MrNeRF (@janusch_patas)
2025-02-28 | โค๏ธ 57 | ๐ 10
ATLAS Navigator: Active Task-driven LAnguage-embedded Gaussian Splatting
Contributions: โข Memory-efficient online language-embedded Gaussian splatting: We introduce a method for Gaussian splatting that incorporates language embeddings efficiently to optimize memory usage. This approach combines dimensionality reduction techniques, such as principal component analysis (PCA), with efficient online updates, ensuring that the map maintains a compact and scalable representation while preserving the necessary geometric and semantic fidelity.
โข Submapping with sparse semantic hierarchical clusters and dense geometry: The proposed map is structured into sparse semantic clusters that represent regions or objects and dense geometric representations for navigation. This design leverages a tectonic structure, which updates submap anchor poses after loop closure. As a result, the map is compatible with external odometry sources and pre-built maps, enabling consistent updates and large-scale navigation.
โข Large-scale (>km) autonomous task-driven navigation with reusable maps: Our approach supports task-driven navigation by dynamically identifying and retrieving map regions relevant to specific objectives. The system is also compatible with real-time, interactive task re-specification, allowing for adaptive exploration and navigation in large-scale environments.
๋ฏธ๋์ด
๐ Related
Auto-generated - needs manual review