CLEAR: A Semantic-Geometric Terrain Abstraction for Large-Scale Unstructured Environments

TL;DR

CLEAR addresses the problem of scalable, long-horizon navigation in large, unstructured terrains by fusing semantic landcover boundaries with elevation geometry. The method couples Boundary-Seeded Decomposition with recursive Elevation Plane Fitting to produce convex, semantically aligned regions encoded as a terrain-aware Graph, with costs that reflect slope, roughness, landcover, and heading. Across maps spanning multiple tens of square kilometers, CLEAR achieves up to 10x faster planning with about 6.7% overhead and yields 6–9% shorter, more reliable paths than baselines, validated in physics-based simulations. This approach enables robust, real-time planning for disaster response, defense, and planetary exploration where traditional abstractions struggle at scale.

Abstract

Long-horizon navigation in unstructured environments demands terrain abstractions that scale to tens of km$^2$ while preserving semantic and geometric structure, a combination existing methods fail to achieve. Grids scale poorly; quadtrees misalign with terrain boundaries; neither encodes landcover semantics essential for traversability-aware planning. This yields infeasible or unreliable paths for autonomous ground vehicles operating over 10+ km$^2$ under real-time constraints. CLEAR (Connected Landcover Elevation Abstract Representation) couples boundary-aware spatial decomposition with recursive plane fitting to produce convex, semantically aligned regions encoded as a terrain-aware graph. Evaluated on maps spanning 9-100~km$^2$ using a physics-based simulator, CLEAR achieves up to 10x faster planning than raw grids with only 6.7% cost overhead and delivers 6-9% shorter, more reliable paths than other abstraction baselines. These results highlight CLEAR's scalability and utility for long-range navigation in applications such as disaster response, defense, and planetary exploration.

Figures (9)

• Figure 1: The CLEAR framework takes in landcover and elevation data (a), applies boundary-seeded decomposition [\ref{['alg:bas']}] (b), performs region-wise surface regression and abstraction [\ref{['alg:plane_fitting']}] (c), and encodes connectivity through a graph representation over the resulting regions (d).
• Figure 2: Illustration of decomposition methods: (a) Grid — 400 cells, (b) Hexagonal — 448 cells, (c) Quadtree — 184 cells, (d) — 105 cells.
• Figure 3: Land-cover reconstruction on the Wharton (W) subregion. Grid, Hex, and Quadtree distort semantic boundaries, whereas better preserves fine classes and irregular transitions.
• Figure 4: Elevation reconstruction on the Wharton (W) subregion with $\sim$650 regions. aligns with terrain contours and yields smooth planar regions, while Grid/Hex show stair-step artifacts and Quadtree exhibits block artifacts.
• Figure 5: Performance across 12 maps of varying complexity. Each point averages 5 decomposition resolutions. Axes report semantic fidelity (mIoU) and geometric fidelity (RMSE). consistently outperforms Grid, Hexagonal, and Quadtree across all complexities.