MoreAI — spatial intelligence for robotsContact
Guide · 3D Data

Objaverse-XL
for Robot Learning.

Objaverse-XL is the largest open collection of 3D assets — over 10 million objects with real geometry, not just pixels. For geometry-primed policies, that changes what "pretraining data" means.

01What Objaverse-XL Is

Objaverse-XL is an open, web-scale dataset of 10M+ 3D objects aggregated from Sketchfab, GitHub, Thingiverse, Polycam, and the Smithsonian. Each entry ships with mesh geometry, and often materials and rigging — the exact substrate a geometry-native model can consume without first inferring depth from a photograph.

Contrast that with pixel-first pretraining: video and teleop frames give a model appearance, but geometry has to be learned back out of RGB. Objaverse-XL skips the round trip.

02Why Geometric Priming Needs 3D Data

A pixel-first foundation learns to hallucinate 3D from 2D. That works for description, but it burns compute on re-deriving what a depth channel already contains. Geometric priming inverts the assumption: pretrain on surfaces, edges, and poses first, then let language and appearance attach.

Objaverse-XL is the scale that makes this feasible outside a lab. Ten million assets is roughly the point at which a geometry-native encoder generalizes across shape families rather than memorizing them.

03A Practical Training Recipe

  • Ingest — Sample meshes across categories, keeping the natural long-tail of Objaverse-XL — cluttered, imperfect assets are the ones that transfer.
  • Render with depth — Rasterize each asset from many viewpoints and export RGB + depth + surface normals. Depth is the supervision signal a geometric encoder needs.
  • Prime the encoder — Train the geometric neural network — GINN in MoreSight™ — to output invariants (position, orientation, velocity) rather than tokens.
  • Adapt on task data — Fine-tune focused action models on minutes-to-hours of teleop against the primed encoder. The heavy lift stays with pretraining.

04What Objaverse-XL Doesn't Give You

Static assets aren't dynamics. Objaverse-XL supplies geometry; physics — friction, contact, mass — still has to come from simulation or real interaction. Treat it as the shape prior, not the physics prior.

It also isn't a policy dataset. A geometry-primed encoder is the foundation; teleop and simulated interaction remain required for closed-loop control.

←BACK TO GUIDES