acceptodds
Under review as a conference paper at ICLR 2027

Strata3D: Efficient and Accurate Agentic Indoor Scene Generation with Layered Floor Plans

Abstract

Simulated indoor scenes are the training ground for embodied agents, and the scenes that matter most are dense: tens of objects stacked on, tucked under and hung from a few pieces of furniture. Recent agentic generators place objects directly in 3D and discover physical and functional constraints by rendering, critiquing and revising — a search that grows expensive as rooms become dense and still leaves many objects in collision. Our key insight is that many of these constraints are not three-dimensional: whether objects overlap, whether a doorway stays clear and whether every object can be reached are properties of the floor plan, provided the plan records which objects may overlap. We introduce Strata3D, a layout-first framework built on a layered floor plan, in which each object is a footprint on one of five vertical layers. The layers turn these constraints into exact, deterministic checks whose findings guide a language model in revising its plan; the validated plan is then lifted to 3D as a contract that the retrieved assets must honour. On dense prompts, against SceneCode and SceneSmith run to completion, Strata3D improves object-count accuracy by more than ten points, halves the share of colliding objects and wins 41 of 46 blind pairwise comparisons, while using two orders of magnitude fewer tokens. Ablations attribute each gain to the component that targets it, and the generated scenes can be used directly as mobile-manipulation environments.

Primary area:
generative models
Keywords:
3D agents, 3D 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 3 reviews, above 27% of the others.
Reviewer_D9rNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Strata3D, an indoor 3D scene generation framework that uses layered floor plans and deterministic geometric checks to guide layout revision, then retrieves and assembles 3D assets while preserving the planar constraints. Experiments on dense rooms demonstrate improved prompt fidelity and fewer object collisions compared with SceneCode and SceneSmith, alongside substantially lower generation time and token consumption.

Critical strengths

S1. Clear motivation and coherent system design. The layered floor plan distinguishes valid stacking from actual collisions and combines deterministic checking, guarded revision, and constrained 3D assembly to address dense scene generation.

S2. Well-controlled comparisons demonstrate gains in quality and efficiency. With a shared backbone model and asset library, comparisons against complete baseline pipelines show that Strata3D improves prompt fidelity, reduces the proportion of colliding objects to 11.9%, and substantially lowers generation costs. These results support the effectiveness of the overall system.

Critical weaknesses

W1. Limited comparisons with the most closely related methods. The paper does not quantitatively compare against closely related structured approaches, such as SpatialGrammar and HDSL. The specific advantages of its five-layer representation and associated verification mechanisms over existing 2D or hierarchical representations therefore remain insufficiently established.

W2. Unclear benefits of some components and inconsistent evaluation conditions in part of the comparison. Removing the declared-relation module or VLM critic leaves several metrics unchanged or improved, so their necessity is not sufficiently demonstrated. Moreover, the published baseline results in Table 2 use a GPT-4o judge, whereas this paper uses GPT-5.6, limiting the direct comparability of judge-dependent metrics. This concern does not apply to the main dense-room experiment, which uses a unified evaluation setup.

Questions for authors

Q1. What are the specific advantages over closely related methods? Could the authors provide comparisons with methods such as SpatialGrammar and HDSL under a unified setup to establish the additional benefits of the five-layer representation and its verification mechanisms?

Q2. Do improvements in planar validation translate into better functionality in the final scenes? How do the authors explain improved internal reachability alongside a decline in the external NAV.55 metric? Does this reflect differences in metric definitions, or a mismatch between the planar validation rules and the actual usability of the resulting 3D scenes?

Reviewer_R23dOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Strata3D, an efficient agentic framework for text-driven indoor 3D scene generation. Instead of directly generating and refining scenes in 3D, Strata3D introduces a layered floor-plan representation that organizes objects into five vertical layers. A deterministic checker validates spatial constraints and guides LLM-based layout refinement, after which the validated plans are lifted into 3D through asset retrieval, placement, and mesh repair. Experiments demonstrate improved scene quality and substantially reduced computational costs compared to existing agentic approaches, including SceneCode and SceneSmith.

Critical strengths
  1. Efficient and Well-Structured Framework. By shifting spatial reasoning and constraint verification from 3D to layered 2D floor plans, Strata3D significantly reduces the computational overhead of iterative scene generation. The reported improvements in token efficiency and generation time are substantial. According to the experiment results, such scene representations may fix the collision problems in the generated 3D scenes.

  2. Deterministic Constraint Verification. The framework incorporates explicit geometric constraints, including object overlap, support relationships, doorway clearance, and reachability. The checker-guided revision process provides more reliable feedback than relying solely on VLM-based visual critiques.

Critical weaknesses
  1. Limited Expressiveness of the Heuristic Layered Representation. Layered floor plans are fundamentally a heuristic simplification of 3D spatial relationships and cannot fully represent complex 3D geometry. For example, placing plants with large overhanging leaves, tall ladders, or objects on multi-level shelves requires reasoning about vertical geometry and spatial occupancy that cannot be captured by the proposed representation. Consequently, 2D verification cannot guarantee 3D physical plausibility. This limitation is also reflected in the experimental results, where 28% of on-host objects topple under gravity, and only 51% satisfy the physical stability criterion.

  2. Insufficient Validation of Applicability to Embodied AI. The experimental evidence does not sufficiently support the paper's claim that the generated scenes can be directly used as mobile-manipulation environments. In the reported experiments, a mobile Franka Panda successfully completes only 28 out of 114 pick-and-place tasks. Although some failures may originate from the robot's planning or control limitations, the low success rate raises concerns about the practical usability of the generated environments.

  3. Insufficient Ablations to Attribute Improvements in 3D Physical Plausibility. Although the paper includes ablations of the layered representation and deterministic checker, the substantial reduction in 3D collision rates appears to come primarily from the subsequent mesh repair stage rather than the proposed planning framework. According to Table 9, the collision rate remains nearly unchanged between the initial plan (32.9%) and the full planner before mesh repair (33.1%), while it drops to 10.9% after mesh repair. Therefore, the experiments do not sufficiently establish how much the proposed layered representation and checker-guided planning independently contribute to the final 3D scene quality.

  4. Missing Comparisons with General-Purpose Coding Agents. The paper primarily compares Strata3D against specialized agentic scene generation frameworks, while overlooking strong general-purpose coding agents, such as GPT with Codex. Given that Strata3D is essentially an agentic workflow, comparisons with these baselines under comparable computational budgets are important to establish the effectiveness and generalizability of the proposed framework.

Questions for authors

How to deal with non-uniform scaling of retrieved 3D assets? Does it introduce geometric distortions or affect their physical stability and overall aesthetic?

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

Strata3D employs a layout-first approach to generate dense indoor scenes: First, a language model proposes a planar graph where each object is assigned to one of five layers. Then, a deterministic checker tests the boundaries of objects of actual width, layer overlap, supports, doorway clearance, and accessibility, returning each violation in textual form. The language model modifies the planar graph based on these checks and implements safeguards to ensure scene asset richness, including restrictions on deleting unnamed objects and protection of declared functional relationships. This paper validates the method on the SceneEval-100 dataset and performs end-to-end comparison tests with SceneCode and SceneSmith on 24 dense reservation cuees, using the same language model and asset library, and scoring the final scene of each system using the same evaluation criteria.

Critical strengths

1.This paper contributes a hierarchical scene design architecture, addressing a novel research question that helps solve the problem of how scene design can be used downstream, such as many issues in robot manipulation.

2.The paper is clearly structured and easy to understand.

3.The experiments are thorough and the comparisons are fair. The dense room comparison test is well-controlled: two agent baseline models are run end-to-end using the same language model and resource library, scored in their respective final scenes, and the number of placed objects is similar, making the end-to-end comparison test fairer. The ablation test results are informative and honestly report components that do not produce measurable differences. The paper also frankly points out legacy issues such as objects tipping over under gravity and the removal of a large number of required objects during room modification.

Critical weaknesses
  1. Experiments show that the complete process is useful, but its major improvements over earlier systems require more specific explanation to avoid the misconception that this paper is incremental. This paper discusses previous agent-based and support-based systems, including SceneSmith and SAGE, and explains differences such as continuous hierarchical footprints, human width accessibility, and the preservation of checked scenarios during asset placement. Rather than relying primarily on overall performance improvements to explain its contribution, it clarifies which problems these differences enable the method to handle that earlier systems could not.

  2. The layout obtained through inspection may still not be applicable to different robots. The initial inspection considered a person with a width of 0.55 meters, while the optional robot mode adds accessibility constraints; the robot demonstration used the mobile Franka robot. Page 9 reports that many unreachable objects were too high or too low. Therefore, it is unclear how many changes would be needed for other robots (e.g., Split ALOHA) and whether the same layout could be preserved. This is crucial when performing different robot tasks using the generated scene.

Questions for authors
  1. It would be helpful if the author could provide a concrete example illustrating how an earlier, proven planning system (such as SceneSmith or SAGE) performed poorly in handling room layouts, while a layered representation or inspector handled them well. This would allow us to see the major improvements in a specific instance, rather than just in a list of layer types.

  2. If the Franka mobile robot were replaced with Split ALOHA or another robot, which body size, working height, and reach rules would need to be changed? Would the layout need to be regenerated? How do the current results affect this adjustment?

  3. Can the assembly place objects on the internal shelves of a multi-tiered cabinet, rather than primarily on the top surface? Section 3.4 describes testing at various support heights; however, the cabinet in Figure 3.study is empty inside. Should this approach be shown in the context of multi-tiered cabinets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.