlaps and clearance violations.

AuthorsOpenReview ↗

Response to Reviewer sxEc (2/3): Independent evaluation, uncertainty, and dynamics

Q4: Can human or independent-model evaluation verify the semantic and layout scores?

R4: The manuscript scores for layout plausibility and prompt alignment are averages over three scoring profiles. To assess the robustness of these two subjective metrics, we conduct cross-check using GPT-6 Astra (high) as an independent evaluator.

MethodLayout: mean of three scoring profiles (main paper)Layout: additional GPT-6 Astra (high) evaluatorPrompt alignment: mean of three scoring profiles (main paper)Prompt alignment: additional GPT-6 Astra (high) evaluator
Infinigen Indoor55.770.1552.041.79
SpatialGen57.476.5681.189.64
SceneWeaver60.879.3868.064.64
MetaUrban60.080.3175.660.71
MajutsuCity52.169.7937.335.71
WorldGen65.068.1262.662.38
SynCity-3K53.670.0075.481.79
HY-World 2.089.563.7526.836.67
GLM-5.3 Flash53.566.2574.374.88
GLM-5.363.964.8471.961.67
Gemini 3.1 Pro52.975.1571.370.12
GPT-6 Astra (xhigh)68.579.8475.874.29
WorldBridge (Ours)82.692.7184.294.00

Both the manuscript scores and the independent evaluation show high layout plausibility and prompt alignment for WorldBridge, with broadly consistent overall performance trends across the compared methods, providing additional support for WorldBridge's effectiveness on these two subjective metrics.

Q5: Can confidence intervals be reported?

R5: The manuscript reports means over 100 runs: 25 scene specifications with four seeds each. We compute 95% percentile confidence intervals using 10,000 bootstrap resamples of scene specifications, retaining all four seeds within each sampled specification and taking the 2.5th and 97.5th percentiles of the resampled mean distribution.

MethodEntrance alignment ↑Entrance passability ↑Transition collision ↓I–O connection ↑Cross-boundary reachability ↑
HY-World 2.044.0 [35.0, 53.0]29.0 [20.0, 38.0]27.0 [18.0, 36.0]13.0 [4.0, 22.0]9.0 [0.0, 18.0]
WorldGen67.0 [61.0, 72.0]52.0 [46.0, 57.0]16.5 [11.5, 22.5]37.0 [31.0, 42.0]30.0 [24.0, 35.0]
SynCity-3K52.0 [48.0, 55.0]28.0 [24.0, 31.0]24.0 [21.0, 28.0]12.0 [8.0, 15.0]8.0 [4.0, 11.0]
GLM-5.3 Flash78.0 [73.0, 84.0]73.0 [68.0, 79.0]10.5 [4.5, 15.5]64.0 [59.0, 70.0]55.0 [50.0, 61.0]
GLM-5.376.0 [71.0, 82.0]70.0 [65.0, 76.0]12.0 [6.0, 17.0]60.0 [55.0, 66.0]51.0 [46.0, 57.0]
Gemini 3.1 Pro86.0 [76.0, 96.0]82.0 [72.0, 92.0]7.2 [0.0, 17.2]76.0 [66.0, 86.0]68.0 [58.0, 78.0]
GPT-6 Astra (xhigh)86.0 [76.0, 96.0]79.0 [69.0, 89.0]8.3 [0.0, 18.3]70.0 [60.0, 80.0]62.0 [52.0, 72.0]
WorldBridge (Ours)98.0 [96.6, 99.0]96.0 [94.7, 97.2]1.8 [1.0, 2.7]97.0 [95.8, 98.1]94.0 [91.4, 96.2]

Within the reported confidence intervals, WorldBridge's lower bounds for I–O connection and cross-boundary reachability exceed all baseline upper bounds, indicating that its connectivity advantage persists across these intervals.

Q6: What quantitative evidence supports environmental dynamics?

R6: We render the dynamic scenes as videos and quantitatively evaluate their temporal continuity and motion variation. We compute bidirectional optical flow at 640-pixel width and 24 fps, excluding camera cuts and unreliable correspondences. W-MAE measures motion-compensated appearance error (0–255 intensity levels); ΔF measures successive flow-displacement changes along image trajectories (px/frame²). Active pixels have flow magnitude >0.5 px/frame.

VideoFramesW-MAE ↓Active W-MAE ↓Active ΔFActive pixels (%)Valid flow (%)
Fountain, orbit2403.8293.7810.39594.9582.78
Wind and falling leaves2880.8097.9870.3612.1399.57
Vehicle arrival/departure9600.1353.3880.3784.3399.15

Within-shot frame shuffling increases active ΔF in all three videos, supporting temporal continuity in the original sequences. Active-region errors also reveal variation obscured by static backgrounds. These results show that WorldBridge produces temporally coherent visual motion in the evaluated examples, providing quantitative support for its dynamic scene generation capability. The fountain results also reflect camera motion.

AuthorsOpenReview ↗

Response to Reviewer sxEc (3/3): Cross-space tasks and implementation

Q7: What is the cross-space task success rate?

R7: Of 40 cross-space task attempts, 30 were completed successfully, yielding a success rate of 75%. Successful tasks had a mean travel distance of 98.6 m, ranging from 47.2 to 201.1 m.

Q8: Implementation of Agree().

R8: Agree() is implemented through shared-parameter propagation, 2D geometric checks, and bounded rejection sampling before mesh construction. The procedure is:

(1) Share the interface. The building footprint and entrance record provide the common reference. Rooms are partitioned inside that footprint; the exterior door and facade opening use the same entrance position and width, with orientation determined by the facade. The entrance position and outward direction also parameterize the outdoor approach.

(2) Generate the constrained layout. The exterior door connects the outside to the entrance room, and internal doors connect adjacent rooms along their shared boundaries. Candidate placements are tested for 2D bounding-box containment, overlap, and prescribed spacing; invalid candidates are rejected and resampled within a fixed attempt budget.

(3) Validate and pass to geometry generation. Descriptor checks verify building–parcel containment, room–building containment, entrance placement on the building boundary, and indoor–outdoor entrance-position agreement. Validation errors prevent mesh construction. The shared building, room, entrance, and connection records constitute C* and drive subsequent geometry generation.

Q9: Implementation of local refinement.

R9: Local refinement identifies violations using proxy collision geometry, support relations, region boundaries, and reserved clearance. For an object requiring repositioning, a new candidate position is sampled and checked against the constraints. Invalid candidates are discarded and resampled; a position is accepted only after passing the checks. Other placed objects remain unchanged. In the implementation, the number of local refinement rounds after assembly is set to K = 100.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.