OSH open-state mask — OOD diagnostic
Does the frozen OpenStateHead movability prior survive out-of-distribution? We render each asset, run OSH, and compare the predicted open-state mask m_open across three regimes.
Key finding. OSH overfits the geometry representation, not the category. On novel mesh categories the mask fires normally (0.073 ≈ in-distribution 0.084); it collapses (~0.02) only on ArtiCraft's primitive geometry. So OSH's benefit is robust to category shift but brittle to representation shift — which is why vanilla beats OSH on ArtiCraft but not on in-distribution data.
| regime | n | mean mask activation | spatial entropy |
|---|---|---|---|
| PNM (in-distribution) trained 7 categories | 6 | 0.084 | 0.755 |
| Held-out PNM (clean category-OOD) novel categories · MESH geom · same schema | 20 | 0.073 | 0.753 |
| ArtiCraft (representation-OOD) primitive geom (box/cyl/sphere) | 12 | 0.019 | 0.644 |
PNM — in-distribution (trained categories) 6
Reference. The OSH mask lands on the movable part; activation ~0.084.
Held-out PNM — clean category-OOD (mesh + same schema, novel category) 20
Mask activation 0.073 ≈ in-distribution. OSH still fires sharply on never-seen mesh categories (Chair, Faucet, Toilet, Door…). Category novelty alone does NOT break OSH.
ArtiCraft — representation-OOD (primitive geometry) 12
Mask activation collapses to ~0.02 (near-empty / mis-fired). The primitive box/cylinder/sphere geometry — not the new category — is what breaks the OSH movability prior.
