maxwildersmith/splataverse-artvip

Splataverse · ArtVIP articulated splats

Articulated 3D Gaussian splats converted from ArtVIP (Apache-2.0): 435 articulated objects and 6 interactive indoor scenes. Every rigid body of the source asset is its own splat, and the asset's own UsdPhysics joints (prismatic and revolute, with pivots, axes and limits) are kept in a USD stage that references the splats. You can scrub any object or scene between joint states, e.g. for articulation estimation from observed states, or for building metric scene graphs from segmented objects.

Converted with splataverse's analytic, training-free mesh-to-splat converter (textures sampled per Gaussian from the source UsdPreviewSurface materials, including their UV transforms). Splats are stored as PLY rather than SPZ so the per-Gaussian id fields travel with them.

Previews

Ten example assets, joints scrubbed open/closed (left: splat colour, right: moving parts tinted), rendered with gsplat from the released files:

Layout

One Gaussian splat per rigid body — the static frame plus each movable part — linked by a USD stage:

objects/<category>/<asset>/
  articulation.usda      # /Furniture/<part> Xforms with splat:file -> parts/*.ply, UsdPhysics joints under /Furniture/joints
  parts/frame.ply        # the static root body
  parts/<part>.ply       # one splat per movable rigid body
  summary.json           # parts, joints, and the id tables below
objects/index.jsonl      # one row per asset: category, path, part/joint/subpart counts, classes present
scenes/<scene>/...        # same layout; every object in the room contributes its own parts
class_names.json         # dataset-wide part-class vocabulary (class_id -> name)
object_class_names.json  # dataset-wide object-class vocabulary (object_class_id -> name)
previews/<asset>.gif     # 10 example assets, joints scrubbed open/closed (left: colour, right: parts tinted)

Per-Gaussian fields (binary little-endian PLY)

Standard 3DGS fields (x y z, f_dc_0..2, opacity as the raw value in [0,1], scale_0..2 as log scales, rot_0..3 as a wxyz quaternion), plus int32 ids so any division of an asset can be rendered or extracted directly:

field	meaning
part_id	rigid body within the asset (0 = static frame); matches summary.json parts and the USD
subpart_id	element inside a body, e.g. a handle on a door, a shelf in the frame (summary.json subparts)
class_id	semantic class of that subpart (door, drawer, handle, shelf, tabletop, button, body, ...) — class_names.json
object_class_id	class of the object the Gaussian belongs to (cabinet, sofa, book, ...) — object_class_names.json
instance_id	object instance within a scene (always 0 for single objects; summary.json instances)
seg_id	unused (0), kept for compatibility with other splataverse splats

Classes come from ArtVIP's own USD semantic labels (SemanticsLabelsAPI:class) where present; the interactive scenes carry no labels, so their classes are derived from prim names (E_drawer1_24 -> drawer, model_book_12 -> book), with unrecognised elements labelled body.

Conventions: Y-up, metres. Part splats are stored in the world (rest-pose) frame; each joint's localPos0/localRot0 is its pivot and frame in that world frame, with the motion axis along the joint's local X. Revolute limits are in degrees, prismatic in metres (UsdPhysics units).

Joint value 0 is the rest pose the splats are stored in, as authored in ArtVIP — usually closed, but not always (laptops, for example, are stored open and close at +90°). To pose an asset: for joint value v, move the child part's Gaussians by v · axis (prismatic) or rotate them by v degrees about axis through pivot (revolute), composing parent motions first for chained parts.

Notes

• Glass and other transmissive materials are rendered from their diffuse colour only (they appear opaque).
• 8 of the source scenes' joints drive bodies with no geometry and are omitted.
• Low-resolution textures on large surfaces were bilinearly upsampled before conversion so panels are densely covered.

Attribution

Source assets: ArtVIP — ArtVIP: Articulated Digital Assets of Visual Realism, Modular Interaction, and Physical Fidelity for Robot Learning (ICLR 2026), arXiv:2506.04941, released under Apache-2.0 at x-humanoid-robomind/ArtVIP. This dataset is a derived conversion under the same license.