PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Map Sharing and Capability-Aware Coverage Planning for AI-Native 6G Robotic Coordination

Abdulqader Dhafer, Qi Wang, Zhou Daniel Hao

arXiv:2609.37666v1cs.ROcs.AI

TL;DR

Heterogeneous SAR coverage planning must connect terrain semantics to platform-specific reachability, while continuous image exchange is costly. The paper combines semantic grids, reachability-constrained planning, and incremental state sharing; it reports higher coverage with no infeasible allocations than LS-MCPP, alongside reduced payload in the high-update case.

  • Problem

    Conventional coverage methods often do not translate perceived terrain into platform-specific reachability, while continuous image exchange imposes high communication cost.

    Source: §I · p. 1

  • Method

    The framework converts aerial observations into semantic grids and couples capability-constrained area decomposition and path planning with incremental semantic-map corrections.

    Source: §I · pp. 1–2, 1) Reachability-Constrained… · p. 3

  • Results

    91.5% coverage with no capability-infeasible allocations, versus LS-MCPP’s 78.8% coverage and 21.5% capability-infeasible allocation rate; high-update semantic corrections reduce payload approximately 82× versus periodic full-map sharing.

    Source: C. Quantitative Planner Benchmark · p. 5, §VI · p. 6, E. Application-Payload Analysis for… · p. 6

  • Takeaways & Limitations

    Semantic corrections update the shared planning state without requiring repeated transmission of the complete map.

    Source: §VI · p. 6

Abstract

from arXiv · show

Search and Rescue (SAR) operations increasingly deploy heterogeneous teams of aerial and ground robots. However, conventional coverage methods typically do not translate perceived terrain into platform-specific reachability, while continuous image exchange imposes a high communication cost. We propose an edge-centric, semantic-aware coverage planning framework that integrates aerial terrain perception, robot-specific traversability reasoning, and payload-efficient semantic state sharing. Aerial observations are converted into compact semantic grid maps, enabling reachability-constrained area decomposition and capability-aware coverage paths that assign only regions admitted by each robot's capability profile. The resulting perception-sharing-planning loop feeds semantic corrections into traversability reasoning and replanning, forming an application-level mechanism motivated by AI-enabled goal-oriented communication envisioned for AI-native 6G networks. For the high-update case, transmitting semantic corrections reduces the application payload by a factor of approximately $82$ relative to periodic full-map sharing. Across matched benchmark scenarios, the proposed method achieved $91.5\%$ coverage with no capability-infeasible allocations, compared with $78.8\%$ coverage and a $21.5\%$ capability-infeasible allocation rate for LS-MCPP. Semantic corrections update the shared planning state without requiring repeated transmission of the complete map.

I. INTRODUCTION

Heterogeneous SAR teams need coordination that connects terrain perception to robot-specific coverage assignments and shares mission-relevant state without continuous imagery. The paper proposes an edge-centric framework coupling semantic perception, reachability-constrained planning, and incremental semantic-map corrections.

I. INTRODUCTION

  • SAR teams combining aerial and ground robots can face spatial mismatches, infeasible assignments, and redundant exploration when terrain exceeds some platforms’ safe reach.Conventional coverage planning often assumes homogeneous mobility and centralized global maps, unlike distributed heterogeneous SAR deployments.Source: §I · p. 1
  • Existing work often treats semantic perception, multi-robot coverage, and communication efficiency separately, leaving a gap in translating terrain semantics into robot-specific assignments and sharing resulting state.Standard planners may use a shared terrain graph without enforcing robot-specific traversability constraints.Source: §I · p. 1
  • The framework couples semantic perception, capability-aware coverage planning, and incremental semantic-state sharing for heterogeneous SAR teams.Its planning method embeds platform-specific traversability constraints in geodesic expansion, while sharing incremental semantic-map corrections rather than continuous imagery.Source: §I · pp. 1–2

II. RELATED WORK

Prior work advances disaster-scene semantic perception and multi-robot coverage, but the reviewed approaches do not translate perception-derived terrain labels into robot-specific reachability constraints for coverage planning.

A. Perception and Semantic Mapping for Disaster Robotics

  • Disaster datasets support semantic segmentation of damage, debris, and flooding, but existing work generally uses those labels for damage assessment rather than robot-specific traversability.Building-focused annotations in xBD do not describe local access conditions needed for ground navigation.Source: A. Perception and Semantic Mapping for… · p. 2

B. Coverage and Path Planning for Heterogeneous MultiRobot Systems

  • Classical and multi-robot coverage planners balance workload or reduce makespan, yet the reviewed methods do not encode perception-derived, robot-specific traversability during allocation.The paper positions its framework as embedding platform-specific traversability directly into semantic-grid decomposition.Source: B. Coverage and Path Planning for… · p. 2

C. AI-Native 6G Semantic Coordination

The paper frames semantic coordination as task-relevant state exchange within an edge-supported distributed perception and planning loop. It represents aerial observations as compact semantic grids and uses them for capability-aware planning.

C. AI-Native 6G Semantic Coordination

  • Semantic communication shifts exchange from reconstructing source data toward delivering task-relevant information for downstream decisions in dynamic SAR environments.The paper situates this approach within a 6G vision of communication and edge computing supporting distributed AI perception and planning.Source: C. AI-Native 6G Semantic Coordination · p. 2

III. PROPOSED APPROACH

  • The framework converts aerial observations into semantic grid maps for reachability-constrained decomposition, local path planning, and incremental state sharing.The shared grid is constructed from aerial imagery using YOLOv11 segmentation; each cell receives its most frequent predicted class, while unlabeled cells are treated as impassable and excluded from the coverage denominator.Source: §III · p. 2, A. Semantic Environment Representation · p. 2

B. Reachability-Constrained Area Decomposition and Path Planning

Coverage regions and paths are formed from semantic terrain information together with each robot’s traversability constraints. A discrete geodesic Voronoi procedure assigns reachable cells while leaving cells unreachable by every robot unassigned.

1) Reachability-Constrained Decomposition:

  • Robot-specific geodesic Voronoi decomposition computes distances on the semantic grid while respecting obstacles and each platform’s terrain traversability.Breadth-First Search expands only through cells admitted by the robot’s capability set, and reachable cells are assigned by minimum geodesic cost.Source: 1) Reachability-Constrained… · p. 3

B. Reachability-Constrained Area Decomposition and Path Planning

2) Local Coverage Path Planning:

Each robot receives an initial path for its assigned region, then executes and revises it locally as observations reveal terrain or obstacle changes. Incremental nearest-cell selection and capability-constrained A* support coverage of assigned reachable cells while reducing backtracking.

2) Local Coverage Path Planning:

  • New terrain or obstacle observations trigger local replanning, while onboard observations update the semantic map during execution.The robot executes its initial path locally using the semantic grid map.Source: 2) Local Coverage Path Planning · p. 3
  • Capability-constrained A* generates feasible routes to the nearest uncovered assigned cell, selected by Manhattan distance.Source: 2) Local Coverage Path Planning · p. 3
  • Uncovered assigned cells encountered along path segments are marked covered, enabling implicit path stitching that reduces backtracking and traversal overhead.Source: 2) Local Coverage Path Planning · p. 3

C. Goal-Oriented Semantic State Sharing

The framework shares compact semantic-cell corrections rather than raw imagery or complete map broadcasts, with local replanning when observations conflict with a robot’s map. Its segmentation dataset combines three sources to represent varied terrain and disaster conditions.

C. Goal-Oriented Semantic State Sharing

  • Only affected cells are shared when traversability changes or obstacles are found, and robots replan locally when observations conflict with potentially incomplete or outdated maps.The authors leave robustness to communication delays and inter-agent map inconsistencies for future work.Source: C. Goal-Oriented Semantic State Sharing · p. 3
  • Semantic-cell updates encode affected locations and revised states for integration into each agent’s local map, avoiding repeated complete-map broadcasts.Source: C. Goal-Oriented Semantic State Sharing · p. 3

IV. DATASETS

  • The custom dataset combines three open-source sources into 5,234 images, split 80/10/10 across training, validation, and testing.RescueNet, the Semantic Segmentation Satellite Imagery Dataset, and FloodNet contribute complementary disaster imagery and annotations.Source: §IV · p. 4

A. Data Pre-processing

The preprocessing pipeline harmonizes source annotations into a 10-class terrain taxonomy designed to preserve distinctions relevant to traversability. On 523 held-out images, segmentation reached a micro F1-score of 0.859, with weaker performance for sparsely represented classes and oblique views.

A. Data Pre-processing

  • Annotations with partially overlapping label granularities were mapped to a common taxonomy of 10 terrain classes for robot-specific traversability reasoning and coverage planning.The taxonomy retains distinctions in structural damage and terrain accessibility.Source: A. Data Pre-processing · p. 4, Table I · p. 4

A. Segmentation Model

  • The segmentation model was trained on a composite dataset spanning heterogeneous spatial scales and viewpoints typical of aerial disaster assessment.Source: A. Segmentation Model · p. 4
  • 0.859 micro F1-score was achieved on 523 held-out test images, alongside precision of 0.883 and recall of 0.837.Sparsely represented classes and building-damage labels under oblique views had lower performance.Source: A. Segmentation Model · p. 4

B. Isaac Sim Feasibility Evaluation

Across the common benchmark configurations, the proposed method achieved higher coverage than the baselines with no capability-infeasible allocations. In Isaac Sim trials, semantic-map updates and local replanning corrected observed traversal conflicts.

B. Isaac Sim Feasibility Evaluation

  • The Isaac Sim scene represented a disaster environment with partially collapsed buildings and debris-obstructed roads for a controlled feasibility evaluation.Figure 3 depicts aerial and ground robots with a discretized semantic grid.Source: B. Isaac Sim Feasibility Evaluation · p. 5, Figure 3 · p. 5
  • Fewer than 10% of assigned cells had local traversal conflicts across ten Isaac Sim trials, and semantic-map updates with local replanning corrected them.Conflicts primarily arose when dominant-class aggregation concealed non-traversable content in cells labeled feasible.Source: B. Isaac Sim Feasibility Evaluation · p. 5

C. Quantitative Planner Benchmark

  • 91.5% coverage with no capability-infeasible allocations was achieved by the proposed method, versus 78.8% coverage and 21.5% capability-infeasible allocation for LS-MCPP.The comparison is over the common configuration set reported in Table II.Source: Table II · p. 5, C. Quantitative Planner Benchmark · p. 5
  • The benchmark evaluated planners on identical 20 × 20 semantic grids across 814 randomized configurations using one ground robot and one UAV.Of 900 attempted configurations, 85 were excluded for DARP failures or non-convergence, and one was incompatible with MFC’s contracted-root representation.Source: C. Quantitative Planner Benchmark · p. 5

D. Discretization Resolution Analysis

Grid resolution trades semantic fidelity against planning makespan: finer grids conceal less non-traversable area but increase path length. The study uses a 20×20 grid in simulation and notes that suitable dimensions depend on covered area and altitude.

D. Discretization Resolution Analysis

  • At 40×40 versus 8×8, intra-cell impurity falls from 15.2% to 3.3% and hidden non-traversable fraction from 7.0% to 1.5%.Figure 4 summarizes the resolution sweep across 80 scenes.Source: D. Discretization Resolution Analysis · p. 6, Figure 4 · p. 6
  • Across the 8×8–40×40 sweep, maximum per-robot path length rises from roughly 59 to 1,380 grid-cell steps.The paper reports that planning makespan increases rapidly with resolution.Source: D. Discretization Resolution Analysis · p. 6, Figure 4 · p. 6
  • The simulations use a 20×20 grid, with 6.9% impurity and 3.1% hidden non-traversable fraction.The appropriate resolution is not fixed; it scales with covered area and altitude, with grid dimensions increased at higher altitude to keep represented cell area approximately consistent.Source: D. Discretization Resolution Analysis · p. 6

E. Application-Payload Analysis for Semantic State Sharing

The analysis compares image streaming, full-map sharing, and event-triggered semantic corrections by application payload. Under the stated high-update assumptions, delta updates sharply reduce payload relative to both full-map baselines, while the reported values are not end-to-end network performance.

E. Application-Payload Analysis for Semantic State Sharing

  • The mechanism sends an update only when onboard observation detects a difference from the stored semantic map, rather than repeatedly sending unchanged cells.Each update represents changed cells plus an application-message header; the expected rate depends on observation frequency and update probability.Source: E. Application-Payload Analysis for… · p. 6
  • Under the high-update assumption q = 0.70, delta updates require 49 B/s, reducing payload by approximately 57× versus event-triggered full-map sharing and 82× versus periodic full-map sharing.The event-triggered full-map baseline requires approximately 2.8 kB/s.Source: E. Application-Payload Analysis for… · p. 6
  • The payload estimates use f = 10 Hz, b = 3 B per changed cell, h = 4 B, and mean update size n̄ = 1; the modeled rate ranges from 0 to 70 B/s.The update frequency and size assumptions determine the reported application-payload values.Source: E. Application-Payload Analysis for… · p. 6
  • The reported values quantify application payload under the stated encoding and update assumptions, not end-to-end network performance.Semantic corrections update terrain state used for traversability reasoning and coverage assignment.Source: E. Application-Payload Analysis for… · p. 6

VI. CONCLUSION

The conclusion reports benchmark gains for capability-aware coverage and lower application payload from semantic corrections, then frames the loop as an application-level example of goal-oriented communication. The data-and-code section says licensing prevents redistribution of the merged dataset, while preprocessing scripts and source-access instructions are available.

VI. CONCLUSION

  • The proposed method achieved 91.5% coverage with no capability-infeasible allocations, versus 78.8% coverage and a 21.5% capability-infeasible allocation rate for LS-MCPP.These results are reported for the common benchmark.Source: §VI · p. 6
  • Under the high-update assumptions, delta updates yield 49 B/s, approximately 82× lower payload than periodic full-map sharing.A 20×20 semantic grid also reduces application payload by approximately four orders of magnitude relative to uncompressed imagery.Source: §VI · p. 6
  • The segmentation model achieved a micro F1-score of 0.859 on 523 held-out images.Source: §VI · p. 6
  • The framework uses aerial terrain labels to constrain robot-specific reachability before spatial decomposition, while incremental corrections update shared planning state.The paper presents the decision loop as an application-level example aligned with the AI-native 6G vision; physical-platform evaluation under communication delay and inter-agent map inconsistency is future work.Source: §VI · p. 6

DATA AND CODE AVAILABILITY

  • Licensing constraints prevent redistribution of the merged dataset and annotations; preprocessing scripts and instructions for obtaining the source datasets are available.Source: DATA AND CODE AVAILABILITY · p. 6
Loading 2609.37666v1…