Active Neural Mapping at Scale

TL;DR

This work tackles the challenge of actively mapping large-scale indoor environments with incomplete observations by marrying implicit neural representations (INRs) and explicit geometric topology. The authors extract a generalized Voronoi graph (GVG) from a continually updated NeRF-based neural map, anchoring uncertain regions to Voronoi vertices to enable adaptive-granularity, safe exploration along traversable paths. A hybrid neural representation and a hierarchical planning framework enable scalable reconstruction across 20+ rooms, with a graph-based planner (Dijkstra) operating on the sparse topology to drive exploration. Experiments on Gibson and Matterport3D datasets show state-of-the-art completeness and efficient exploration (7–9 Hz) with robust large-scale reconstruction and comprehensive ablations validating the ROI anchoring, visibility guidance, and bootstrap strategies.

Abstract

We introduce a NeRF-based active mapping system that enables efficient and robust exploration of large-scale indoor environments. The key to our approach is the extraction of a generalized Voronoi graph (GVG) from the continually updated neural map, leading to the synergistic integration of scene geometry, appearance, topology, and uncertainty. Anchoring uncertain areas induced by the neural map to the vertices of GVG allows the exploration to undergo adaptive granularity along a safe path that traverses unknown areas efficiently. Harnessing a modern hybrid NeRF representation, the proposed system achieves competitive results in terms of reconstruction accuracy, coverage completeness, and exploration efficiency even when scaling up to large indoor environments. Extensive results at different scales validate the efficacy of the proposed system.

Figures (9)

• Figure 1: The reconstructed mesh through autonomous exploration with a continually-learned neural map (MP3D-zsNo4 Matterport2017_3DV with 23 rooms). The synergistic integration of geometry, appearance, topology, and uncertainty information within the neural map leads to complete and accurate reconstruction of large-scale environments.
• Figure 2: The structuring process through Voronoi tessellation. The Generalized Voronoi graph (GVG) (b) is extracted as equidistant samples to the zero-crossing surfaces (a). The visible regions of interest (ROIs) guided by perturbed map parameters are anchored to neighboring vertices (c).
• Figure 3: The fine-grained local horizon (the dark area near the frustum that indicates the nearby distance values) guarantees safe exploration, while the rest of the map maintains the global regions of interest coarsely.
• Figure 4: The reconstruction results of small scenes through active exploration.
• Figure 5: The complex scene with severe occlusions addresses challenges to trade-offs between efficient exploration and accurate reconstruction.