LEXIS-Net
A discrete vocabulary for interaction

input human mesh — SMPL-X body
human articulation $\theta$ $\mathbb{R}^{21\times 3}$
input object mesh — canonical suitcase
canonicalize $f^o$ shape embed
Encoder $E_\varphi$
$\theta$, $f^o$ → $\hat{z}$
ẑ
LEXIS
$\mathcal{Z}$
c1
c7
c42
c89
c171
c250
c413
c682
…
discrete · $\mathbb{R}^{2048 \times 128}$
z
Decoder $D_\psi$
$z$, $f^o$ → IF
triplane query — body + object
body & object triplane query
LEXIS-Net output — body + object with InterField proximity heatmap proximity colorbar
InterField
body & object surfaces
$f^o$
Manifold of valid signatures
  • discrete prior over plausible interaction signatures
  • captures structured body & object proximity
Pose & action driven
  • a code captures how body pose and object shape encode interaction
Continuous InterField
  • Surface-topology agnostic (mesh, point cloud etc.)
  • proximity queryable at any 3D point on body or object
LEXIS · Antić et al. 2026