README.md

metadata

license: cc-by-4.0
pretty_name: GraspNet-Infinite object pool (Objaverse-LVIS, manipulation-reasonable subset)
task_categories:
  - robotics
tags:
  - 3d
  - objaverse
  - grasping
  - synthetic-data
  - scene-generation
size_categories:
  - 1K<n<10K

GraspNet-Infinite object pool: 5,074 Objaverse-LVIS objects for tabletop scene generation

5,074 rigid, closed, single-object 3D models (GLB with textures where the source had them) selected from Objaverse-LVIS as manipulation-reasonable tabletop objects for a parallel-jaw gripper: containers, bottles and packaging, tools, toys and figurines, personal and household items, food, hardware. Every object carries a JSON record with its identity, licence and author, LVIS label, mesh statistics, native extent, a typical real-world size and the scale factor to metres. The set is meant for composing multi-object tabletop scenes (drop-in for BlenderProc, Isaac Sim, Kubric, PyBullet, trimesh/pyrender).

Layout

objects/A_containers_kitchenware.zip      <lvis_category>/<uid>.glb + <lvis_category>/<uid>.json
objects/B_bottles_cans_packaging.zip
objects/C_tools_utensils.zip
objects/D_toys_figurines_animals.zip
objects/E_personal_office_household.zip
objects/F_food.zip
objects/G_hardware_parts.zip
manifest.csv / manifest.jsonl              one record per object (same fields as the per-object JSON)
attribution.csv                            title, author, source URL, licence per object (keep with any redistribution)
load_example.py                            trimesh example: load, scale to metres, rest on a table

Unzip everything into one folder and the file field of each record is a relative path.

Per-object record

field	meaning
uid	Objaverse / Sketchfab model id; file is <lvis_category>/<uid>.glb
name, description, tags, author, author_url, source_url, published_at, licence	Sketchfab metadata; licence is CC-BY-4.0 or CC0-1.0 (nothing else was admitted)
lvis_category	the LVIS label Objaverse-LVIS assigns (automatic; a few percent of labels are loose)
family, family_letter	our grouping: A containers & kitchenware, B bottles, cans & packaging, C tools & utensils, D toys, figurines & animals, E personal, office & household, F food, G hardware & parts
shape_class	box, orb, cylinder or complex (coarse, assigned per LVIS category)
faces, components, watertight, textured, animated, bytes	mesh statistics after texture-seam merging; animated means the GLB carries an animation (rest pose loads fine)
native_extent, native_longest	axis-aligned bounding-box extent in the GLB's own units (glTF units are nominally metres, but Objaverse scales are arbitrary)
typical_longest_cm	typical real-world longest dimension of this kind of object (per LVIS category, from our size table)
scale_to_metres	multiply vertex coordinates by this to bring the object to typical_longest_cm in metres: (typical_longest_cm / 100) / native_longest
scale_source	the size is a per-category prior, not a per-instance measurement; a per-object estimate is in progress
size_note	small, kept for the size algorithm marks categories under 4 cm (coins, beads, caps, berries)
graspnet_nearest_object, graspnet_chamfer, graspnet_leakage_radius	nearest GraspNet-1Billion object (id 0..87) under a scale-normalised, PCA-aligned Chamfer distance, and the radius below which objects were excluded as near-copies

How the objects were selected

1. Label pass over all 1,156 LVIS categories: kept the seven families above; removed vehicles, furniture, buildings, appliances, people, plants, textiles and deformables, liquids, fixtures, weapons, flat slabs (plates, trays, books, cards, phones, flatbreads), glass and eyewear, objects typically longer than 27 cm, and categories whose LVIS label does not describe the contents (traffic cones filed as cone, barrels as wine_bucket, game characters as mascot). Small objects were kept on purpose. Per-object name rules removed logos, lighting-catalogue uploads, barrels, architecture, weapons under tool labels, NSFW; one object per uploader-and-mesh-count group (duplicates).
2. Licence: CC-BY and CC0 only.
3. Mesh gates: loads as at most 200 bodies, watertight after merging texture seams, at most three bodies, 500 to 3,000,000 faces.
4. Benchmark leakage: Chamfer distance to every GraspNet-1Billion object at least 0.01 (scale-normalised).

Not applied yet, so expect them in the set: label noise at the instance level (a few percent), rigged characters and human figures (kept as figurines), untextured meshes (36% have no texture), arbitrary resting orientation (use stable-pose computation or physics settling), and the per-category size prior (a mug record says 9 cm whether the model is an espresso cup or a beer stein).

Loading

import json, trimesh
rec = json.loads(open("A_containers_kitchenware/mug/<uid>.json").read())        # after unzipping
m = trimesh.load(rec["file"], force="mesh")          # merges the glTF scene into one mesh, keeps the texture
m.apply_scale(rec["scale_to_metres"])                 # now metres at the typical size
m.apply_translation(-m.bounds[0] * [0, 0, 1])         # (example) rest on z = 0; glTF is y-up, so pick your table normal first

load_example.py does this for a few objects and prints their extents. For scene composition: scale to metres, let a physics engine settle the objects on the table (or sample stable poses with trimesh.poses.compute_stable_poses), render RGB-D with GraspNet-1Billion intrinsics. The graspnet_* fields let you keep or drop objects by similarity to the benchmark.

Licence and attribution

Each model keeps its Sketchfab licence (licence field): CC-BY 4.0 requires attribution to the author; CC0 does not. attribution.csv lists title, author, author URL, source URL and licence for every object; keep it with any redistribution or publication. The Objaverse metadata is used under its ODC-By 1.0 licence. Please cite Objaverse:

@inproceedings{deitke2023objaverse, title={Objaverse: A Universe of Annotated 3D Objects},
  author={Deitke, Matt and Schwenk, Dustin and Salvador, Jordi and Weihs, Luca and Michel, Oscar and VanderBilt, Eli and
  Schmidt, Ludwig and Ehsani, Kiana and Kembhavi, Aniruddha and Farhadi, Ali}, booktitle={CVPR}, year={2023}}

Selection, metadata and size table: Peize Li, 2026 (GraspNet-Infinite project). 5,074 objects, 20.5 GB zipped. Family counts: A_containers_kitchenware 580, B_bottles_cans_packaging 317, C_tools_utensils 211, D_toys_figurines_animals 1809, E_personal_office_household 1062, F_food 1058, G_hardware_parts 37.

GraspNet-1Billion example scene

The folder graspnet_example/ holds one complete GraspNet-1Billion scene (scene_0000, RealSense, all 256 views: RGB, depth, labels, per-view object poses, camera trajectory) with its 9 object meshes and reconstruct_scene.py, a standalone script that rebuilds the scene from the stored poses and checks it against the depth image. It is the reference layout for rendering multi-object scenes in the GraspNet style. That folder is CC BY-NC-SA 4.0 (GraspNet team, MVIG, SJTU), not CC-BY; see its README.