Hugging Face's logo

Datasets:
airoa-org
/
yubi-real2sim-assets

Modalities:
3D
Languages:
English
Size:
< 1K
ArXiv:
Tags:
robotics
simulation
3d-gaussian-splatting
mujoco
real2sim
airoa
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1

Access the Real2Sim assets on Hugging Face

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

The 3DGS scenes are adaptations of third-party captures licensed under CC BY 4.0; keep the attribution in each scene's ATTRIBUTION.txt when you use or share them. The object and FR3 meshes are licensed under the Apache License 2.0. To access the assets, log in to Hugging Face, review the terms and click below. Access requests are reviewed manually.

Log in or Sign Up to review the conditions and access this dataset content.

Real2Sim assets

Real2Sim evaluates openpi pi0.5 policies on a Franka FR3 arm with the bimanual YUBI gripper in the MuJoCo simulator. A 3D Gaussian Splatting (3DGS) capture supplies the photorealistic background. This repository holds the inputs the Real2Sim code repository does not track: four 3DGS background scenes and the object and FR3 meshes. The policies are in airoa-org/yubi-real2sim-policies.

Together with the policies they cover the 24 release profiles in config/release/fr3/{flat,duo}/<room>_<task>_<token>.yaml: two FR3 rigs (flat, duo), two rooms (reading_room, a_room) and three tasks (box_pen, box_tape, cup).

Layout 

README.md
scenes/
  flat_reading_room/   flat_a_room/   duo_reading_room/   duo_a_room/
assets/
  trellis_cup_plate/   trellis_office/   fr3_visual/   fr3_collision/

Scenes

3DGS rooms registered to an FR3 rig. Each folder holds the point cloud (PLY), background.json (the registration), table_collision/ (the simulator's table box), ATTRIBUTION.txt and LICENSE-APACHE-2.0. The Flat and Duo folders of a room share the same PLY file.

Folder Contents Size Install to Used by
scenes/flat_reading_room/ "Reading Room" registered to the Flat rig 1.35 GB {inputs_root}/scenes/flat_reading_room fr3/flat/reading_room_* (6)
scenes/flat_a_room/ "A Room" registered to the Flat rig 0.71 GB {inputs_root}/scenes/flat_a_room fr3/flat/a_room_* (6)
scenes/duo_reading_room/ "Reading Room" registered to the Duo rig 1.35 GB {inputs_root}/scenes/duo_reading_room fr3/duo/reading_room_* (6)
scenes/duo_a_room/ "A Room" registered to the Duo rig 0.71 GB {inputs_root}/scenes/duo_a_room fr3/duo/a_room_* (6)

Assets

Meshes that the simulator loads from the repository's robosuite_env/assets/. Every profile uses all four.

Folder Contents Size Install to
assets/trellis_cup_plate/ cup and plate: textured visual mesh (OBJ + PNG) and convex collision parts 475 MB {repo_root}/robosuite_env/assets/trellis_cup_plate
assets/trellis_office/ box, packing tape, pen and LAN cable: textured visual mesh (OBJ + PNG) and convex collision parts 353 MB {repo_root}/robosuite_env/assets/trellis_office
assets/fr3_visual/ FR3 arm visual meshes: 28 binary STL files and manifest.json (colours, triangle counts) 12 MB {repo_root}/robosuite_env/assets/fr3_visual
assets/fr3_collision/ FR3 arm collision meshes: 8 STL files, unmodified 0.1 MB {repo_root}/robosuite_env/assets/fr3_collision

{repo_root} is your checkout of the Real2Sim repository, and {inputs_root} is $REAL2SIM_INPUTS_ROOT (default {repo_root}/local-inputs).

Download

The Real2Sim installer downloads only what the given profiles need, verifies the digest of each installed folder and writes REAL2SIM_INPUTS_ROOT to .env:

.venv/bin/python scripts/download_public_assets.py \
  --profile config/release/fr3/flat/reading_room_cup_cup1trim.yaml

Manual download:

# one scene
hf download airoa-org/yubi-real2sim-assets --repo-type dataset --include "scenes/flat_reading_room/*" --local-dir yubi-real2sim-assets

# everything
hf download airoa-org/yubi-real2sim-assets --repo-type dataset --local-dir yubi-real2sim-assets

The repository is gated: log in with hf auth login and accept the terms on this page first. The launcher refuses symlinked runtime files, so copy real files into place and do not edit the installed trees.

Licences and attribution

Each folder carries its own licence files, which take precedence over this summary.

Scenes: CC BY 4.0

Changes from the originals: each room starts from the PLY obtained on 2026-11-04 with the scene page's Download button. The spherical-harmonic fields are stored at degree 3 (zero-padded for "Reading Room"; for "A Room", the downloaded level-of-detail-0 PLY with bands 1-3 set to zero), the vertex properties are reordered, and a registration transform (background.json) and the simulator's table collision box (table_collision/) were added. The Gaussians themselves are not edited. ATTRIBUTION.txt in each folder lists the changes. The creators do not endorse this adaptation or its use. CC BY 4.0 does not license other rights, such as publicity, privacy or moral rights, or rights in works visible in a capture.

Third-party works visible in a scene (for example the framed photographs on the Reading Room wall) belong to their own rights holders and are not covered by CC BY 4.0 or by any other licence stated in this repository. To ask for such a work to be removed, contact ext-oss@airoa.org.

The table box and background.json were made by the AI Robot Association (AIRoA) and are not part of the original captures. They are licensed under the Apache License 2.0; each scene folder includes the full text as LICENSE-APACHE-2.0.

TRELLIS object meshes: Apache-2.0, AIRoA

assets/trellis_cup_plate/ and assets/trellis_office/ were generated by AIRoA from its own photographs with TRELLIS.2 (Microsoft, MIT), split into convex collision parts with CoACD (Xinyue Wei and Minghua Liu, MIT) and converted with trimesh (MIT). They are distributed by AIRoA under the Apache License 2.0 (see LICENSE, NOTICE and README.txt in each folder). If you use the collision parts, please cite Wei et al., "Approximate Convex Decomposition for 3D Meshes with Collision-Aware Concavity and Tree Search", ACM Transactions on Graphics 41(4), SIGGRAPH 2022.

TRELLIS.2's image-to-3D pipeline also ran DINOv3 (Meta, DINOv3 License) for image features, RMBG-2.0 (BRIA AI, CC BY-NC 4.0) for background removal, and nvdiffrast (NVIDIA, non-commercial licence) for texture baking. None of their code or weights is included here; the meshes and textures are outputs of that pipeline.

FR3 meshes: Apache-2.0, Franka Robotics GmbH

assets/fr3_visual/ and assets/fr3_collision/ come from franka_description at commit 2fd6aaa8c5e3bbfe857b8fb19978694cdc41e6ea (tag 2.7.1), Copyright 2023-2024 Franka Robotics GmbH, Apache License 2.0. Each folder includes the upstream LICENSE and NOTICE unchanged.

  • fr3_collision: the collision STL files, unmodified.
  • fr3_visual: modified. The upstream COLLADA (.dae) visual meshes were converted to one binary STL file per material with the node transforms baked in; colours are in manifest.json. README.txt states the changes.

Rendering

Every release profile renders the 3DGS background through the Inria Gaussian-Splatting rasterizer, whose licence allows non-commercial research and evaluation use only. Running the full pipeline is therefore currently limited to non-commercial use, whatever the licences of the files in this repository.

Citation 

@article{nishizawa2026yubi,
  title={YUBI-Real2Sim: 3DGS Scene Verification for Reliable Policy Evaluation},
  author={Nishizawa, Hiroki and Ohkawa, Takehiko and Scianca, Nicola and Matsushima, Tatsuya and Morishima, Shigeo and Ota, Kei},
  journal={arXiv preprint arXiv:2610.17521},
  year={2026}
}
Downloads last month
45
Total file size:
4.96 GB

Paper for airoa-org/yubi-real2sim-assets