奔驰马克一点零

# GEAI-Benckmark-1.0 Curated benchmark package for `file_2` to `file_11` from the cleaned hi_scene benchmark set. The package is organized for scene inspection, asset review, prompt tracing, VLN task design, VLA task design, and packaged video review. ## What is included - 10 cleaned multi-room scenes under `scenes/file_2` to `scenes/file_11`. - Original `scene.json` prompts plus packaged `scene.xml`, textures, mesh assets, and per-scene asset manifests. - Final overview videos under `videos/` for fast visual screening. - Documentation for scene usage, prompt interpretation, and VLN/VLA task design under `docs/`. ## Quality and integrity snapshot - Final manual and scripted cleanup produced zero reported counts for front-to-wall orientation errors, out-of-bounds assets, below-ground assets, floating assets, scene hinge mismatches, and raw broken articulated assets across `file_2` to `file_11`. ## Scene index | Scene | Title | Rooms | Assets | Articulation candidates | Quality reward | Scene README | Video | | --- | --- | ---: | ---: | ---: | ---: | --- | --- | | `file_2` | Duplex Apartment with Dining and Study Support Chains | 7 | 50 | 14 | 0.9432 | [scene card](scenes/file_2/README.md) | [video](videos/file_2/scene_overview.mp4) | | `file_3` | Family Townhouse with Mudroom, Nook, and Home Office | 7 | 43 | 19 | 0.9770 | [scene card](scenes/file_3/README.md) | [video](videos/file_3/scene_overview.mp4) | | `file_4` | Compact City Flat with Reading Alcove | 6 | 36 | 13 | 0.9495 | [scene card](scenes/file_4/README.md) | [video](videos/file_4/scene_overview.mp4) | | `file_5` | Artist Live-Work Home with Gallery and Studio | 6 | 35 | 11 | 0.9723 | [scene card](scenes/file_5/README.md) | [video](videos/file_5/scene_overview.mp4) | | `file_6` | Maker Apartment with Workshop and Loft Bedroom | 6 | 36 | 19 | 0.9799 | [scene card](scenes/file_6/README.md) | [video](videos/file_6/scene_overview.mp4) | | `file_7` | Book-Filled Home with Tea Corner and Study Room | 6 | 41 | 16 | 0.9552 | [scene card](scenes/file_7/README.md) | [video](videos/file_7/scene_overview.mp4) | | `file_8` | Suburban Two-Level Home with Craft Nook | 8 | 55 | 23 | 0.9502 | [scene card](scenes/file_8/README.md) | [video](videos/file_8/scene_overview.mp4) | | `file_9` | Design Studio Apartment with Meeting and Sample Zones | 7 | 49 | 21 | 0.9432 | [scene card](scenes/file_9/README.md) | [video](videos/file_9/scene_overview.mp4) | | `file_10` | Chef-Focused Apartment with Pantry and Upper Sleeping Room | 7 | 10 | 5 | 0.9462 | [scene card](scenes/file_10/README.md) | [video](videos/file_10/scene_overview.mp4) | | `file_11` | Research Home Office Suite with Lab-Style Workspace | 6 | 50 | 18 | 0.9675 | [scene card](scenes/file_11/README.md) | [video](videos/file_11/scene_overview.mp4) | ## Recommended uses - `VLN`: multi-room routing, landmark grounding, vertical circulation, room-to-room localization, and long-horizon navigation evaluation. - `VLA`: tabletop relation chaining, support-surface reasoning, container interaction, articulation-aware manipulation, and workspace organization tasks. - `Scene QA`: checking geometry integrity, prompt-to-layout alignment, and articulated-object consistency. ## Repository structure ```text GEAI-Benckmark-1.0/ ├── scenes/ # file_2 to file_11 scene packages ├── videos/ # overview videos and index.html └── docs/ # scene index, task design guide, operation guide ``` ## Quick start ### 1. Read prompt, assets, and task ideas for a scene - Open `scenes/file_X/README.md` for prompt, room structure, asset summary, and VLN/VLA task suggestions. - Open `scenes/file_X/assets_hub/assets_manifest.json` for asset-level prompts, routes, and task profiles. - Open `videos/file_X/scene_overview.mp4` for a quick visual pass before task selection. ## Documentation - [Scene Index](docs/SCENE_INDEX.md) - [Task Design Guide](docs/TASK_DESIGN_GUIDE.md) - [Operation Guide](docs/OPERATION_GUIDE.md) - [Chinese Overview](README_CN.md) ## Notes - Prompts and asset-generation traces are preserved in English because the original scene and asset prompts were authored in English. - The package prioritizes scene fidelity for VLN/VLA benchmarking rather than minimizing storage footprint.