Placement_Assets / docs /external_source_catalog.md
yuchi233's picture
Upload folder using huggingface_hub
bcbeda2 verified
|
Raw
History Blame Contribute Delete
4.9 kB

External Source Catalog

This catalog tracks candidate external asset sources for placement and articulated-object tasks. Entries here are link-only references unless a concrete asset is physically copied into assets/ and indexed in manifest/assets.jsonl.

Intake priority

Priority Source type Examples Reason
P0 Already simulation-ready or generated with reliable articulation RoboCasa, DISCOVERSE, Infinigen-Articulated, MuJoCo Menagerie Fastest path to reusable MJCF/URDF assets after validation.
P1 Articulated datasets or task layers with joints/semantics PartNet-Mobility, GAPartNet, ManiSkill, BEHAVIOR-1K/OmniGibson, UniDoorManip, AdaManip, DoorGym Useful candidates, but usually require license review, conversion, collision/dynamics QA, and policy mapping.
P2 Static or weakly structured large-scale object pools Objaverse, Google Scanned Objects, SketchUp 3D Warehouse, PartNet static Good for long-tail visual variety; not automatically articulated or simulation-ready.
P3 Generative/reconstruction methods Hunyuan3D, PartCrafter, Articulate-Anything, Real2Code, FreeArt3D, SINGAPO, DIPO, PhysNAP, NAP, CAGE, ArtFormer Useful for proposals, bootstrapping, or candidate geometry, but outputs must still pass physical and task validation.

Candidate sources

Source URL Expected readiness before local adaptation Placement relevance
RoboCasa https://robocasa.ai L1 candidate Kitchen fixtures, objects, scenes, and appliance tasks.
DISCOVERSE https://github.com/discoverse-dev/DISCOVERSE L1/L2 candidate by subasset MJCF task environments, robots, manipulation objects, and scenes.
MuJoCo Menagerie https://github.com/google-deepmind/mujoco_menagerie L1 candidate Robot arms, hands, grippers, and MJCF models.
Infinigen / Infinigen-Articulated https://github.com/princeton-vl/infinigen L1/L2 candidate after export Programmatic doors, cabinets, drawers, windows, appliances, containers.
PartNet-Mobility https://huggingface.co/datasets/haosulab/PartNet-Mobility L2 candidate Articulated furniture, containers, appliances.
GAPartNet https://pku-epic.github.io/GAPartNet/ L2 candidate Actionable parts, handles, drawers, doors, lids.
ManiSkill https://github.com/haosulab/ManiSkill L1 task-layer candidate SAPIEN tasks over PartNet-Mobility assets, especially cabinets/drawers.
BEHAVIOR-1K / OmniGibson https://behavior.stanford.edu/knowledgebase/objects/index.html L1/L2/L3 candidate by object status Household objects, articulated furniture, appliances, receptacles.
UniDoorManip https://github.com/sectionZ6/UniDoorManip L1/L3 candidate Door/body/handle combinations and open-door manipulation assets.
DoorGym https://github.com/PSVL/DoorGym L1 generated worlds, L3 handles Door/handle geometry and MuJoCo door tasks.
AdaManip https://github.com/yuanfei-Wang/AdaManip L1 candidate, license must be checked Articulated manipulation objects and mechanisms.
Articulate3D https://huggingface.co/datasets/INSAIT-Institute/Articulate3D L2 annotations, paired scene mesh required Real-scene articulated part annotations.
MultiScan https://3dlg-hcvc.github.io/multiscan/ L2 candidate, gated/NC constraints likely Real scanned indoor articulated objects.
OPD https://github.com/3dlg-hcvc/OPD L2 candidate Openable-part detection assets and annotations.
AKB-48 https://liuliu66.github.io/articulationobjects/ L2 candidate Articulated object knowledge base.
Objaverse / Objaverse-XL https://objaverse.allenai.org L4/L5 by default Large static long-tail object pool.
Google Scanned Objects https://app.ignitionrobotics.org/GoogleResearch/fuel/collections/Google%20Scanned%20Objects L4 by default Scanned household objects for placement/receptacle scenes.
SketchUp 3D Warehouse https://3dwarehouse.sketchup.com L3/L4/L5 by object Long-tail furniture, appliances, rooms; redistribution must be reviewed per model.

Deduplication rule

Many sources are derived from the same upstream geometry. Do not sum their counts directly.

Use a source-family identity when possible:

source_family_id + source_asset_id + raw_mesh_sha256 + joint_graph_sha256

Examples:

  • PartNet -> PartNet-Mobility -> GAPartNet / ManiSkill / OPDSynth
  • UniDoorManip / AdaManip / DoorGym may reuse PartNet or 3D Warehouse geometry
  • A new task annotation over the same mesh is an annotation/task increment, not a new geometry asset

Reporting rule

For each source or import campaign, report these counts separately:

candidate_count
license_pass_count
downloaded_count
load_pass_count
physics_pass_count
policy_pass_count
L0_count

Do not report downloaded files as production-ready assets.