PaperlayerView exact revision on arXiv

Source-linked AI summary

TASG-Explore: Traversability-Aware Sector-Guided Exploration for Ground Robot on Uneven Terrain

Shaocong Wang, Shiliang Shao, Ting Wang, Guangjie Han, Lianqing Liu

arXiv:2609.08512v2cs.RO

TL;DR

Uneven-terrain exploration requires ground robots to balance coverage efficiency with terrain safety and local detail. TASG-Explore combines hierarchical traversability analysis, sector segmentation, and dynamic-roadmap planning; benchmarks report superior overall exploration performance and real-world experiments validate its practicality.

  • Problem

    On uneven terrain, compact region-level guidance can miss narrow passages and locally complex areas, while detailed terrain analysis can reduce exploration efficiency.

    Source: §I · p. 1, C. Uneven-Terrain Exploration Methods · p. 4

  • Method

    TASG-Explore combines variable-voxel ground fitting and adaptive 8-bit obstacle encoding with sector-guided planning over a dynamic roadmap.

    Source: §I · pp. 1–2

  • Results

    Benchmarks in caves, forests, and rugged hills show superior overall exploration performance over existing baselines, and real-world experiments validate practicality.

    Source: §X · p. 19

  • Takeaways & Limitations

    The framework supports efficient global exploration while retaining traversability-aware completeness and safety in uneven outdoor environments.

    Source: §X · p. 19

Abstract

from arXiv · show

Autonomous exploration on uneven terrain requires ground robots to balance exploration efficiency, coverage completeness, and terrain safety. Detailed tsrrain reasoning improves local reliability but can slow large-scale exploration, whereas coarse region guidance expands quickly in open areas but can miss narrow passages and irregular traversable boundaries. To address this challenge, this paper presents TASG-Explore, a traversability-aware sector-guided exploration framework for ground robots. The framework first performs hierarchical traversability analysis using variable-voxel ground fitting and adaptive 8-bit obstacle encoding. It then splitting cost map into sectors, incrementally updates sector clusters, extracts terrain-coupled frontier viewpoints, and maintains a dynamic topological roadmap with unknown topological hypotheses. Finally, a sector-guided planner selects region targets and inserts local viewpoints to generate efficient exploration routes. Benchmark experiments in diverse challenging environments, including caves, forests, and rugged hills, show that TASG-Explore achieves the best overall performance among six representative state-of-the-art planners. The proposed traversability analysis improves processing efficiency by 6.3 times while maintaining high accuracy, and the exploration planner improves exploration efficiency by 51% and increases coverage by up to 2.95 times in rugged hill scene. Large-scale real-world experiments further demonstrate the practical value of the proposed method.

I. INTRODUCTION

Ground-robot exploration on uneven terrain must balance route efficiency with terrain-aware safety and coverage. TASG-Explore combines hierarchical traversability analysis, incremental sector segmentation, and roadmap-based planning to address this tradeoff.

I. INTRODUCTION

  • Terrain-aware ground exploration must balance efficient global expansion with safety and completeness, since planar paths can cross unsafe terrain and coarse guidance can miss narrow or irregular areas.Fine-grained analysis can also make exploration conservative, motivating adaptive coarse-to-local reasoning.Source: §I · p. 1
  • TASG-Explore combines variable-voxel ground fitting and adaptive 8-bit obstacle encoding so planning can reason coarsely in open areas and refine terrain details where safety and completeness matter.The framework also incrementally segments unknown regions and generates terrain-coupled reachable viewpoints.Source: §I · pp. 1–2
  • A dynamic roadmap uses verified traversable samples for safe known-space connectivity and unknown samples as temporary hypotheses toward distant passages, while sector targets guide coverage and local viewpoints recover observations.Multi-scale sparsification reduces redundancy in open regions while retaining vertices important to narrow structures and terrain transitions.Source: §I · p. 2
  • Benchmark experiments across caves, forests, and outdoor hills report the best overall performance across environment types, with real-world experiments also validating the system.The comparison uses the same robot, sensor, map, and termination settings for the proposed method and state-of-the-art planners.Source: §I · p. 2, Figure 1 · p. 2

A. Autonomous Exploration Methods

  • Exploration research has shifted from local frontier or information-gain choices toward global route reasoning, but many global methods target aerial or general 3-D exploration and omit ground traversability.Local methods typically select viewpoints using immediate information gain, visibility, and travel distance.Source: A. Autonomous Exploration Methods · pp. 2–3

B. Ground-Robot Exploration Methods

Ground-robot exploration methods commonly reason over planar free space, while uneven terrain makes support and traversability central to reachability. TASG-Explore links terrain modeling, sector segmentation, and global planning to manage that constraint.

B. Ground-Robot Exploration Methods

  • Many ground-robot planners use 2-D frontiers, regions, or topological graphs, but geometrically short planar paths and high-value frontiers may be unsafe or unreachable on uneven terrain.Related approaches include persistent roadmaps, skeleton graphs, frontier processing, and hierarchical region-based planning.Source: B. Ground-Robot Exploration Methods · p. 3

C. Uneven-Terrain Exploration Methods

  • Uneven-terrain exploration must balance coverage efficiency with terrain fidelity: compact region or graph methods can miss narrow passages and small reachable areas, while detailed analysis can reduce exploration efficiency.Existing terrain-aware work includes navigation, target selection, and large-region or multi-resolution representations.Source: C. Uneven-Terrain Exploration Methods · pp. 3–4

A. Problem Definition

  • The problem is to build a map online and choose safe, informative future observation positions to complete coverage in minimum exploration time.Candidate targets must lie on locally supportable terrain, and their paths must avoid non-traversable, unreliable terrain and obstacles.Source: A. Problem Definition · p. 4

B. Proposed Framework Overview

  • The framework updates terrain models from local submaps, segments unknown space into sectors, extracts terrain-coupled frontier viewpoints, then maintains a dynamic roadmap and jointly orders region and viewpoint targets.It repeats the update-and-replan process until no valid unknown regions or frontier viewpoints remain.Source: B. Proposed Framework Overview · pp. 4–5

IV. HIERARCHICAL TRAVERSABILITY ANALYSIS

  • Variable-resolution voxels model large ground surfaces, while obstacle encoding adds fine obstacle information; their fused traversability costs support region abstraction, connectivity queries, and path-cost evaluation.Traversability cost c distinguishes unknown terrain, known traversable terrain, and higher-cost or non-traversable terrain.Source: §IV · p. 5, Figure 3 · p. 5

A. Variable-Voxel Ground Fitting

Variable-voxel ground fitting adapts spatial resolution to terrain structure, while adaptive 8-bit obstacle encoding adds fine vertical occupancy evidence to the traversability cost map.

A. Variable-Voxel Ground Fitting

  • Open smooth ground remains represented by coarse stable planes, while voxels with non-planar structure split recursively until a stable plane is found or maximum depth is reached.This uses finer resolution around steps, obstacle boundaries, and rough regions.Source: A. Variable-Voxel Ground Fitting · p. 5
  • Each 2-D cell is initialized from its lowest local surface; cells without a valid lower plane remain unknown, while initialization restores traversable surfaces and reduces map holes.The lower-plane traversability decision uses slope and normalized roughness criteria.Source: A. Variable-Voxel Ground Fitting · p. 5

B. Adaptive 8-bit Obstacle Encoding

  • Each cell's adaptive 8-bit height code records obstacle evidence across vertical layers, with encoding range and scale updated dynamically as observed height changes.The code represents one byte of occupancy data, and the previous code is reprojected rather than cleared when the scale changes.Source: B. Adaptive 8-bit Obstacle Encoding · pp. 5–6
  • The refined local cost marks sufficiently occupied cells as obstacles, treats near-vertical height discontinuities as obstacles, and assigns higher cost—not exclusion—to continuous rough slopes.Reprojecting prior occupancy by interval overlap preserves accumulated obstacle evidence when vertical scale changes.Source: B. Adaptive 8-bit Obstacle Encoding · p. 6

C. Terrain Cost Incremental Update

The global terrain map incrementally fuses local cost observations and, for traversable cells, terrain-surface information. Obstacle reliability and consecutive-free observations help remove transient false obstacles while retaining stable terrain boundaries.

C. Terrain Cost Incremental Update

  • Local cell costs are incrementally fused into the fixed-size global cost map, with observation reliability and robot distance determining the local fusion weight.The same projection updates height, ground normal, and plane offset for traversable cells.Source: C. Terrain Cost Incremental Update · p. 6
  • Traversable-cell projections also update global point-cloud height, ground normal, and plane offset for terrain queries and viewpoint generation.The map supports obstacle and free-space queries, terrain-coupled viewpoint generation, and topological edge validation.Source: C. Terrain Cost Incremental Update · p. 6
  • Obstacle reliability rises with occupied observations and falls with free observations; repeated free observations can remove an old obstacle below the retention threshold.This filter is intended to keep real walls and terrain breaks stable while allowing short-term obstacles from moving objects or sparse measurements to disappear.Source: C. Terrain Cost Incremental Update · p. 6

V. REGION SEGMENTATION AND SAMPLING

The segmentation module organizes the updated terrain map into known traversable samples, unknown-region clusters, terrain-coupled frontier viewpoints, and hypothetical samples. Sector regions encode the direction, depth, and angular width of space relative to the robot.

V. REGION SEGMENTATION AND SAMPLING

  • The module outputs traversable samples, unknown-region clusters, terrain-coupled frontier viewpoints, and hypothetical samples in unknown regions.It samples the global cost map in a sliding window centered on the robot to obtain a local traversability map.Source: §V · p. 6
  • A sector is defined by inner and outer radii, start angle, angular width, and a state distinguishing known from unknown cells.Known sectors contain free or occupied cells; unknown sectors contain unknown cells.Source: §V · p. 6
  • Sector regions preserve observation angle, radial depth, and angular width, describing the distribution of unknown space within local sensor range.Source: §V · pp. 6–7

A. Recursive Sector Region Splitting

Recursive sector splitting adapts local regions to known–unknown boundaries through state transitions and state flipping. Known-region sampling then uses sector geometry and traversability checks to generate samples.

A. Recursive Sector Region Splitting

  • Recursive splitting detects state transitions while scanning sectors in polar coordinates and subdivides the outer region for further processing.If the initial state yields no valid samples, the algorithm flips the state and recomputes; this supports detailed segmentation of complex boundaries.Source: A. Recursive Sector Region Splitting · p. 7

1) Known-Region Sampling:

  • Known-region samples use polar sector geometry, with fixed radial steps and angular intervals that narrow with increasing radius.This design produces denser angular coverage at longer range.Source: 1) Known-Region Sampling · p. 7
  • A sample is retained only if it meets the exploration-boundary constraint, lies in a free cell, has no nearby occupied cells, and remains within local-map range.The corresponding grid-cell ground height and sensor height are used in the sampling calculation.Source: 1) Known-Region Sampling · pp. 7–8

2) Unknown-Region Filtering:

The global unknown-sector set is updated incrementally, and filtering removes regions that provide insufficient coverage or observation value, including isolated holes.

2) Unknown-Region Filtering:

  • New unknown sectors replace historical sectors within their γs-neighborhood, and an unblocked inner boundary moves the representative point closer to known space.Source: 2) Unknown-Region Filtering · p. 8
  • Regions with less than a 30% radial unknown ratio are removed, avoiding a full-cell traversal of each candidate region.Source: 2) Unknown-Region Filtering · p. 8
  • A region behind an obstacle is removed when all three outward rays are blocked, while an isolated hole is removed when eight rays reach known cells within a fixed distance.Source: 2) Unknown-Region Filtering · p. 8
  • Filtering also removes regions with too small a coverage radius, too narrow an angular span, or centers too close to obstacles.Source: 2) Unknown-Region Filtering · p. 8

C. Terrain-Coupled Frontier Viewpoints and Unknown Region Clustering

Terrain-coupled viewpoints connect visible frontiers to reachable ground, while incremental clustering organizes unknown sectors into stable region targets for planning.

1) Viewpoint Extraction:

  • Frontier viewpoints are retained only when visibility and terrain-reachability constraints both hold; their gain counts associated frontiers.The reachability test constrains the viewing direction relative to the local ground plane.Source: 1) Viewpoint Extraction · p. 8

2) Unknown Region Clustering:

  • Sector clusters persist across planning cycles by reusing nearby historical centers, then grouping unassigned sectors through obstacle-free adjacency and merging compatible clusters.Cluster centers and sector sets form the unknown-region targets used for sector-guided planning.Source: 2) Unknown Region Clustering · pp. 8–9
  • Unknown clusters supply roadmap hypotheses through a cluster-center sample and up to five region-center samples, limiting redundant vertices in large clusters.Clusters with fewer than five regions use all region centers; larger clusters use farthest-point sampling.Source: 2) Unknown Region Clustering · p. 9

VI. TOPOLOGICAL ROADMAP AND EXPLORATION PLANNING

The dynamic roadmap combines verified traversable vertices with temporary unknown-space hypotheses, updating and sparsifying the graph while preserving useful connectivity.

A. Dynamic Roadmap Maintenance

  • The roadmap connects verified known-space samples and temporary unknown-space hypotheses, inserting samples only when sufficiently separated and linking collision-free neighbors.Its two vertex sets represent explored traversable space and possible connectivity toward unknown regions.Source: A. Dynamic Roadmap Maintenance · p. 9

1) Unknown Vertex Updating:

  • Unknown vertices transfer to the known set when free, are removed when occupied or obstacle-crossing, and survive reachability pruning only if connected to known space within three steps.This removes isolated hypotheses while retaining useful connectivity near the exploration boundary.Source: 1) Unknown Vertex Updating · pp. 9–10

2) Local Graph Pruning:

  • Near the robot, pruning removes leaves and bypassable degree-two vertices but retains junctions, preserving local branches and bridges for narrow-space paths.The local window within radius rmax is exempt from global sparsification.Source: 2) Local Graph Pruning · p. 10

3) Global Graph Merging:

  • Remote known and unknown vertices are merged separately in progressively finer voxels, with blocked merges deferred and bridge vertices restored to preserve major topology.Open regions are compressed more aggressively while obstacle boundaries and complex passages retain finer resolution.Source: 3) Global Graph Merging · p. 10

1) Guided Path Generation:

The planner orders unknown regions using a sector-guided route, then inserts reachable frontier viewpoints without changing that regional order. In maze-like scenes, it reports near-best coverage with less exploration time than the second-best method.

1) Guided Path Generation:

  • Unknown-region gain combines angular breadth and radial depth, and A* roadmap paths use terrain cost and heading consistency to guide region ordering.The final TSP orders region centers from the robot’s current position.Source: 1) Guided Path Generation · pp. 10–11

2) Viewpoint-Augmented Planning:

  • Frontier viewpoints are assigned to reachable region centers and inserted at the lower-cost interval, preserving the region order while adding local observations.The method contrasts this organization with viewpoint-only and mixed region-viewpoint TSP approaches.Source: 2) Viewpoint-Augmented Planning · p. 11

VII. SIMULATION BENCHMARK RESULTS AND ANALYSIS

  • The benchmark covers seven scenes spanning narrow passages, complex connectivity, open terrain, and elevation variation, with completion defined as coverage above 95%.Each of six state-of-the-art planners is tested 10 times per scene using exploration time, distance, final coverage, and average velocity.Source: §VII · pp. 11–12

1) Classical Maze-Like Scenes:

  • 19.6% and 16.0%: TASG-Explore reduces exploration time versus the second-best method in Classical Parking and Large Maze while keeping coverage near the best level.Average exploration times are 264.5 s and 631.4 s, respectively.Source: 1) Classical Maze-Like Scenes · p. 12

2) Uneven Open Environments:

Across uneven open and large-scale scenes, the reported results favor TASG-Explore in exploration time, coverage, and terrain analysis; its roadmap and planning mechanisms target both efficient expansion and local terrain detail.

2) Uneven Open Environments:

  • 48.5% and 23.8%: TASG-Explore reduces exploration time versus the second-best method in Small Terrain and Complex Terrain, completing them in 58.6 s and averaging 141.7 s.The comparison concerns uneven open environments where 2-D distance does not capture true motion cost.Source: 2) Uneven Open Environments · p. 12, Table III · p. 13

3) Large-Scale Complex Environments:

4) Planning Efficiency:

  • In most scenes, per-iteration runtime stays below 100 ms with a concentrated distribution and small variance; Rugged Hill shows no obvious long tail.Fig. 12 reports runtime distributions over 10 repeated experiments in seven scenes.Source: 4) Planning Efficiency · pp. 13–14, Figure 12 · p. 15
  • Sector targets, near-dense/far-sparse roadmap maintenance, and viewpoint insertion after regional ordering control planning scale while retaining local observations.The organization separates global coverage decisions from supplementary local observations.Source: 4) Planning Efficiency · p. 14

B. Ground Traversability Analysis Evaluation

  • 0.868, 0.954, and 0.856: Average Trav. IoU, Cov. Ratio, and Acc. Ratio; average runtime is 15.29 ms, about 6.3 times faster than the fastest baseline average of 96.05 ms.The method uses variable-resolution ground fitting and 8-bit height codes to represent traversable ground and local obstacles.Source: B. Ground Traversability Analysis… · pp. 15–16, Table IV · p. 16

VIII. ABLATION STUDY

The ablations isolate how unknown-region processing, roadmap maintenance, and hierarchical path organization affect exploration. Together, they show that region-level guidance and local viewpoints serve complementary roles, while roadmap and clustering choices affect exploration cost and connectivity.

VIII. ABLATION STUDY

  • The ablations use the same simulator and metrics, changing only the tested design component to isolate its contribution.Source: §VIII · p. 16

1) Unknown-Region Processing Ablation:

  • Keeping all unknown regions increases exploration time and per-iteration runtime in both scenes; Euclidean clustering raises Classical Parking exploration time by 13.3% and movement distance by 9.7%.Euclidean clustering lacks historical inheritance, so target centers can become unstable in narrow passages.Source: 1) Unknown-Region Processing Ablation · p. 16

2) Roadmap Update Ablation:

  • Without unknown vertices, exploration time increases by 7.5% and 17.4% across the two scenes, while disabling dynamic roadmap management increases runtime and allows redundant nodes to accumulate.The proposed roadmap remains visibly sparser as coverage grows while maintaining stable update speed.Source: 2) Roadmap Update Ablation · p. 16

3) Path-Organization Ablation:

  • The proposed planner first orders sector regions, then inserts local viewpoints; replacing this hierarchy with frontier-only or mixed-target TSPs increases exploration time or movement distance.F-TSP raises exploration time by 49.8% and movement distance by 27.2% in Classical Parking; RF-TSP raises exploration time there by 22.3%.Source: 3) Path-Organization Ablation · p. 16

B. Ground Traversability Ablation

Real-world trials demonstrate online exploration across open lawn, uneven woodland, and large rugged terrain, while the ablations distinguish the roles of variable voxels and 8-bit obstacle encoding. The conclusion notes that the supporting-ground representation is not suited to vertically overlapping traversable levels.

B. Ground Traversability Ablation

  • Removing 8-bit height encoding leaves coverage ratio high but lowers Trav. IoU and Acc. Ratio, especially in Uneven Forest, where they are 0.182 and 0.167.The full method records obstacle occupancy within each cell, improving obstacle classification without a large computation increase.Source: Table VI · p. 17, B. Ground Traversability Ablation · p. 17
  • Fixed-voxel fitting lowers average Trav. IoU by 0.509, with Uneven Forest reaching only 0.041; variable voxels adapt to open ground and dense obstacles.The full method is slightly slower than simplified variants, with RT between 11.30 ms and 17.86 ms, while segmentation performance is best or near best.Source: B. Ground Traversability Ablation · p. 17

IX. REAL-WORLD EXPERIMENTS

  • The method replans online across three outdoor scenes, with average iteration times of 13.89 ms, 17.14 ms, and 20.71 ms, and trajectories that follow traversable areas.The largest rugged scene completes in 1524 s over a 757 m trajectory; the lawn and woodland scenes complete in 365 s and 835 s, respectively.Source: §IX · pp. 18–19, Figure 17 · p. 19

X. CONCLUSION

  • The conclusion combines hierarchical traversability analysis, incremental sector segmentation, and dynamic topological-roadmap planning for large-scale exploration.The roadmap estimates unknown connectivity, while sector-guided planning supports efficient global exploration.Source: §X · p. 19
  • The method is designed for outdoor scenes represented by a supporting ground and is not well suited to vertically overlapping traversable structures.Supporting multi-level scenes remains future work.Source: §X · p. 19

Related papers

Loading 2609.08512v2…