projectsim/warehouse-manipulation-and-states

Warehouse Manipulation & States

Try it before you download. Every item in this dataset is live in projectsim Lab: drag the box flaps, roll the cage casters, then copy the API call that generates a set like it. Jump straight to a class: Boxes · Cages · Pallet Boxes.

Procedurally generated, sim-ready 3D warehouse assets - cardboard boxes, pallet-and-box stacks, and wire logistics cages - shipped as OpenUSD (.usda) and MuJoCo XML (.xml, MJCF) with per-zone PBR maps, plus a static glTF (.glb) for tooling that doesn't read either. Every variant carries real per-link mass, inertia, colliders, and, where present, articulation with joint limits and friction latches.

Motivation

Warehouse robotics work - pick-and-place of boxes, container manipulation, roll-cage navigation - needs assets that (a) look like the real thing, (b) sim-correctly in Isaac Sim / MuJoCo, and (c) come in enough parametric variation that a policy sees the distribution of shapes and materials, not one shape re-textured. Public warehouse assets are either scanned singletons (no variation, no reliable physics) or category-generic props (no articulation, guessed masses). This dataset is our procedural pipeline's output: every variant has a spec that produces watertight geometry, sim-authored physics, and PBR textures.

What's included

Class	Count	Articulation	Materials
cardboard_box	75	4 top flaps, each on a revolute joint with a 0.01 N·m light hold	body + four flap zones
pallet_boxes	75	static (pallet + stacked boxes prop)	pallet, boxes, tape, plus a printed shipping-label zone
logistics_cage	75	4 swivel casters + 4 wheel axles (8 revolute joints) per variant	painted-metal deck (industrial finish palette), brushed-steel wire, black-rubber casters

Total: 225 variants across 3 classes, ~12 GB.

Per-variant contents

assets/<class>/<vid>/
├── <vid>.usda            OpenUSD, textures referenced as ./textures/*.png
├── <vid>.glb             glTF binary, embedded textures, static pose
├── textures/*.png        PBR maps (base_color, normal, roughness, metallic; +labels on pallet_boxes)
└── mjcf/
    ├── <vid>.xml         MJCF, drop-in for MuJoCo `simulate` (`simulate <vid>.xml`)
    ├── visuals/*.obj     per-zone visual meshes, textured
    ├── visuals/textures/ PBR maps referenced by the MJCF
    └── collision/*.obj   per-component convex hulls (or primitives, per authored collision type)

The MJCF mirrors the USD authoring: primitives (box, sphere) where the rig authored them, per-component convex hulls elsewhere; intra-asset self-collision excluded so flaps and moving parts return to their home pose; visual geoms in group 2 (shown by default), collision geoms in group 3 (press 3 in simulate to reveal).

Parametric variation covered

Every class is authored from a spec.json sampled deterministically from a (count, seed) pair; anchor rows are the class's presets and dimension extremes.

• cardboard_box - 3 mm single-wall corrugated RSC, always articulated (the has_top_tape param is pinned to false at generation). Body: 250–500 mm × 156–364 mm × 200–350 mm, bevel-edged panels, optional stadium hand-hole cutouts.
• pallet_boxes - wooden EUR pallet loaded with 3–10 stacked cardboard boxes; pallet 1200 × 800 mm ±10 %, per-box dims sampled independently, tape strips on selected boxes, printed shipping-label region on one visible face.
• logistics_cage - 4-sided wire cage with swivel casters; two presets are gated (nestable-frame, drop-gate). Base dims 720–1500 × 900–1700 × 700–1000 mm. All 75 variants roll (base_style is pinned to swivel_casters - the sampler cannot draw the "welded rigid" alternative).

The exact per-variant parameter values are not shipped in v1.

Generate your own

This dataset is sample output from the projectsim pipeline. For sim-ready assets of your objects, use the projectsim API:

• Standard asset - one image in, a sim-ready asset out
• Measured asset - mm-accurate, from a ChArUco-board capture
• Variation sets - procedural or high-fidelity families from a single object

Self-serve: get an API token and order your first set in minutes. Need 1,000–10,000 assets of a category? Request a larger dataset →

Intended use

• Training and evaluating manipulation policies (grasp, place, stack, open flap)
• Perception (segmentation, category ID, pose estimation) from synthetic renders
• Domain-randomization sources for warehouse pipelines
• Sim-based experimentation before scanning real assets
• Prototyping in Isaac Sim / MuJoCo / any USD-aware DCC

Tested simulators

Verified end-to-end:

• NVIDIA Isaac Sim 6.0.1 - loads <vid>.usda headless and GUI. Doors/flaps open under applied effort, casters roll, taped boxes stay one rigid body, latches hold under gravity.
• MuJoCo 3.11.0 - loads mjcf/<vid>.xml directly (simulate mjcf/<vid>.xml). Joint dynamics, mass/inertia, and per-component collision hulls all match the authored rig.
• Blender 5.0 - USD and glTF import both work.

Physical properties - authored vs estimated

Property	How	Notes
Volume, COM, inertia	Computed from watertight mesh	Falls back to the convex hull for open-shell parts (cage frame)
Mass	Volume × density	Density per zone, classified by a material-analysis VLM. The VLM inspects the object's renders together with the zone names in context — e.g. reading pallet slats as pine, box walls as kraft cardboard, cage wire as galvanised steel, casters as rubber — and looks each zone's density up in a shared material table (wood ~600, kraft ~180, steel ~7850, rubber ~1200 kg/m³). Per-link mass = mesh volume × classified density.
Collision shape	Auto-classified per link	Author-time collision types per link, taken directly from the rig: • primitives (box, sphere) where the rig is geometric — wheels, casters, simple frames • convexHull for sealed volumes with a single graspable shape (cardboard flaps, cage forks) • convexDecomposition with PhysX shrink-wrap for multi-part rigid props whose interior matters (pallet_boxes: pallet blocks + individual boxes each become their own hull at load) • sdf (sdfResolution 256) for thin, non-convex shells where a decomposition would seal openings (logistics_cage frame, cardboard_box body) MJCF ships the same intent: primitives verbatim, per-connected-component convex hulls where the USD uses convexHull / convexDecomposition, and a single mesh geom for sdf shells (MuJoCo auto-hulls it).
Joint limits	Authored in the rig	Explicit degrees / meters per joint
Joint friction / latches	Authored	Cardboard flaps: 0.01 N·m soft latch (opens by hand, doesn't self-open). No other latches in v1
Textures	4K PBR maps (cardboard_box, pallet_boxes); procedural PBR from an industrial finish palette (logistics_cage)	Zone-mapped so a downstream engine can rebind materials per zone
max_mass_kg	Cavity volume × water density	Meta-only hint for pick-and-place - not the physics mass

Limitations

• Physics lives in the .usda and .xml, not the .glb. glTF has no physics equivalent - the .glb is a visual copy (static pose, single mesh per link, no joints, no colliders). Use it for previews, tooling that reads neither USD nor MJCF, Blender/Unity/three.js/model-viewer; use .usda for Isaac Sim, mjcf/<vid>.xml for MuJoCo.

License

Files: CC BY 4.0 - free to use, share, and adapt for any purpose, including commercial use; attribution required.

Citation

@dataset{kaedim_warehouse_manipulation_and_states_2026,
  author = {{Kaedim} and {projectsim}},
  title = {Warehouse Manipulation \& States},
  year = {2026},
  publisher = {Hugging Face},
  version = {1.0},
  url = {https://huggingface.co/datasets/projectsim/warehouse-manipulation-and-states}
}

Maintenance

New classes and variants are added as separate top-level assets/<class>/ folders; the on-disk <vid> prefix is stable across future commits - a v03_pallet_boxes_6c1b6d today will be the same asset next release. Regressions are republished as an additive commit; assets are only ever removed with a version bump.

Issues / questions: open a Discussion on this dataset repo, or email team@projectsim.ai.