Real2Sim Lab Room
A procedurally generated OpenUSD model of one real laboratory room (10 x 7.5 m inside) with its furniture, built
for physics simulation in NVIDIA Isaac Sim 5.1 / Isaac Lab. Every file is written by the Python generators in
source/ from YAML parameter files; nothing was scanned or modelled by hand, and there are no external asset
dependencies (the few textures are procedural and included).
中文摘要:本数据集是一间真实实验室(内净尺寸 10 × 7.5 m)的程序化 USD 数字孪生,面向 NVIDIA Isaac Sim 5.1 / Isaac Lab 的物理仿真。内容包括:带向室内开启的铰接双开门的房间壳体;8 个带抽屉和键盘托(移动副)的 L 形工位;8 把带回转座面和万向脚轮的 办公椅(articulation);4 张实验桌;1 个钢制货架;一台可操作的台式示波器(4 个按键、6 个旋钮、BNC 接口)和 11 件桌面道具; 关节、抓取帧与布局的元数据 JSON;用于导航的占据图、净空图和高度图;全部生成脚本、检查脚本和 YAML 配置。平面尺寸来自平面图和 家具图纸,层高、吊顶、窗、门的细部以及全部质量和摩擦参数是估计值,详见下文"已知局限"。不包含任何机器人模型或策略。
The scene_v1look_* renders show an earlier static version of the scene: the pedestal drawers have no bar
handles there. No finished render of the current furniture exists; the current drawers with handles are visible
only in the collider debug renders (previews/scene_v2_*.jpg). See previews/README.md.
Contents
| Path | What |
|---|---|
scenes/scene.usd (.usda) |
production scene: room + 21 furniture items. 2718 prims, 640 colliders, 130 rigid bodies, 122 joints, 8 articulations |
scenes/scene_manip.usd (.usda), scene_manip.meta.json |
scene.usd as a sublayer plus the oscilloscope and 11 props on table_2, a workstation desk and a rack deck. 3280 prims, 719 colliders, 152 rigid bodies, 133 joints, 9 articulations |
scenes/room.usd (.usda), room.meta.json |
room shell alone (floor, walls, columns, carpet, windows, ceiling, lights, door) |
scenes/room_latched.usd (.usda), room_latched.meta.json |
the same room with both door hinges limited to 0..0. No composed scene references it |
scenes/assets/*.usd(a) + *.meta.json |
table_120, shelf, drawer_cabinet, workstation, workstation_mirrored, chair, oscilloscope |
scenes/assets/props/ |
scope_probe, bnc_plug, multimeter, pcb_board, screwdriver, parts_box, storage_bin, notebook, mug, cube_4cm, ball_5cm, index.json |
scenes/textures/, scenes/assets/**/textures/ |
procedural textures (carpet, blinds, oscilloscope panel, multimeter LCD) |
maps/ |
occupancy maps (ROS map_server PGM + YAML, PNG, NPY), clearance maps, collider height map of scene.usd |
source/ |
generators, local checkers, YAML configs, requirements.txt; optional Isaac-side test scripts in source/isaac_tests/ |
validation/ |
result JSON files behind the numbers quoted below |
docs/ |
technical report (Chinese, public edition, MD + PDF), PROPS.md, CONVENTIONS.md |
previews/ |
renders used on this card |
MANIFEST.json, THIRD_PARTY.md |
file list with sizes and sha256; third-party notes |
Each .usd (binary crate) has a .usda (text) twin with the same content.
Loading
Open scenes/scene.usd (or scenes/scene_manip.usd) in Isaac Sim 5.1. All references are relative; keep the
scenes/ folder together. Stage: metersPerUnit = 1, upAxis = Z, default prim /World, which contains
/World/PhysicsScene (gravity -Z, 9.81), /World/Room, /World/Furniture/<name> and, in the manipulation
scene, /World/Props/<name>.
from isaacsim import SimulationApp
app = SimulationApp({"headless": True})
import omni.usd
from isaacsim.core.api import SimulationContext
omni.usd.get_context().open_stage("scenes/scene.usd")
sim = SimulationContext(physics_dt=1 / 200, rendering_dt=1 / 200, stage_units_in_meters=1.0,
device="cpu", backend="numpy")
sim.reset()
for _ in range(1000):
sim.step(render=False)
Settings used in the validation runs:
- Physics step 200 Hz (dt 0.005 s); every test was also run at 120 Hz.
- CPU PhysX is recommended for a single scene. As authored,
scene.usdsteps in 0.86 ms (CPU, 200 Hz) against 2.8 ms on the GPU pipeline; once the chairs sleep, 0.28 ms on the CPU. On the GPU pipeline PhysX contact reports were empty and bodies do not sleep. - Solver iterations: the chair articulations carry
solverPositionIterationCount = 32,solverVelocityIterationCount = 1in the USD; nothing else overrides the simulator defaults. Lowering the chairs to 4 position iterations is faster on the GPU but increases the rest drift (0.7-1.0 mm in 5 s on the CPU). - The chairs start 1 mm above the floor and the props 1 mm above their supports, and settle in the first steps.
- Rendering: the lights are three
UsdLux.RectLights, tuned for RTX Real-Time with auto exposure off.
Isaac Lab: scene.usd has been loaded as the terrain USD of an Isaac Lab scene. Its /World/PhysicsScene then
sits next to Isaac Lab's own /physicsScene; stepping and gravity were normal in those runs. If the default prim is referenced under
another prim, replace the leading /World of every path given in the meta files.
Coordinate frame and conventions
- Units metres;
+Zup. World origin: floor-level corner between the front wall (y = 0, the wall with the door) and the left wall (x = 0), on the inner wall faces.+Xalong the 10 m side,+Ytowards the back wall (y = 7.5). - Asset frame: origin at the centre of the footprint on the floor, local
-Y= back (wall side),+Y= front / access side. Chair: origin on the swivel axis at floor level,+Y= direction of the sitter's knees. - Each file:
/<Name>(default prim) withvisuals,collisions,Looks; movable parts are separate Xforms with their ownvisuals/collisions, joints underjoints. - Colliders are boxes, convex meshes and spheres (chair casters) with
UsdPhysics.CollisionAPIandpurpose = guide, so they are invisible in renders. Visual meshes carry no collision. There are no triangle mesh colliders. - Materials are
UsdPreviewSurface. PhysX-specific attributes are authored as raw attributes (physxArticulation:*,physxJoint:*,physxRigidBody:*) because the files were written with plain usd-core. - Friction (static / dynamic): floor 0.8 / 0.7, carpet 1.0 / 0.9, furniture 0.6 / 0.5, caster wheels 0.8 / 0.8. These are assumed values.
More detail: docs/CONVENTIONS.md, docs/TECH_REPORT.md section 3.
Articulated parts
"Friction" of every joint below is a velocity drive with target 0 and a force / torque cap (a capped damper),
not PhysX jointFriction and not true static friction. Limits are in USD units (m, degrees); the PhysX tensor API
and Isaac Lab report revolute joints in radians.
| Part | Count | Joint prim (in scene.usd) |
Type, limits | Parameters | Metadata |
|---|---|---|---|---|---|
| Door leaves | 2 | /World/Room/door/joints/hinge_left, hinge_right |
revolute about Z; left 0..100 deg, right -100..0 deg, 0 = closed, opens into the room | 28 kg per leaf (estimated), torque cap 4 N m; no latch, handles do not turn | scenes/room.meta.json: hinge positions, limits, handle grasp points, approach directions |
| Pedestal drawers | 24 (3 per workstation) | /World/Furniture/ws_{left,right}_{1..4}/pedestal/joints/drawer_{top,mid,bottom}_slide |
prismatic, 0..0.393 m | 1.5 / 1.5 / 2.5 kg, friction 6 / 6 / 7 N | scenes/assets/drawer_cabinet.meta.json, workstation.meta.json (drawers, handle_grasp) |
| Keyboard trays | 8 | /World/Furniture/ws_*/joints/keyboard_tray_slide |
prismatic, 0..0.2 m, diagonal towards the chair | 1.5 kg, friction 4 N | scenes/assets/workstation.meta.json (keyboard_tray) |
| Chairs | 8 articulations, 12 links / 11 DoF each, floating base | /World/Furniture/chair_{left,right}_{1..4}/joints/swivel, caster_{0..4}_swivel, wheel_{0..4}_axle |
revolute, unlimited | 12.0 kg total; swivel torque cap 0.30 N m; wheel axle cap 0.0193 N m (about 3.5 N rolling resistance) | scenes/assets/chair.meta.json |
Oscilloscope (scene_manip.usd only) |
1 articulation, 10 joints | /World/Props/oscilloscope/joints/{power,autoset,single,run_stop}_button_slide, {horizontal_scale,ch1_scale,ch2_scale,ch1_position,ch2_position,trigger_level}_knob_turn |
keys prismatic 0..3 mm with spring return; knobs revolute, limited (0..288, 0..270, -135..135 deg) | keys about 2 N; knobs 0.03-0.04 N m | scenes/assets/oscilloscope.meta.json, scenes/scene_manip.meta.json; docs/PROPS.md |
BNC attach joint (scene_manip.usd only) |
1 | /World/Props/bnc_plug_attach |
fixed joint, authored disabled | enabled at run time by the helper in source/isaac_tests/props_bnc_attach.py |
scenes/scene_manip.meta.json (bnc_attach) |
The workstation metas also give the recommended chair pose and knee-space boxes; table_120.meta.json and
shelf.meta.json give the usable top rectangle and every rack deck with its height and clearance;
scene_manip.meta.json gives the world start pose of every prop, supports, targets and suggested stand points.
source/scripts/r2s/instrument_state.py turns oscilloscope joint positions into a logical instrument state
(power, run / stop, time base, V/div, connected channel).
Validation summary
Two kinds of checks were made. Result files are in validation/; tables and method are in
docs/TECH_REPORT.md (sections 6 and 7) and docs/PROPS.md.
Local geometry checks (usd-core only, reproducible from source/). Re-run for this package: room 2458 / 2458,
latched room 2458 / 2458, scene "21 furniture items, 0 problems", office assets 4139 / 4139, storage assets all
passed, props 2076 / 2076, manipulation scene 132 / 132. Regenerating from source/ reproduces all 73 files under
scenes/ byte for byte.
PhysX runs in Isaac Sim 5.1 (headless). The whole-scene runs used copies whose md5 sums equal those of the
files under scenes/. The part-level tests (door, drawers, chairs, props) were run on staged copies during
development; that those copies were byte-identical to the final files was not recorded.
| Test | Result | Limits of the test |
|---|---|---|
Load scene.usd |
0 unresolved references, 0 physics warnings; 130 dynamic bodies, 222 static shapes, 8 articulations | 4 configurations: CPU / GPU PhysX at 200 / 120 Hz |
| Settle 5 s | chairs drift 0.004-0.11 mm (CPU), 0.016-0.048 mm (GPU); tilt below 0.0002 deg | chairs only; on the GPU pipeline bodies keep about 0.1 mm/s residual velocity |
| Initial penetration | 0 penetrating collider pairs (separating-axis test on the USD); 0 static overlaps deeper than 1 mm (PhysX) | smallest dynamic-static gap 6.0 mm (drawer front to back wall), drawer to carcass 2.0 mm |
| Support surfaces | 101 probe cubes (5 cm) on tables, desks, rack decks, carpet, floor: z error below 0.001 mm, 0 failures | cubes only, no heavier loads |
| Room closure | rays from inside leak only through the modelled door gaps (4 mm between the leaves, 5 mm under them) | windows are solid colliders |
| Occupancy / reachability | free area 52.7 m2 with chairs; one connected component; a 0.3 m radius robot reaches 38.7 m2; the corner right of the door is not reachable | geometric analysis on a 1 cm grid, not a robot test |
| Door | 16 cases pass: breakaway between 7 and 8 N at the handle, 20 N for 1 s opens to the 100 deg limit, 60 kg box at 0.3 m/s pushes the unlatched door open, latched variant holds 200 N | forces applied to the leaf or a driven box; no hand, no handle mechanism. Below breakaway the leaf creeps (0.33 deg in 5 s at 7 N) |
| Drawers, keyboard trays | all cases pass: breakaway 6.35 N (top, mid), 7.45 N (bottom), 4.35 N (tray); stay where released; cubes stay inside | applied forces only; no hand |
| Chairs | pass: start to roll between 3 and 4 N; after a 20 N x 1 s push they coast 1.13 m in 2.5 s; 60 s rest drift below 0.6 mm/min and 0.12 deg/min | empty chair. With a 70 kg load on the seat the chair had not stopped after 5.5 m: rolling resistance does not grow with load. Tips over under a 60 N sideways push at z = 0.85 m |
Oscilloscope and props (docs/PROPS.md) |
16 cases pass at 200 and 120 Hz on a test table: key thresholds 1.9-2.0 N, knob breakaway 0.031-0.041 N m, BNC insertion 0.4-0.7 N | CPU PhysX only. Actuated by applied forces, force-driven pads and kinematic bodies, never by a robot hand |
scene_manip.usd settle 5 s |
props end at most 0.0011 mm (in the plane) from the authored rest pose; no warnings | settle only: nothing was manipulated in the composed scene |
Step time, scene.usd as authored |
CPU 0.86 ms (200 Hz), GPU 2.8 ms | one shared machine with two RTX 4090 D; indicative only |
Not tested: any robot interacting with the doors, drawers, trays, chairs, oscilloscope or props inside this package's scope; sensors or cameras; RTX renders of the current furniture (the acceptance renders predate the drawer handles); the GPU pipeline for the props; multi-environment cloning; any quantitative comparison of appearance with the real room (the photo comparison is qualitative, camera fit error 12.6-15.0 px).
Robot navigation and manipulation experiments exist but depend on third-party robot models and policies. They
are work in progress and are not part of this package (see THIRD_PARTY.md).
Known limitations and estimated values
Dimensions that were not measured in the real room:
- Room height 2.9 m, door height 2.15 m, all window extents, projection screen, the whole ceiling (visible ceiling at 2.60 m, soffits at 2.50 m, light panels, grilles, AC unit, motion-capture cameras) are estimated from three photos. The visible ceiling has no collider: between 2.60 m and the slab at 2.9 m there is nothing.
- The door bottleneck between
table_4andws_right_1is 0.7035 m in the model (other measures of the same gap: 0.683 m on the 1 cm map, 0.673 m as twice the path clearance). It follows from the floor plan and the asset sheets and has not been confirmed against the real room. - Rack length: the model uses 3.30 m (asset sheet); the floor plan says 3.23 m ("derived"). The row of tables is therefore 0.07 m closer to the door than on the floor plan.
- Rack handedness (which end has the 0.80 m bay) and workstation handedness are unconfirmed. The pedestals stand on the floor although the sheet shows them hanging under the desk.
- Drawer bar handles are a modelling choice for operability; the cabinet sheet shows handle-less fronts.
- All masses, inertias, joint friction values and surface friction coefficients are assumed or tuned for plausible behaviour, not measured.
Behaviour:
- No latch and no turning handle on the door: about 8 N opens it. The left leaf fully open is 13.5 cm from
table_4. - Joint friction is a capped damper: below the breakaway force parts creep instead of holding still (door: 0.33 deg in 5 s at 7 N; chair: 4.3 cm in 3 s at 2 N).
- Chair rolling resistance is independent of the load on the seat; height, backrest and armrests are not adjustable.
- Each chair is parked 5-6 cm from its workstation and blocks the drawers: it has to move back about 0.3 m before a drawer can open.
- Skirting, projection screen, sockets, thermostat, motion-capture cameras and rack foot plates are visual only. Windows are opaque and solid; nothing exists outside the room except a corridor floor patch behind the door.
- The oscilloscope and props are designed objects, not models of real items in the room. Controls are scaled up for 2 cm fingertips; knobs are limited, not endless; no cable; BNC retention is friction only (about 0.55 N) unless the attach helper is used.
- The maps describe
scene.usdwith everything at its start pose. They were not recomputed forscene_manip.usd. - Some comments in the YAML configs are stale (they describe earlier values); the values themselves are the ones
used. See
source/README.md.
docs/TECH_REPORT.md section 10 and appendix A list these points in full.
Provenance of dimensions
- Floor plan of the room (drawio, 125 px = 1 m): room size, wall thickness, columns, door opening, carpet, furniture layout.
- Eight dimensioned furniture sheets: outer dimensions of workstation, drawer cabinet, table, rack, chair; several sub-dimensions were measured on the sheets by pixel counting.
- Three retouched photos of the room: everything about height, windows, ceiling and colours, by eye and by a least-squares pinhole camera fit.
- The door opening direction (into the room) was confirmed by the owner of the room.
The photos, the floor plan and the sheets are not included in this package.
Regenerating
cd source
python -m pip install -r requirements.txt # Python 3.11
export PYTHONPATH=scripts
python scripts/build_room.py && python scripts/build_room.py --latched
python scripts/assets/build_table.py && python scripts/assets/build_shelf.py
python scripts/assets/build_chair.py && python scripts/assets/build_workstation.py
python scripts/assets/build_oscilloscope.py && python scripts/assets/build_props.py
python scripts/build_scene.py && python scripts/build_manip_scene.py # output in source/out/
See source/README.md for the checkers and for what differs from the development tree.
Citation and contact
@misc{real2sim_lab_room,
title = {Real2Sim Lab Room: a procedural USD digital twin of a laboratory for Isaac Sim},
author = {Real2Sim Lab Room authors},
year = {2026},
note = {https://huggingface.co/datasets/JiachengliXX/real2sim-lab-room}
}
Contact: via the repository discussions on Hugging Face.
Licence
- Generated assets, metadata, maps, previews and documentation in this repository (everything outside
source/andoptional/): CC BY 4.0. - Generator and checker code in
source/: Apache-2.0. optional/holds components with their own terms; read each component's README before use:optional/reference_inputs/: photos, floor plan and furniture sheets of the real room (owner's material, CC BY 4.0).optional/robot_look_layers/: appearance override layers authored for this project (CC BY 4.0); they contain no robot geometry but their mesh face-index lists were derived from third-party robot meshes.optional/robot_g1_rh56e2/: a Unitree G1 with Inspire RH56E2 hands, derived from third-party models (Unitree G1 description, BSD-3-Clause; Inspire RH56E2 URDF export obtained via a third-party repository tagged Apache-2.0, upstream licence not independently confirmed). Not covered by the CC BY 4.0 grant above.optional/nav_manip_wip/: navigation and manipulation code and results, work in progress (Apache-2.0 for our code). Reported manipulation success rates were obtained with arm effort limits above the vendor values (40 / 12 N·m used versus 25 / 5 N·m in the vendor URDF) and have not been re-measured with vendor limits.
- Downloads last month
- -





