cs.CVMar 12, 2026
SaveMANSION: Multi-floor lANguage-to-3D Scene generatIOn for loNg-horizon tasks
Organizations: Tsinghua University · AgiBot · McGill University, MILA - Quebec AI Institute
Abstract
Real-world robotic tasks are long-horizon and often span multiple floors, demanding rich spatial reasoning. However, existing embodied benchmarks are largely confined to single-floor in-house environments, failing to reflect the complexity of real-world tasks. We introduce MANSION, the first language-driven framework for generating building-scale, multi-floor 3D environments. Being aware of vertical structural constraints, MANSION generates realistic, navigable whole-building structures with diverse, human-friendly scenes, enabling the development and evaluation of cross-floor long-horizon tasks. Building on this framework, we release MansionWorld, a dataset of over 1,000 diverse buildings ranging from hospitals to offices, alongside a Task-Semantic Scene Editing Agent that customizes these environments using open-vocabulary commands to meet specific user needs. Benchmarking reveals that state-of-the-art agents degrade sharply in our settings, establishing MANSION as a critical testbed for the next generation of spatial reasoning and planning.
Figures & tables
| Method | Type | Bdry. | Topo. | Vert. core | Room type |
| Graph2Plan [ 16 ] | model-based | ✓ | ✓ | resident | |
| HouseDiffusion [ 37 ] | model-based | ✓ | resident | ||
| GSDiff [ 17 ] | model-based | ✓ | resident | ||
| ProcTHOR [ 10 ] | rule-based | ✓ | resident | ||
| Holodeck [ 48 ] | LLM | open-vocab | |||
| AnyHome [ 14 ] | LLM+model | ✓ | resident |
Table 1: Comparison of floorplan generation methods. Bdry. / Topo. / Vert. denote Boundary/Topology/Vertical structure. Boundary/Topology: controllable conditioning at test time. Vertical structure indicates cross-floor aligned cores (walls/rooms/regions) that persist across floors. Room type: resident vs. open-vocab.
Figure 2: Overview of the MANSION framework: a multi-agent-driven pipeline for generating multi-story 3D buildings from natural language. The process includes: (A) Whole Building Planning, (B) Per-Floor Planning, (C) Floorplan Synthesis, and (D) Scene Instantiation.
Figure 3: MansionWorld statistics: functional composition and floor-area distributions across different floor counts.
Figure 4: The “Check-and-Provision” workflow of our Task-Semantic Scene Editing Agent. The agent first decomposes a high-level instruction (“bring a snack and a drink to the sofa”) into preconditions. It then sequentially performs a (a) Path Connectivity Check, an (b) Object Availability Check, and an (c) Object Provisioning & Scene Edit to ensure the task is executable before generation.
| Method | Micro-IoU | Macro-IoU |
| Obj-GAN [ 27 ] | 10.68 | 8.44 |
| CogView [ 11 ] | 13.30 | 11.43 |
| Imagen [ 36 ] | 12.17 | 14.96 |
| T2D | 54.34 | 53.30 |
| CHD (moonshot) | 60.09 | 56.09 |
| CHD (gemini-2.5-pro) | 76.34 | 72.24 |
Table 2: IoU scores under different configurations on T2D
| Method | Micro-IoU | Macro-IoU |
| CHD (gemini-2.5-pro) | 29.36 | 22.25 |
| CHD (8minus-MA) | 36.12 | 26.14 |
| CHD (MA) | 33.49 | 25.39 |
| Ours (gemini-2.5-pro) w/o hierarchical splitting | 45.65 | 42.42 |
| Ours (gemini-2.5-pro) | 63.56 | 61.65 |
| Ours (MA) | 76.74 | 76.64 |
Table 3: IoU scores under different configurations on Resplan-1k
| Method | Bedroom (4 4 m, rect.) | Classroom (8 8 m, rect.) | ||||||||||||
| #Obj | #OB | #CN | #Rch | Real. | Div. | Lay. | #Obj | #OB | #CN | #Rch | Real. | Div. | Lay. | |
| LayoutGPT | 8.3 (0.0) | 0.1 | 2.6 | 95.3 | 3.9 | 33.3 | 7.8 | 14.7 (0.0) | 0.0 | 0.5 | 98.6 | 3.9 | 3.9 | 13.7 |
| Holodeck | 17.5 (7.1) | 0.0 | 0.0 | 88.7 | 41.2 | 9.8 | 39.2 | 64.4 (25.2) | 0.0 | 0.0 | 80.0 | 0.0 | 51.0 | 5.9 |
| Ours | 22.6 (9.2) | 0.0 | 0.0 | 100.0 | 54.9 | 56.9 | 52.9 | 57.3 (19.8) | 0.0 | 0.0 | 100.0 | 96.1 | 45.1 | 80.4 |
| Method | Restaurant (polygon) | Library (polygon) | ||||||||||||
| #Obj | #OB | #CN | #Rch | Real. | Div. | Lay. | #Obj | #OB | #CN | #Rch | Real. | Div. | Lay. | |
Table 4: object placement quantitative comparison. We report the average number of placed objects (#Obj, with small items in parentheses), out-of-boundary objects (#OB), Layout-level collided object pairs (#CN), floor-object reachability (#Rch, %), and user-study preference scores (%) for Realism (Real.), Diversity (Div.), and Layout (Lay.).
Figure 5: object placement qualitative comparison.
| Method | Single fl. | Double fl. | Four fl. |
| COME [ 50 ] | 30 (50, 30) | 20 (50, 20) | 0 (0, 0) |
| BUMBLE [ 39 ] | 40 (50, 40) | 20 (30, 40) | 0 (0, 0) |
| BUMBLE w/ object type | 60 (90, 70) | 60 (80, 60) | 0 (0, 0) |
Table 5: Task success rates from 10 trials. Progress score is reported in the brackets in the format (Object retrieval success, Navigation success)
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Qualitative floorplan comparison with CHD.
Figure 7: Qualitative comparison between Holodeck and MANSION under high-level semantic building prompts.
Figure 8: Structural realism in MANSION. (a) Cross-floor load-bearing walls (blue) are preserved to maintain vertical consistency. (b) A villa with a protruding room and garage shows controllable per-floor outer contours.
Figure 9: Additional details of the MansionWorld ecosystem.
Figure 10: Qualitative examples of non-residential buildings in MansionWorld .
Figure 11: Qualitative examples of entertainment and residential buildings in MansionWorld .
Figure 12: Illustration of the single-floor, topology-driven pipeline. (a) Input room topology graph. (b) Cut-round construction and hierarchical splitting over the free region. (c) Final 3D scene instantiation in AI2-THOR after applying structure, objects, and lighting.
Figure 13: System Architecture of the Task-Semantic Scene Editing Agent. The system operates via a ReAct Controller (top) that iteratively plans and issues JSON tool requests. A Tool Invoker (middle) serves as an execution bridge, routing perception tasks to the fast Static Semantic State (bottom left) and action tasks to the On-Demand Physics Engine (bottom right). The dashed arrow highlights the Hybrid State Management mechanism, where physical simulation results are synchronized back to the static scene JSON to ensure consistency.
Figure 14: Sample screenshots from a task execution. The robot begins on the second floor, takes the elevator to the first floor to retrieve a cloth, and then returns to the sofa. Task query: I want to clean my sofa. Go get a cloth from the first floor and come back near the sofa.
| Environment | Prompt |
| Single-floor apartment | Find a box on the bed and bring it to the bathroom. |
| Double-floor office | Find a cellphone on a blue couch on the first floor and bring it to the round table on the second floor. |
| Four-floor office | Go to the third floor, find a laptop on the desk in the meeting room on the third floor, and take it to the restroom on the fourth floor. |
Table 6: Task Settings and Prompts
Figure 15: Single-floor apartment layout.
Figure 16: Two-floor office layout.
Figure 17: Four-floor building layout.
Figure 18: Panorama view of the service shaft room.
Figure 19: Failure case 1
Figure 21: Reachability visualization in a library scene.
Figure 22: Sample from the user study for the classroom scene category.
Figure 23: Prompt template for whole-building program planning.
Figure 24: Prompt template for single-floor topology generation.
Figure 25: Prompt template for LLM-guided seed box planning for cutting.