PaperlayerView exact revision on arXiv

Source-linked AI summary

Centralized Multi-UAV Exploration and 3D Reconstruction Using Single-UAV Planners

João Félix Mendes, Meysam Basiri, Rodrigo Ventura

arXiv:2609.40208v1cs.RO

TL;DR

Extending single-UAV exploration to multiple UAVs raises shared-mapping and safety challenges. This paper adapts existing sampling-based planners within a centralized framework, and simulations find that Separated Starts deliver faster exploration and improved coverage than Joint Starts.

  • Problem

    Extending single-UAV exploration to multi-UAV teams raises challenges in consistent mapping, safe navigation, and coordination.

    Source: §I · p. 1, A. System Architecture · p. 3

  • Method

    A centralized framework adapts voxblox for shared multi-UAV TSDF mapping and integrates collision avoidance and robot filtering while retaining planners’ core logic.

    Source: §I · p. 2

  • Results

    Separated Start deployments achieve faster exploration and improved coverage than Joint Starts across the evaluated planners.

    Source: B. Simulated Experiments · p. 6, §V · p. 8

  • Takeaways & Limitations

    Initial deployment strategy plays a critical role in multi-UAV exploration performance even with identical planners and without explicit task allocation.

    Source: §V · p. 8

Abstract

from arXiv · show

Extending single Unmanned Aerial Vehicles (UAVs) exploration methods to multi-UAV teams can improve coverage speed and robustness, but introduces challenges such as consistent mapping, safe navigation, and deployment strategy. In this work, we present a centralized multi-UAV exploration framework that enables the use of existing single-UAV sampling-based planners in a multi-UAV setting. The proposed architecture allows multiple UAVs to collaboratively explore unknown environments using a shared global Truncated Signed Distance Field (TSDF) map and centralized planning. Building on the voxblox library, we adapt its mapping pipeline to support real-time fusion of depth measurements from multiple UAVs into a common TSDF representation. In addition, inter-UAV collision avoidance and robot self-filtering mechanisms are integrated into the system to ensure safe navigation and prevent reconstruction of other UAVs as static obstacles. The framework is evaluated in simulation using four sampling-based exploration planners - RH-NBVP, KRH-NBVP, AEP, and KAEP - whose core sampling logic is preserved, with only system-level adaptations for multi-UAV operation. Experiments are conducted across multiple environments and under two deployment configurations: Joint Start (JS), where UAVs are initialized in close proximity, and Separated Start (SS), where UAVs are initialized in distinct locations. Results show that SS deployments consistently achieve faster exploration and improved coverage across all planners, highlighting the importance of the deployment strategy in multi-UAV exploration performance.

I. INTRODUCTION

The paper extends single-UAV sampling-based exploration to a centralized multi-UAV framework, addressing shared mapping and safety challenges while retaining existing planners’ core logic. Its simulations report faster exploration and improved coverage with Separated Starts than Joint Starts.

I. INTRODUCTION

  • Centralized planning provides access to global information for map fusion and conflict-aware coordination across agents.The paper presents this architecture as a way to study multi-UAV behavior and extend single-UAV planners without explicit coordination methods.Source: §I · pp. 1–2
  • Single-UAV methods do not directly provide shared mapping, inter-UAV collision avoidance, or protection against reconstructing other UAVs as obstacles.These gaps complicate multi-UAV deployment of planners such as RH-NBVP and require careful multi-sensor fusion to preserve map consistency.Source: §I · p. 1
  • Separated Starts consistently achieve faster exploration and improved coverage than Joint Starts across the evaluated planners.The study compares four planners and defines JS as close-proximity initialization and SS as spatially distinct initialization.Source: §I · p. 2
  • The framework adapts voxblox for real-time fusion of multiple UAVs’ depth measurements into a shared TSDF map.It also integrates collision avoidance and pointcloud filtering, while the planners’ core sampling logic remains unchanged.Source: §I · p. 2
  • The work evaluates centralized map fusion and the effect of deployment strategies on exploration efficiency.The stated experimental evaluation covers accuracy and real-time performance of map fusion.Source: §I · p. 2

II. RELATED WORKS

Prior exploration work includes frontier-based and sampling-based planners, while multi-robot approaches often require specifically designed coordination. This paper focuses on adapting established single-UAV planners and voxblox for centralized multi-robot exploration.

II. RELATED WORKS

  • Frontier-based exploration can miss informative viewpoints deeper in free space because it is limited to frontier regions.Prior extensions cluster frontier voxels to reduce computation or constrain selection to the sensor field of view for high-speed aerial exploration.Source: §II · p. 2
  • Sampling-based planners evaluate candidate viewpoints in free space, with methods spanning RRT-based, global-coverage, and trajectory-centric approaches.Examples include RH-NBVP, RRT*-inspired planners, KRH-NBVP, and KAEP.Source: §II · p. 2
  • Multi-robot work commonly uses task allocation or decentralized planning, while direct adaptation of single-robot pipelines without changing core planning logic has received less attention.Centralized architectures can aggregate system knowledge and enable reuse of single-UAV planners such as RH-NBVP.Source: §II · p. 2
  • Existing sub-mapping and pose-graph mapping extensions can impose computational overhead, leaving lightweight real-time multi-robot mapping underexplored.Native voxblox is designed for single-sensor operation, despite its accuracy and computational efficiency in single-UAV frameworks.Source: §II · pp. 2–3
  • This work adapts voxblox for real-time multi-robot depth fusion and evaluates deployment strategies without modifying underlying planning algorithms.It contrasts this approach with prior work that proposes new exploration algorithms or explicit coordination strategies.Source: §II · p. 3

III. PROPOSED APPROACH

The proposed centralized architecture combines onboard sensing and mapping with server-side fusion and planning. It addresses shared mapping, collision avoidance, and prevention of mutual UAV reconstruction.

III. PROPOSED APPROACH

  • Scaling centralized multi-robot exploration requires a shared global map, collision-free motion among robots, and filtering that prevents robots from becoming mapped obstacles.These are identified as three fundamental system challenges.Source: A. System Architecture · p. 3
  • Each UAV filters other agents from depth data, integrates the pointcloud into a local voxblox sub-map, and periodically sends incremental updates for server fusion.Peer-to-peer state sharing supports onboard filtering; the ground station fuses the updates into a unified map.Source: A. System Architecture · p. 3, Figure 2 · p. 3
  • When a UAV nears its current goal, it requests a path asynchronously; the server plans using the global map and shared active-path cache to avoid inter-UAV collisions.The resulting collision-aware path is sent to the onboard controller for execution.Source: A. System Architecture · p. 3, Figure 2 · p. 3
  • The implementation reuses core inter-robot collision-avoidance logic from an existing codebase but redesigns its integration for centralized mapping and planning.The cited original publication addresses single-robot exploration, although its codebase includes multi-robot structures.Source: A. System Architecture · p. 3

B. Background on Evaluated Planners

The evaluated planners cover receding-horizon and hybrid local-global strategies, with kinodynamic variants producing dynamically feasible trajectories. Their planner-specific mechanisms provide distinct approaches to viewpoint selection and exploration guidance.

B. Background on Evaluated Planners

C. Centralized Multi-Robot Mapping

The framework extends voxblox to fuse multiple UAVs’ local TSDF maps into a centralized map for planning, prioritizing computational efficiency for real-time multi-UAV exploration.

C. Centralized Multi-Robot Mapping

  • The extension addresses voxblox’s native lack of direct multi-robot mapping support by enabling volumetric map information from multiple robots and sensors to be fused.Voxblox provides efficient pointcloud integration and accurate reconstruction, but its native architecture is limited to single-sensor integration.Source: C. Centralized Multi-Robot Mapping · p. 3
  • Local voxblox TSDF sub-maps are periodically merged into a centralized map used for planning.Each UAV integrates high-frequency sensor data locally before fusion; locally integrated sub-maps make fusion less sensitive to small temporal misalignments or network latency, assuming standard synchronization such as NTP.Source: C. Centralized Multi-Robot Mapping · p. 4
  • Fusion combines corresponding voxel weights and computes a weighted average of signed distances, preserving accumulated confidence as observations are incrementally integrated.Each voxel stores a signed distance to the nearest surface and a weight reflecting measurement reliability.Source: C. Centralized Multi-Robot Mapping · p. 4
  • Compared with c-blox and voxgraph, the centralized fusion strategy prioritizes computational efficiency for real-time multi-UAV mapping.The passage does not report a quantitative comparison.Source: C. Centralized Multi-Robot Mapping · p. 4

D. Multi-Robot Collision Avoidance

The planner checks candidate paths against static surfaces in the centralized TSDF map and other UAVs’ planned paths, pruning paths that violate the minimum clearance.

D. Multi-Robot Collision Avoidance

  • Candidate paths are checked against static obstacles using the centralized TSDF and against other UAVs’ shared planned paths.For inter-robot checks, the minimum distance between points on the respective paths is computed.Source: D. Multi-Robot Collision Avoidance · p. 4
  • Paths closer than dcollision to either the static environment or another robot are marked non-traversable and pruned from the planning tree.Only collision-free paths are retained for evaluation.Source: D. Multi-Robot Collision Avoidance · p. 4

E. Multi-Robot Pointcloud Filtering

Pointcloud filtering removes observations around other UAVs before map integration, preventing their reconstruction as static obstacles and preserving the representation of free space.

E. Multi-Robot Pointcloud Filtering

  • Dynamic UAV observations can become persistent false obstacles in the map, unnecessarily constraining future paths and reducing traversable free space.Source: E. Multi-Robot Pointcloud Filtering · p. 4
  • Before integration, the system removes pointcloud data within a dcollision-radius sphere centered at each other robot’s estimated pose.Because planning enforces the same minimum clearance from real obstacles, the filter does not remove valid measurements of those obstacles.Source: E. Multi-Robot Pointcloud Filtering · p. 4
  • With filtering enabled, the TSDF visualization no longer shows regions from mutual UAV reconstruction and reflects only the static ground plane.Without filtering, cyan and green regions around each UAV correspond to observations of the other UAV; filtering removes these regions in the two-UAV demonstration.Source: E. Multi-Robot Pointcloud Filtering · p. 4

IV. EXPERIMENTAL EVALUATION

The evaluation tests four sampling-based planners in simulation across three environments and multiple UAV initialization strategies. It reports exploration, reconstruction, path-length, and velocity measures while retaining the single-UAV study’s planner settings.

IV. EXPERIMENTAL EVALUATION

  • The evaluation uses three DJI F450 UAVs in simulation, with simulated RTK-GPS state estimation and Model Predictive Controller trajectory tracking.Exploration is decoupled from localization by assuming sufficiently accurate state estimation; the framework can use any localization backend if pose estimates are sufficiently consistent for TSDF integration.Source: §IV · pp. 4–5
  • Four sampling-based planners—RH-NBVP, AEP, KRH-NBVP, and KAEP—are evaluated across a maze, police station, and school environment.The maze tests narrow-passage navigation; the police station represents compact reconstruction in an open-air environment, and the school is a larger, more geometrically complex task.Source: §IV · p. 5
  • Separated Start initializes UAVs in distinct regions, while Joint Start places them in the same region; the maze evaluates only Separated Start.The study describes Joint Start in the maze as forcing single-file motion and redundant exploration; both configurations are evaluated in the police station and school.Source: §IV · p. 5, Figure 4 · p. 5
  • Evaluation measures exploration rate at 25%, 50%, and 95% completion, Full Exploration, total path length, and average velocity, with five repeated experiments reporting means and standard deviations.In the multi-UAV setting, total path length sums distances traveled by all UAVs, and average velocity is the mean of individual UAV average velocities.Source: §IV · p. 5
  • Planner, simulation, and sensor settings match the single-UAV study, with deployment configuration as the addition; the maze uses dreplan = 0.4 m and dcollision = 0.7 m.The passage states these tighter maze constraints accommodate narrow passages; Table I is identified as summarizing key experimental parameters unchanged from.Source: §IV · p. 5, Table I · p. 6

A. Multi-Robot Map Fusion Accuracy and Performance

The centralized map preserves environment geometry at roughly voxel-scale error while distributing mapping work between UAVs and the central server. Centralized fusion remains fast in the reported school-environment measurements.

A. Multi-Robot Map Fusion Accuracy and Performance

  • Map accuracy and processing times were evaluated for a three-UAV KAEP team in the police station and school environments under Joint Start.The experiments ran on a workstation with an Intel Core i9 (14th Gen), 32 GB RAM, and an NVIDIA GeForce RTX 4060 GPU.Source: A. Multi-Robot Map Fusion Accuracy and… · p. 6
  • The centralized TSDF map achieves RMSE on the order of the 0.2 m voxel resolution in both evaluated environments.The authors report that this preserves overall geometric structure for motion planning.Source: A. Multi-Robot Map Fusion Accuracy and… · p. 6
  • In the school environment, onboard local TSDF and ESDF updates average 87.0 ms and 10.0 ms, while centralized fusion takes 8.6 ms and 10.1 ms.Local integration times are a conservative upper bound because the workstation also runs the simulation, multiple mapping pipelines, and central server.Source: A. Multi-Robot Map Fusion Accuracy and… · p. 6
  • The mapping workload is split between onboard high-frequency pointcloud integration and server-side processing of incremental sub-map updates.The authors report that this offloading prevents the central server from becoming a computational bottleneck and enables real-time multi-UAV exploration.Source: A. Multi-Robot Map Fusion Accuracy and… · p. 6

B. Simulated Experiments

Across the simulations, centralized multi-UAV exploration accelerates coverage and mitigates some single-UAV failure modes, while initial deployment strongly affects performance. Separated Starts reduce early overlap, and planner speed rankings remain consistent across environments.

B. Simulated Experiments

  • KAEP achieves the fastest exploration rates, followed by KRH-NBVP, AEP, and RH-NBVP across environments, deployment strategies, and coverage stages.The reported stages include E25%, E50%, and E95%.Source: B. Simulated Experiments · p. 6
  • Separated Start is faster than Joint Start at E25%, E50%, and E95% in all environments, with a larger effect in larger environments.Separated Starts reduce early-stage overlap by placing UAVs in spatially distinct regions; Joint Starts produce redundant initial reconstruction.Source: B. Simulated Experiments · p. 6
  • In the maze, multi-UAV deployment enables RH-NBVP and KRH-NBVP to achieve full exploration despite prior single-UAV local-minimum failures.The authors attribute the improvement to implicit spatial task allocation under Separated Start, without explicit inter-robot coordination.Source: B. Simulated Experiments · p. 6
  • KAEP has the highest average velocity, followed by KRH-NBVP, AEP, and RH-NBVP, while multi-UAV average velocities are generally lower than in single-UAV experiments.The authors relate the lower velocities, particularly in the maze, to back-and-forth movement within assigned exploration sections.Source: B. Simulated Experiments · p. 6
  • Centralized multi-UAV exploration accelerates coverage and mitigates single-UAV failure modes, with initial deployment playing a significant role in performance.The conclusion is supported by faster exploration under Separated Start and maze exploration gains for planners prone to local minima.Source: B. Simulated Experiments · p. 6

V. CONCLUSION

The framework extends established single-UAV planners to centralized multi-UAV exploration through mapping and safety integrations. In simulation, separated starts outperform joint starts across the tested planners, while centralized control has reliability and scalability trade-offs.

V. CONCLUSION

  • The framework adapts voxblox for shared TSDF mapping and adds collision avoidance and point-cloud filtering to prevent reconstructing UAVs as static obstacles.The mapping and safety components support multi-UAV operation within the centralized planning pipeline.Source: §V · p. 7
  • SS initialization yields faster exploration and improved coverage than JS across RH-NBVP, KRH-NBVP, AEP, and KAEP.The deployment comparison holds across the four evaluated planners.Source: §V · p. 8
  • Centralized control introduces a single point of failure and limits scalability as the number of UAVs increases.The conclusion identifies real-world multi-UAV experiments and decentralized planning as future directions.Source: §V · p. 8

Related papers

Loading 2609.40208v1…