---
title: Multi-Robot Dynamic Coverage Path Planning
url: https://www.emergentmind.com/topics/multi-robot-dynamic-coverage-path-planning-mdcpp
type: topic
---

# Multi-Robot Dynamic Coverage Path Planning

Searching arXiv for recent papers on multi-robot dynamic coverage path planning and closely related formulations.
Multi-robot Dynamic Coverage Path Planning (MDCPP) denotes a class of coverage path planning problems and methods in which multiple robots must collectively cover an area of interest while adapting to changing workloads, robot capabilities, team composition, environmental knowledge, communication constraints, or sensing objectives. In the literature represented here, MDCPP appears both as an explicit problem label and as a broader umbrella over online, decentralized, connectivity-aware, and workload-adaptive multi-robot coverage systems. Common concerns include complete coverage, balanced workload distribution, low overlap, collision avoidance, adaptation to unknown or changing environments, and practical deployability on aerial and ground robot teams [2509.23705], [2403.10460], [2210.07514].

## 1. Definition and problem scope

MDCPP extends Multi-robot Coverage Path Planning (MCPP) beyond static partitioning assumptions. A central motivation is that conventional MCPP typically assumes fixed robot velocities and static area partitioning to balance the workload among robots, whereas real deployments involve variable task demands, robot heterogeneity, dynamic failures, unknown maps, and communication limits [2509.23705]. A closely related formulation treats efficient coverage of unknown environments as requiring robots to adapt their paths in real time based on on-board sensor data [2503.00647]. Another line of work emphasizes decentralized execution so that multi-robot coverage remains robust under uncertain malfunctions [2210.07514].

Across the cited work, MDCPP-like systems differ primarily along four axes. First, they differ in what changes online: target-density-induced workload, robot capability, connectivity structure, map knowledge, or sensing geometry. Second, they differ in coordination architecture, ranging from centralized synchronizing machines to fully decentralized policies and localized coordination [2204.09726], [2011.10034]. Third, they differ in the object being allocated: cells, subareas, tours, or continuous trajectories. Fourth, they differ in whether the environment is known, partially known, or unknown, and whether they model obstacle avoidance, path deconfliction, or terrain-aware sensing explicitly [2411.01707], [2507.17519].

A recurrent objective is to reduce the completion time of the last robot while maintaining full coverage and reducing redundant traversal. This appears explicitly as coverage time in CAP and Multi-CAP, as mission completion time in online centralized planners, as workload balance under capability-weighted partitioning, and as turn- or overlap-aware makespan reduction in grid-based methods [2503.00647], [2509.14941], [2303.00047], [2212.02231].

## 2. Dynamic workload adaptation and heterogeneous capability models

One major MDCPP formulation treats dynamic adaptation as workload redistribution driven by variable robot performance. In "Coverage Control for a Multi-robot Team with Heterogeneous Capabilities using Block Coordinate Descent (BCD) Method" [2204.09726], each robot \( i \) is assigned a non-negative capability factor \( C_i \), representing remaining energy, mobility speed, or sensing strength. The paper uses \( \ell_1 \)-normalized capabilities,
\[
c_i = \frac{C_i}{\sum_{k=1}^n C_k}, \quad \forall i \in \mathcal{N},
\]
and aims to allocate portions of the environment proportionally to robot capabilities so that a robot with higher capability takes a larger portion of the environment while a robot with lower capability takes a smaller one [2204.09726].

The partitioning mechanism in that formulation is a power diagram, with each cell \( W_i \) defined by a position \( p_i \) and weight \( w_i \). The stated objective is to match area ratios to capability ratios,
\[
\frac{A_i}{A_j} = \frac{C_i}{C_j}, \quad \forall i,j,
\]
through the cost
\[
H = \sum_{i=1}^n \left( \frac{A_i}{c_i} - A \right)^2.
\]
The approach alternates between optimizing generator positions and generator weights via block coordinate descent, while using distributed gradient computation and centralized synchronization [2204.09726]. The reported behavior is monotonic decrease of \( H \), with reallocation when capability changes or a robot is removed [2204.09726].

A more explicit workload-adaptive formulation appears in "MDCPP: Multi-robot Dynamic Coverage Path Planning for Workload Adaptation" [2509.23705]. There, the central claim is that robot velocities are variable due to interactions with targets, hardware differences, or environmental factors, so static partitioning leads to imbalanced assignments and increased completion time. The algorithm dynamically estimates each robot's remaining workload by approximating the target distribution with Gaussian mixture models and assigns coverage regions using a capacity-constrained Voronoi diagram [2509.23705]. The workload assigned to robot \( r_i \) in region \( S_{i,t} \) is written as
\[
w_{i,t} = \Phi_i \cdot \alpha_i \sum_{j=1}^{n_{g,i}} \|q_{i,t} - g_j\|, \quad g_j \in S_{i,t},
\]
where \( \Phi_i \) is the sum of estimated coverage times, \( \alpha_i \) reflects robot capability, and \( q_{i,t} \) is the current robot position [2509.23705]. The optimization target is
\[
\min \max_{i=1,\dots,n_r} w_{i,t}.
\]

These two lines of work are closely related but not identical. The BCD-based method directly reallocates geometric area according to capability factors [2204.09726], whereas the workload-adaptive MDCPP formulation estimates target-dependent coverage effort via Gaussian mixtures and then balances predicted workload under a capacity-constrained Voronoi assignment [2509.23705]. This suggests two complementary notions of “dynamic” coverage balancing: one based on robot-side capability variation, the other on environment-side workload variation.

## 3. Online planning in unknown environments

A second major MDCPP theme concerns coverage in unknown workspaces where paths must be updated online. "Online On-Demand Multi-Robot Coverage Path Planning" [2303.00047] presents a centralized, horizon-based, synchronous, on-demand algorithm for large, complex, unknown workspaces. Its distinguishing mechanism is that only robots that have traversed previously computed paths entirely are replanned, while the remaining paths for other robots are reused [2303.00047]. The paper formally proves complete coverage of the unknown workspace and reports that the method scales to hundreds of robots in large complex workspaces, consistently beating a state-of-the-art online centralized algorithm in terms of the time needed to achieve complete coverage [2303.00047].

"Online Concurrent Multi-Robot Coverage Path Planning" [2403.10460] addresses a different inefficiency in receding-horizon systems: when path planning occurs for robots with no paths, robots with outstanding paths do not execute, and when robots execute, planning does not occur for robots yet to get new paths. The proposed remedy is a centralized algorithm that is not horizon-based. It plans paths at any time for a subset of robots with no paths while the rest execute their outstanding paths, thereby enabling concurrent planning and execution [2403.10460]. The algorithm uses a global discrete clock synchronized with robot clocks, reserves non-participants’ remaining path segments, proves complete coverage of an unknown workspace, and reports up to \(1.6\times\) speedup relative to a state-of-the-art horizon-based algorithm [2403.10460].

A hierarchical online formulation appears in CAP, "A Connectivity-Aware Hierarchical Coverage Path Planning Algorithm for Unknown Environments using Coverage Guidance Graph" [2503.00647]. CAP incrementally constructs a coverage guidance graph whose nodes represent disconnected uncovered subareas and the robot’s current position, and whose edges represent collision-free paths between subareas [2503.00647]. On this graph, a global Traveling Salesman Problem determines an efficient order for visiting open subareas, while local planning uses a greedy back-and-forth pattern for exploring subareas and local TSP solving for explored subareas [2503.00647]. The main paper focuses on single-robot implementation, but the architecture is described as lending itself to MDCPP through shared or distributed coverage guidance graphs, dynamic reallocation, and possible connectivity maintenance [2503.00647].

Multi-CAP generalizes this architecture explicitly to multiple robots [2509.14941]. It dynamically constructs and maintains an adjacency graph, keeps subareas as connected components of free space, and formulates subarea assignment as an Open Multi-Depot Vehicle Routing Problem to compute disjoint tours that minimize redundant travel [2509.14941]. Each robot then executes its assigned tour and adapts its local coverage strategy in real time within each subarea. Simulation and hardware experiments are reported to show improvements in coverage time, total path length, and path overlap ratio relative to state-of-the-art methods [2509.14941].

## 4. Centralized, decentralized, and distributed coordination architectures

MDCPP research exhibits a wide spectrum of coordination architectures. At one end are centralized planners and synchronizers. The BCD-based heterogeneous-capability controller uses a centralized machine to synchronize the robots, receive updated capabilities, compute and broadcast normalized capabilities and the optimization state, and transition operation between initialization and patrolling [2204.09726]. The on-demand and concurrent online planners also rely on centralized planning servers that fuse local views, assign goals, and distribute timestamped paths [2303.00047], [2403.10460].

At the other end are decentralized formulations. "Decentralized Coverage Path Planning with Reinforcement Learning and Dual Guidance" [2210.07514] introduces a decentralized reinforcement learning framework with dual guidance in which each robot independently plans coverage without a central controller. The framework augments local observations with artificial potential fields and heuristic guidance so that each agent can infer both environmental complexity and other agents’ intentions [2210.07514]. The task is cast on a grid graph \(G=(V,E)\), agents iteratively expand territories in round-robin fashion, and area division is optimized under
\[
\min \max \;\; ||T_i||
\quad
\text{s.t.}\;\; O \cup T = V,\; O \cap T = \emptyset,
\]
with a team reward penalizing uncovered nodes and overlap [2210.07514]. Reported comparisons show at most 10 percent lower overlap rates than a state-of-the-art decentralized method while maintaining similar efficiency in comparable environments [2210.07514].

A different decentralized architecture appears in "Decentralized Task and Path Planning for Multi-Robot Systems" [2011.10034]. The DTPP framework separates a task allocation module from a localized path planning module. Tasks are modeled as MDPs or MOMDPs, and expected pure reward maximization is solved on a factor graph with the max-sum algorithm. Potential conflicts are then handled by localized forward dynamic programming on local adjacency subgraphs [2011.10034]. Although the paper addresses general dynamic task allocation rather than only coverage, it is presented as suitable for MDCPP when “cover this zone” or similar tasks emerge dynamically [2011.10034].

A distributed, range-constrained implementation is central to workload-adaptive MDCPP [2509.23705]. Here, robots exchange job assignments and positions only with neighbors within communication range and perform local pairwise cell swaps. Repartitioning is triggered when a robot is about to finish its assigned cells, and the paper states that full connectivity yields globally optimal assignments equivalent to centralized capacity-constrained Voronoi partitioning, whereas limited range induces local rather than global workload balancing [2509.23705]. This communication-range dependence is one of the clearer examples in the surveyed work where the distributed implementation itself becomes an object of performance analysis [2509.23705].

## 5. Path generation, deconfliction, and coverage quality criteria

Once work is partitioned, MDCPP methods still differ substantially in how they generate actual coverage routes and how they handle overlap, turning, and conflicts. Several works adopt classical single-robot coverage primitives inside dynamically assigned subareas. The BCD-based heterogeneous-robot framework explicitly states that robots patrol their assigned region using single-robot coverage path planning, for example standard CPP algorithms, once allocations converge [2204.09726]. Multi-CAP likewise uses back-and-forth sweeping in exploring subareas and shortest open-loop coverage paths in fully explored uncovered subareas [2509.14941].

Spanning-tree-derived approaches remain prominent for exact grid coverage. The improved DARP method combines an A*-based task assignment with STC and an Up-First refinement so that obstacle-aware task allocation reduces inefficient assignments and robot-level complete coverage compensates for blind spots caused by coarse cells [2304.09741]. TMSTC* explicitly optimizes turn reduction by partitioning the map into minimum bricks, reducing the problem to finding the maximum independent set of a bipartite graph, then connecting bricks greedily into a tree so that the corresponding circumnavigating coverage path makes fewer turns [2212.02231].

Large-scale grid coverage with explicit path deconfliction is treated in "Large-Scale Multirobot Coverage Path Planning on Grids With Path Deconfliction" [2411.01707]. That work reformulates MCPP directly on a 4-neighbor grid \(G\), establishes NP-hardness results, introduces Extended-STC (ESTC) for partially obstructed \(2\times2\) blocks, and integrates ESTC into LS-MCPP, a local-search framework with grow, deduplicate, and exchange operators [2411.01707]. It also introduces a MAPF-based post-processing stage, applying Multi-Agent Path Finding techniques to MCPP for the first time to resolve inter-robot conflicts and accommodate turning costs [2411.01707]. The path cost is written as
\[
c(\pi_i) = \sum_{e \in E(\pi_i)} w_e + \sum_{j=2}^{|\pi_i|} c_t(\psi_{j-1}, \psi_j),
\]
so both traversal and turning are explicit optimization terms [2411.01707].

Overlap and redundancy are recurring quality criteria. CAP evaluates coverage time, path length, and path overlap ratio and reports lower overlap than five baselines [2503.00647]. DODGE uses overlap count in the terminal reward and reports overlap rates below those of AWSTC and GM-VPC on large maps [2210.07514]. Multi-CAP reports improvements in path overlap ratio, while LS-MCPP’s deduplication operators target the elimination of unnecessary overlaps directly [2509.14941], [2411.01707]. This suggests that “coverage completeness” in MDCPP is typically not considered sufficient on its own; redundancy control is a coequal design goal.

## 6. Sensing objectives, application variants, and practical constraints

MDCPP is not limited to planar occupancy coverage. In "Terrain-Aware Adaptation for Two-Dimensional UAV Path Planners" [2507.17519], a modular extension retrofits 2D multi-UAV planners for terrain-aware planning by adjusting altitude and camera orientations. The demonstrated system extends DARP to DARP-3D, beginning with baseline 2D paths at constant altitude, reconstructing a coarse 3D model, then adapting waypoint altitude and camera pitch and yaw to local terrain and occlusion [2507.17519]. The paper reports improved 3D reconstructions, particularly around vertical features, and notes that the modular algorithm is applicable to fixed-altitude planners more generally [2507.17519]. Although this work is centered on reconstruction rather than classical planar coverage, it shows that MDCPP can also refer to dynamic adaptation of sensing geometry rather than only spatial region assignment.

Resource-constrained worker-station systems provide another application-specific variant. "Learning to Coordinate for a Worker-Station Multi-robot System in Planar Coverage Tasks" [2208.02993] formulates multi-robot coverage with workers of limited energy and one mobile station for replenishment as a fully cooperative multi-agent reinforcement learning problem. The objective is to simultaneously solve coverage planning for workers and rendezvous planning for the station in the presence of random dynamic interferers [2208.02993]. Worker energy is updated according to whether the worker is within proximity of a station, and shared reward combines newly covered area, energy penalties, collision penalties, and time penalties [2208.02993]. The paper states that the learned policy reduces the influence of random dynamic interferers on planning and performs competitively in task finish time [2208.02993].

A different practical emphasis appears in "Coverage Path Planning with Budget Constraints for Multiple Unmanned Ground Vehicles" [2306.04083]. That work proposes a multi-modal, state-machine-based algorithm integrating variable-formation coverage path planning and collaborative active sensing. Formation switching is triggered by block size, LiDAR-detected obstacles, and loss or addition of agents; the planner trades off coverage percentage, path length, and time budget [2306.04083]. The use of V-, U-, and queue formations, together with safe-angle obstacle avoidance, shows a version of MDCPP in which formation control and mode switching are integral to coverage execution rather than post hoc control layers [2306.04083].

These variants indicate that the “dynamic” in MDCPP can denote different adaptive targets: geometry of sensing, replenishment logistics, or motion formation. A plausible implication is that the field is increasingly defined less by a single canonical formulation and more by a shared requirement of online reallocation under coupled motion-sensing-resource constraints.

## 7. Evaluation patterns, assumptions, and open tensions

Evaluation across the cited literature is heterogeneous but reveals a consistent set of dominant metrics: coverage time, mission completion time, path length, path overlap ratio, coverage percentage, turnaround time, and in some UAV settings precision, recall, F1-score, and cloud-to-cloud distance for reconstructed models [2503.00647], [2303.00047], [2306.04083], [2507.17519]. Hardware validation is common: ROS and Gazebo simulations with TurtleBots and quadcopters appear in online centralized planners [2303.00047], [2403.10460]; Spot is used in CAP experiments [2503.00647]; DJI hardware appears in DARP-3D [2507.17519]; Multi-CAP and worker-station MARL methods also report real-robot experiments [2509.14941], [2208.02993].

A persistent tension concerns centralization versus robustness. Centralized planners and synchronizers often provide stronger guarantees on completeness and cleaner workload balancing [2204.09726], [2303.00047], [2403.10460], whereas decentralized and distributed approaches emphasize resilience to malfunctions, communication limits, or dynamic failures [2210.07514], [2011.10034], [2509.23705]. Another tension concerns static optimality versus online scalability: horizon-based replanning of all robots can be costly, motivating on-demand or concurrent replanning strategies that may sacrifice some path optimality for reduced mission time [2303.00047], [2403.10460].

A further recurring issue is the gap between known-map and unknown-map assumptions. Improved DARP and TMSTC* are optimized for static known environments [2304.09741], [2212.02231], whereas CAP, Multi-CAP, and online centralized planners explicitly target unknown workspaces [2503.00647], [2509.14941], [2403.10460]. In some cases this gap is stated directly: the improved DARP paper notes that when environmental information is unknown, the algorithm cannot perform reasonable task allocation and path planning, and that these issues need further research [2304.09741]. Conversely, online methods that handle unknown environments do not always model rich robot heterogeneity or fine-grained workload prediction.

Taken together, the surveyed work presents MDCPP as an evolving synthesis of coverage control, online task allocation, graph partitioning, multi-agent learning, and path deconfliction. The strongest common thread is not a single algorithmic template but a recurring systems objective: complete coverage with dynamically rebalanced effort, low redundancy, and operational feasibility under changing robots, maps, and sensing demands [2509.23705], [2509.14941], [2411.01707].

Source: https://www.emergentmind.com/topics/multi-robot-dynamic-coverage-path-planning-mdcpp