Overview
This direction explores how 3D Gaussian Splatting scene representations can support instance navigation and visual localization. The page is intentionally high-level. It does not disclose information beyond the public boundary of the 3DGS/patent-related work associated with Tsinghua’s vehicle-engineering context.
Problem
A representation designed for rendering is not automatically a scene model suitable for robot decisions. The research question is how view representation, spatial queries, instance semantics, and localization requirements can be placed within an evaluable system boundary.
System Architecture
The public boundary contains only four layers: scene capture and representation → queryable 3D information → navigation/localization interface → task-level evaluation. Pre-filing algorithm steps, parameters, data, and implementation details are withheld.
Results
Current status: patent disclosure in preparation. This site does not claim a granted patent, a published paper, or performance results without public evidence.
Limitations
This is not a technical disclosure. The core approach, data, comparisons, and contribution boundary will be updated only when publication is permitted and evidence has been reviewed.