Event
SAGE-3D proposes 3DGS semantic-physical enhanced navigation method
1 reports · 1 sources
The SAGE-3D authors propose a 3DGS environment construction method for vision-language navigation: 3DGS handles realistic rendering, the source mesh handles reliable collision, and manual semantic annotation handles instruction grounding; after decoupling, the three are assembled into an executable navigation environment, which is modeled as a semantically and physically enhanced POMDP. Collision bodies are extracted from the source mesh via CoACD convex decomposition and then embedded into the 3DGS scene, so that the robot is subject to realistic collision constraints. The authors also release the InteriorGS dataset, containing 1K object-level annotated 3DGS indoor scenes, 554K object instances, and 755 categories. The authors state that 3DGS scenes are harder to converge but generalize strongly.
Report timeline
- ICLR 2026 paper SAGE-3D adds collision and semantics to 3DGS scenes to support vision-language navigation
The ICLR 2026 paper SAGE-3D proposes decoupling and assembling 3DGS rendering, source-mesh collision bodies, and manual semantic annotations into an executable navigation environment, modeled as a semantics- and physics-enhanced POMDP. Its object-level semantic anchoring provides category, instance ID, and bounding box, while physics-aware execution assembles collision bodies extracted from source meshes via CoACD convex decomposition and embeds them into the 3DGS scene.
3D视觉工坊 · 微信公众号 ·