Generalization beyond planar Push-T manipulation

Establish whether Spatial Language Modeling generalizes from planar Push-T manipulation to other object geometries, manipulation tasks, and three-dimensional scenes.

Background

The paper evaluates Spatial Language Modeling only on the planar Push-T task, in which a pusher moves a T-shaped block toward a target pose. Although the shared coordinate-token representation is intended to encode scene geometry, actions, and future states in a task-structured sequence, the experiments do not test whether the representation and training procedure transfer to different object shapes, manipulation settings, or three-dimensional environments. The authors therefore identify broader generalization as an unresolved issue.

References

Third, evaluation focuses on planar Push-T, so generalization to other object geometries, manipulation tasks, and three-dimensional scenes remains to be established.

— Unifying Policy Learning and State Prediction through Spatial Language Modeling  (2610.12172 - Wu et al., 8 Oct 2026) in Section 5, Limitations