spatialgym/SpatialGym_Figural

SpatialGym Figural Objects

Object packs for the figural (tabletop) tasks of SpatialGym — Pose Matching, Mental Rotation and Active Closure. Each pack is a set of self-contained USD objects baked from one curated catalog of source meshes (RoboTwin 2.0 rigid objects + YCB google_16k scans), plus a manifest.json the environments read; mental_rotation also holds 323 props from NVIDIA SimReady, and pose_matching 52 of them at real size. The other six figural packs (polycubes, CLEVR objects, spatial-relation solids, assembly, net folding) are generated locally by assets/download.sh and are not part of this repository.

Overview

Pack	Objects	Sources	What the environment gets	Download	Unpacked
pose_matching	461	RoboTwin 352 · YCB 57 · NVIDIA SimReady 52	rest pose, metric size, PBR textures; object sets and shipped question lists	1.6 GB	1.8 GB
mental_rotation	459	RoboTwin 118 · YCB 18 · NVIDIA SimReady 323	longest axis 1.7 m; per object its body axis and which tasks ask it; shipped question lists	3.8 GB	7.4 GB
active_closure	307	RoboTwin + YCB	voxelised body (cube PointInstancer) + 4-way options	4 MB	20 MB

Files

SpatialGym_Figural/
├── README.md
├── pose_matching.tar.gz      # unpacks to pose_matching/{manifest.json, objects/<id>.usd, objects/textures/, simready/, sets.json, questions/}
├── mental_rotation.tar.gz    # unpacks to mental_rotation/{manifest.json, objects/<id>.usd, objects/textures/, simready/, questions/}
└── active_closure.tar.gz     # unpacks to active_closure/{manifest.json, <id>.usd}

SpatialGym_IsaacLab/assets/download.sh figural in the SpatialGym repository fetches the three packs into assets/data/figural/, builds the six generated packs next to them and checks every pack against the version its code expects.

Packs

pose_matching — objects/<id>.usd, one per object, in its canonical rest pose at real-world size; textures externalised under objects/textures/. 409 come from the catalog above (two near duplicates left out) and 52 are the NVIDIA SimReady props the mental_rotation manifest keeps for Pose Matching (tasks.pose_matching): each keeps its SimReady files under simready/<asset>/, and objects/<asset>.usd takes the mental_rotation wrapper with its scale removed (the props are authored in metres) and its bounding box centred.

_{Pose Matching: one task frame per scene, difficulty 1–3.}

Key	Meaning
id, usd	Object id and USD path relative to the pack
label_l1, label_l2	Coarse / fine object labels shown to the agent
footprint_cm, height_cm	Rest-pose footprint [w, l] and height
yaw_symmetric	true ⇒ the object looks the same at every yaw and is excluded from the manipulable pool
source, orig_id, variant	Provenance: source dataset, original id, texture / shape variant

sets.json — the 1152 object sets Pose Matching's questions name, {"sets": [...]}: each set has an id (s<scene>t<task>_<n>), its scene and task_rungs, the objects on the table (objects, the first num_manipulables of them the ones a question moves) and their usd per scene slot. questions/level1.jsonl, level2.jsonl, level3.jsonl — the shipped questions, 80 a level, one per line {manifest, id, difficulty, seed}, id naming a set. Level 1: 4 objects on the table, 1 moved 1–2 steps from its goal; level 2: 5 objects, 2 moved 3–4 steps; level 3: 6 objects, 4 moved 7–8 steps.

mental_rotation — 459 objects in one pool, every one baked so its longest axis is 1.7 m: 136 from the catalog above, re-baked upright (rest rotation zeroed), and 323 NVIDIA SimReady props, each keeping its SimReady files under simready/<asset>/ (textures at 512 px) and wrapped in objects/<asset>.usd, which centres it, scales it and switches its physics off. Each object records which tasks ask it (tasks), so one manifest serves Mental Rotation and Pose Matching, whose generator bakes the kept SimReady props at real size. No per-object difficulty: object complexity is not an axis of this task.

_{Mental Rotation: top row the locally generated polycube pool (not in this repository), bottom row some of this pack's curated objects.}

Key	Meaning
id, usd	Object id and USD path relative to the pack
scene	Scene rung: -1 for every object (an unrated pool)
label_l1, label_l2	Coarse / fine object labels
source, orig_id, variant	Provenance: robotwin, ycb or simready; the id in that source; its texture / shape variant
dims_m	Bounding-box extent [x, y, z] in metres of the baked USD (longest edge 1.7)
real_dims_m	The same box at the object's real size, metres
axis_kind	How a person reads the object: up (it stands one way: a cup, a kettle), front (it lies along a long axis with a front or working end: a shoe, a hammer, a toy car), none (no axis people agree on)
body_axis	That axis as a unit vector [x, y, z] in the object's frame, pointing up or to the front / working end; null when axis_kind is none. Mostly a signed axis of the frame; for a few objects stored a few degrees off it, the measured direction of the handle
body_axis_origin	A point [x, y, z] its line runs through (along-axis part zero), checked to pass through the middle of the object; null with no axis
axis_why	One line: why this axis, or why none
tasks	One {keep, why} per subset a task asks (below); a task reads its own key only

tasks key	Objects kept	What it is
mental_rotation	392	Asked by Mental Rotation. Left out: 43 that look the same after a turn, and 24 unfit for every task (duplicates, several things stuck together, poor quality)
mental_rotation_axis	144	Asked by Mental Rotation with the body-axis arrow drawn: kept by mental_rotation and has a body_axis
pose_matching	186	Used by Pose Matching at real size. Left out: objects too tall, too large, too small, flat or thin for its table, ones whose facing cannot be told, and the 24 unfit for every task. A kept SimReady entry adds Pose Matching's own label_l1 / label_l2

import json
objects = json.load(open("mental_rotation/manifest.json"))["objects"]
with_arrow = [o for o in objects if o["tasks"]["mental_rotation_axis"]["keep"]]  # 144

questions/level1.jsonl, questions/level2.jsonl — Mental Rotation's shipped questions, 40 a level, one per line {manifest, id, difficulty, seed}. Drawn by spatial_gym.core.questions.Gen from the test split (manifest_test.json, which assets/download.sh writes from the repository's committed 308 / 151 train / test split), among the 129 test objects tasks.mental_rotation keeps, order shuffled with seed 1101, with the arrow off. Level 1 is task rung 0 (at least one turn), level 2 task rung 2 (at least three turns); both levels ask the same 40 objects. SpatialGym_IsaacLab/configs/eval_mental_rotation.yaml plays them; the oracle solves all 80.

active_closure — one <id>.usd per object holding a PointInstancer of cubes (the voxelised body, 348–30 221 voxels), floated in a booth for the closure question.

_{Active Closure: one object per column, visibility rungs 1–3 down the column.}

Key	Meaning
id, usd, category	Object id, USD path, category label
num_voxels, size_m	Voxel count and longest-axis size of the voxelised body
options, answer_idx	A 4-way option set and the correct index. The environment draws its own option set at load; these are kept for older code

All figures are frames the tasks themselves render (oracle runs, the environments' own camera and lighting).

Sources and licensing

The objects are derived from three source datasets:

• RoboTwin 2.0 object library — code and dataset released under the MIT license. Chen et al., RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation, arXiv:2506.18088, 2025.
• YCB Object and Model Set (google_16k scans) — Calli et al., The YCB Object and Model Set: Towards Common Benchmarks for Manipulation Research, ICAR 2015. Model files are distributed under CC BY 4.0.
• NVIDIA SimReady assets (mental_rotation and pose_matching, entries with "source": "simready") — nvidia/simready-assets at revision 4763fbfaa4f8e97658e68b57d01b4c86659e08b5, licensed under CC BY 4.0. Changed for SpatialGym: textures reduced to at most 512 px, only the files each object's USD layers reach are kept, and each object is wrapped in a USD that centres it, switches off its physics and, in mental_rotation, scales it to 1.7 m (pose_matching keeps its real size); the original files are otherwise unedited. Each object's licence file ships beside it, and mental_rotation/simready/ATTRIBUTION.md repeats this notice. No endorsement by NVIDIA is implied.

The baked USD packs in this repository are provided for research use; please cite SpatialGym and the source datasets above when you use them.