iARCS Synthetic Indoor Scenes

12,000 3D indoor scene layouts generated with iARCS: 4,000 bedrooms, 4,000 living rooms and 4,000 dining rooms.

• Paper: iARCS: Iterative Agentic RL for Controllable 3D Scene Generation
• Project page: https://saugat2002.github.io/iarcs/
• Code: https://github.com/thenaivekid/iARCS

Contents

Room	Scenes	Floor plans	Mean objects / scene
Bedroom	4,000	162	4.92
Living room	4,000	192	9.10
Dining room	4,000	177	13.66

Floor plans come from the 3D-FRONT test split used by MiDiffusion. Each floor plan is reused for several generated scenes.

Files

data/<room>/scenes.parquet       one row per scene
data/<room>/floor_plans.parquet  one row per floor plan
json/<room>/scenes.jsonl         same scenes as JSON Lines
json/<room>/floor_plans.json     same floor plans as JSON
json/<room>/categories.json      object category list
raw/<room>/results.pkl           original MiDiffusion output (needs the iARCS / MiDiffusion code to load)
raw/<room>/config.yaml           generation config

Fields

scenes

Field	Description
scene_index	0–3999
room_type	bedroom, livingroom or diningroom
floor_plan_id	3D-FRONT room id; join with floor_plans
num_objects	number of objects
objects	list of objects (below)

objects

Field	Description
category	object class, e.g. double_bed
translation	object centre [x, y, z] in metres; Y is up
half_extents	half the bounding-box size [x, y, z] in metres
angle	rotation around the Y axis, radians
jid	3D-FUTURE model id, retrieved as the same-category model closest in size

floor_plans

Field	Description
floor_plan_id	3D-FRONT room id
vertices	floor mesh vertices, centred on centroid (same frame as object translations)
faces	floor mesh triangles (vertex indices)
centroid	floor-plan centroid in the original 3D-FRONT frame

Usage

from datasets import load_dataset

scenes = load_dataset("Saugat20021/iARCS", "bedroom", split="train")
plans = load_dataset("Saugat20021/iARCS", "floor_plans_bedroom", split="train")

s = scenes[0]
print(s["floor_plan_id"], s["num_objects"])
for o in s["objects"]:
    print(o["category"], o["translation"], o["jid"])

To rebuild textured 3D scenes, place each 3D-FUTURE model jid at translation, rotate it by angle around Y, and scale it to half_extents. 3D-FRONT and 3D-FUTURE must be obtained separately under their own licenses. No CAD assets are included here.

Notes

• A few living-room (4) and dining-room (2) scenes have no objects.
• The data is synthetic and contains no personal information. It inherits the coverage and furnishing style of 3D-FRONT.

License

CC BY-NC 4.0. The layouts come from a model trained on 3D-FRONT and refer to 3D-FUTURE assets, so the non-commercial research terms of those datasets also apply.

Citation

@article{adhikari2026iarcs,
  title   = {iARCS: Iterative Agentic RL for Controllable 3D Scene Generation},
  author  = {Adhikari, Saugat and Neupane, Ashok Prasad and Paudel, Pramish
             and Chhatkuli, Ajad and Paudel, Danda Pani},
  journal = {arXiv preprint arXiv:2608.06161},
  year    = {2026}
}