Multi-Robot Exploration with Communication Constraints

• The paper introduces MRE-CCIC frameworks that combine opportunistic and scheduled rendezvous to ensure efficient map fusion amid intermittent connectivity.
• It employs an entropy-field-based method and MILP scheduling to balance exploration efficiency, bandwidth limits, and robust information sharing.
• Experimental validations show up to 21% faster exploration, reduced overlap, and improved rendezvous efficiency, demonstrating practical impact.

Multi-Robot Exploration with Communication Constraints and Intermittent Connectivity (MRE-CCIC) focuses on the efficient coordination of multiple robots tasked with exploring unknown environments under realistic communication limitations. Unlike traditional approaches assuming persistent and unlimited communication, MRE-CCIC frameworks explicitly model both bandwidth and range constraints, tolerate disconnections, and devise mechanisms for robots to periodically reestablish communication (rendezvous) to synchronize global information. Methods in this domain jointly optimize exploration efficiency, robustness, scalability, and the minimization of redundant coverage.

1. Communication-Aware Network Models

MRE-CCIC systems formalize robot communication as a dynamic, time-varying connectivity graph. A two-layer model emerges as a standard: persistent, low-rate (e.g., pose-only) channels are always available, while high-bandwidth data exchange (including full map-merging) requires robots to be physically close, within a defined communication radius $r_\text{comm}$. Formally,

$$G(t) = (V, E(t)), \quad E(t) = \{(i, j) \mid d_\text{cur}(i, j) < r_\text{comm}\}$$

where $V$ is the set of robot identities and $d_\text{cur}(i, j)$ is Euclidean distance. Upon establishing a high-speed link, immediate map-merging is performed using a commutative, associative, idempotent operator $\boxplus$, ensuring eventual consistency under SLAM (Pongsirijinda et al., 29 May 2025).

This model supports both opportunistic and scheduled rendezvous schemes, balancing the cost of physical travel for reconnection versus the benefit of information fusion. Some architectures supplement local-to-local peer relaying with droppable radio infrastructure for multi-hop communication extension, but most MRE-CCIC systems fundamentally rely on in-range, explicit synchronization events (Saboia et al., 2022).

2. Exploration Algorithms Under Communication Constraints

MRE-CCIC requires exploration strategies that are highly robust against partitioned information. A key innovation is the entropy-field-based method, in which robots rank viewpoints according to a combined metric of expected exploration gain (frontier entropy) and implicit rendezvous desirability (robot entropy). Formally, the total entropy at candidate viewpoint $p$ near frontier centroid $q$ is

$H_\text{total}(i; p, q) = H_f(p, q) + H_r(i, p)$

where $H_f$ captures information-theoretic uncertainty over unexplored frontiers, and $H_r$ encodes a soft potential for rendezvous by increasing when nearby robots could merge maps (Pongsirijinda et al., 29 May 2025). Each robot greedily minimizes $H_\text{total}$ over all candidate viewpoints, triggering intrinsic regrouping as exploration proceeds and the frontier set dwindles.

Alternative paradigms include:

• Deep reinforcement learning policies, using hierarchical graph state encodings and curriculum learning to internalize long-term trade-offs between isolated exploration and reconnection (Tan et al., 7 Sep 2024).
• Job-shop or MILP-based scheduling for rendezvous, explicitly optimizing exploration “jobs” assigned to robots and synchronizing meeting events to minimize waiting time and maximize information flow (Silva et al., 15 Nov 2025, Silva et al., 2023).
• Lightweight, fully decentralized approaches relying on wireless-range localization pings to avoid explicit map sharing, instead penalizing utility for visiting frontiers likely already observed by others (Jadhav et al., 27 Dec 2024).

3. Rendezvous Planning and Execution

Two dominant modes of reconnection arise:

• Implicit (emergent) rendezvous: As in entropy-field and reinforcement-learning approaches, robots naturally converge when the information utility of reconnection exceeds that of further solo exploration. The rendezvous mechanism is encoded as part of the utility potential (Pongsirijinda et al., 29 May 2025, Tan et al., 7 Sep 2024).
• Explicit scheduled rendezvous: Here, rendezvous events are assigned via global plans, derived from MILP, job-shop scheduling, or mission-based protocols. Each robot tracks the global "plan" and switches from exploration to travel-to-rendezvous when a time threshold is reached, minimizing idle time at meeting points (Silva et al., 15 Nov 2025, Gao et al., 2021).

Dynamic adaptation is critical: in MILP-based schemes, the actual rendezvous location is chosen at runtime based on current local maps, using distributed leader election or consensus rules to ensure all assigned robots converge on the same next exchange location (Silva et al., 15 Nov 2025). Mission-based protocols such as Meeting-Merging-Mission alternate between centralized decisions at rendezvous and decentralized exploration in between, exploiting both global optimality and robustness to disconnection (Gao et al., 2021).

4. Information Sharing, Map Fusion, and Robustness

Map fusion is performed only during high-rate communication windows. In two-layer systems, robots continuously exchange low-rate pose or relative localization data, enabling anticipation of future encounters and efficient scheduling. Upon reconnection, a commutative map-merging operator ensures global map consistency.

Robust decentralized (no-global-map) methods, such as WiSER-X, employ only lightweight pings containing range and angle-of-arrival data to estimate neighbor positions using EKF. Each robot penalizes frontier utilities proportionally to the probability neighbors have already explored candidates, significantly reducing coverage overlap and exploration redundancy, empirically achieving 58% lower overlap compared to fully independent strategies (Jadhav et al., 27 Dec 2024).

Asynchronous self-termination is supported: individual robots locally decide when further exploration would only yield areas already likely covered by others, using observation counts and statistical thresholds, thus achieving near-complete coverage with minimal data exchange.

5. Scheduling, Task Synthesis, and Formal Guarantees

Several approaches use formal methods to encode MRE-CCIC objectives:

• Temporal logic synthesis: The exploration objective and intermittent meeting requirements are written as LTL specifications over state–task automata. Distributed synthesis follows a two-stage process: offline scheduling of conflict-free communication events and online joint path planning for each robot, guaranteeing deadlock-free, correct-by-construction satisfaction of both exploration and information-sharing constraints (Kantaros et al., 2017).
• MILP/ILP formulations: Decision variables encode participation and timing for rendezvous "jobs." The MILP minimizes exploration time error and duration variance, with constraints ensuring synchronization, allocation, and timing bounds. Execution tracks rendezvous windows robustly using real-time estimators for travel time, allowing deviation due to obstacles or unexpected delays (Silva et al., 15 Nov 2025, Klaesson et al., 2019).

Tabular summary of guarantee types:

Framework Type	Guarantee	Scalability
Entropy/Potential-driven (MEF-Explore)	Empirical near-optimality	2–15 robots
MILP/JSSP-based (explicit scheduler)	Schedule interpretability	2–10 robots
Temporal logic/distributed automata	Task satisfaction, deadlock-free	up to 100+ robots
RL-based (IR²)	Empirical efficiency, adaptivity	2–6 robots
Decentralized (WiSER-X)	Coverage completeness under error	2–5 robots

6. Experimental Validation and Performance Metrics

Extensive simulation and hardware studies validate MRE-CCIC frameworks:

• Exploration speed and consistency: MEF-Explore achieves 10–40% faster exploration than strong baselines, with 5–15% lower variance and up to 16% higher success rate. Real-world TurtleBot2 deployments yield 21.32% faster exploration and 16.67% more reliable completion (Pongsirijinda et al., 29 May 2025).
• Rendezvous efficiency: MILP-based scheduling (plus RTUS execution) matches planned rendezvous to within a few seconds, with average meeting wait 10× lower than heuristic baselines. Coverage remains optimal and is unaffected by rendezvous tracking (Silva et al., 15 Nov 2025).
• Bandwidth usage: Star-convex polytope and Super Frontier Information compression in Meeting-Merging-Mission reduces communication cost 30–60% versus GMM-based map sharing, supporting bandwidths orders-of-magnitude below SLAM-map streaming (Gao et al., 2021).
• Overlap and coverage: Decentralized, ping-only WiSER-X reduces overlap by 58% over zero-sharing approaches while maintaining ≈93% mean coverage at self-termination—robust even under large range-finding and angle errors (Jadhav et al., 27 Dec 2024).
• Scalability: Temporal-logic and network-flow–based methods scale to tens or even hundreds of robots in large environments by hierarchical clustering and team decomposition; the flow-ILP guarantees intermittent connectivity and information consistency efficiently (Kantaros et al., 2017, Klaesson et al., 2019).

7. Extensions, Limitations, and Open Directions

MRE-CCIC approaches have demonstrated practical robustness, scalability, and efficiency, but several challenges persist:

• Congestion and collision: Scaling beyond 1 robot per 50 m² brings new challenges, especially physical collision risk, which is seldom fully addressed by high-level frameworks (Pongsirijinda et al., 29 May 2025).
• Parameter tuning: Many frameworks rely on scenario-specific coefficients (entropy potential fields, thresholds, etc.); adaptive or learning-based parameterization is an open avenue (Pongsirijinda et al., 29 May 2025).
• Multi-hop and store-and-forward: Most implemented systems only exploit single-hop reconnection; integrating buffer-based store-carry-forward or explicit relay management could increase resilience in extreme environments (Saboia et al., 2022).
• Integration with operator-in-the-loop control: Recent frameworks such as MoRoCo enable adaptive mode switching among parallel exploration, migration, and chain relays to respond to human requests and dynamically updated task priorities under communication restrictions (Tian et al., 11 Aug 2025).
• Theoretical analysis under non-ideal SLAM/localization: Most methods assume reliable pose estimation and localization; future work is needed to quantify the impact of unmodeled NLOS errors and SLAM drift on rendezvous scheduling and map fusion.

MRE-CCIC continues to be actively studied, with research combining information-theoretic task assignment, formal methods for connectivity, bandwidth-aware mapping, secure protocol design for intermittent regimes, and lightweight real-world deployment optimization. The current consensus is that scheduled, information-driven rendezvous, robust decentralized estimation, and opportunistic peer-to-peer exchange are collectively essential for robust multi-robot exploration in realistic, communication-limited regimes (Pongsirijinda et al., 29 May 2025, Tan et al., 7 Sep 2024, Silva et al., 15 Nov 2025, Jadhav et al., 27 Dec 2024, Kantaros et al., 2017).