Multi-Robot Connectivity-Aware Planner

• Multi-CAP is a framework for multi-robot planning that explicitly incorporates communication constraints to maintain continuous connectivity.
• It employs centralized and distributed methods—such as mixed-integer MPC and consensus ADMM—to optimize trajectory planning under stringent connectivity requirements.
• Empirical validations demonstrate improved path efficiency, reduced tracking errors, and robust connectivity through dynamic reconfiguration and online learning.

A Multi-robot Connectivity-Aware Planner (Multi-CAP) is a trajectory planning and decision framework designed for teams of mobile robots that must maintain explicit inter-robot communication connectivity while performing individual or distributed tasks. The key characteristic of a Multi-CAP is its explicit integration of connectivity constraints—arising from SNR, RSSI, or algebraic properties of the robots’ communication graph—directly into task allocation, motion planning, and control loops. The Multi-CAP paradigm has been developed to address heterogeneous robot deployment, coverage, goal assignment, and dynamic reconfiguration, with distributed or centralized architectures and a spectrum of connectivity metrics.

1. Formal Problem Definition and Core Constraints

The defining feature of Multi-CAP is the explicit representation and maintenance of a robot team’s communication network as a formal mathematical object. This object is typically a time-varying, undirected graph $G=(V,E,W)$, where $V$ indexes robots, $E$ encodes feasible communication links (usually based on distance, visibility, SNR, or RSSI), and $W$ assigns edge weights reflecting instantaneous link quality, channel uncertainty, or control-theoretic margins.

Common formal requirements include:

• Connectivity Constraint: The communication graph $G$ must remain connected at all times, typically enforced by ensuring the second smallest eigenvalue of the graph Laplacian, $\lambda_2(L(G))$ (the Fiedler value), satisfies $\lambda_2 \geq \lambda_{\text{min}}$ for a system-dependent $\lambda_{\text{min}}>0$ (Mikkelsen et al., 2024, Shetty et al., 2020, Yang et al., 2024).
• Pairwise Barrier Certificates: Many planners enforce safe sets $\{x\ |\ h_{ij}(x) \geq 0\}$ for each robot pair $(i,j)$, where $V$0 is a control-barrier function derived from inter-robot distances, link strength, or learned communication models (Lin et al., 2020, Yang et al., 2024).
• Uncertainty & Robustness: Advanced planners account for Gaussian position/sensing uncertainty, which propagate through the connectivity graph (via edge weights of the form $V$1) and into the constraint on $V$2 (Shetty et al., 2020).
• Task/Assignment Variables: Decision variables may include allocation matrices $V$3 for robot-task assignment, or VRP/MIP encodings for combinatorial allocation under connectivity and collision constraints (Lin et al., 2020, Caregnato-Neto et al., 2022, Shen et al., 18 Sep 2025).

2. Optimization Frameworks and Solution Architectures

Multi-CAPs implement connectivity constraints within various optimization and control frameworks, which can be categorized according to centralization, decision granularity, and real-time capabilities:

• Centralized Mixed-Integer MPC: A global controller solves a robust MPC with embedded connectivity-encoding (either through binary edge variables $V$4 or algebraic connectivity degree-constraints), achieving real-time, collision-free, robustly-connected task allocation and motion (Caregnato-Neto et al., 2022).
• Distributed Consensus-ADMM: Connectivity constraints, especially $V$5 constraints, are separated and linearized so that robots can trade local “communication budgets” (derived from sensitivity of $V$6 to robot motions) and solve local quadratic programs exchanged via ADMM loops (Mikkelsen et al., 2024, Shetty et al., 2020). This is particularly effective for scalable, SPOF-free deployment.
• Bi-level Barrier Function QPs with Online Learning: The connectivity maintenance problem is solved at each step as a QP that blends nominal control with collision, connectivity, and communication-barrier constraints, where the latter are learned online via Gaussian Process regression over RSSI or SNR measurements (Yang et al., 2024). A spanning-tree subgraph is often selected to reduce constraint redundancy.
• Greedy and Heuristic Methods for Heterogeneous Teams: In scenarios requiring rapid, suboptimal allocation (e.g., when demands and robot capabilities are only revealed upon area exploration), adaptive greedy assignment and single-integrator control are coupled with connectivity-preserving QPs (Lin et al., 2020).
• Manipulation-Theoretic Planning: Some frameworks map the connectivity enforcement to open-chain serial manipulators in configuration space, enabling direct enforcement of connectivity via kinematic constraints corresponding to maximum allowed inter-robot distances (Santos et al., 2024).

3. Connectivity-Aware Task Allocation and Coverage

The assignment and coverage strategies in Multi-CAP are shaped by the need to maintain connectivity during the entire mission. Key techniques are:

• Coverage Path Planning via VRP: The environment is represented as a dynamic adjacency graph of connected subareas; a Vehicle Routing Problem (VRP) solver assigns tours for robots to fully cover these areas while the connectivity-aware adjacency ensures each assigned zone is a single connected component, leading to reduced path length and overlap (Shen et al., 18 Sep 2025).
• Relay-Enhanced Assignment: Task allocation explicitly schedules robots both to primary goals and to act as “relays” positioned at critical points to maintain the communication backbone, with assignment often solved via cost-matrix optimization or Hungarian algorithms (Marchukov et al., 24 Mar 2025).
• Dynamic Budget Trading: In distributed frameworks, robots may trade prospective “communication budget” (based on $V$7 sensitivity), thus enabling agents performing critical tasks to pull connectivity support from others (Mikkelsen et al., 2024).
• Association Path Planning: For environments with fixed APs and wireless coverage, path planning is formulated as a joint robot-path and AP-association assignment problem, where cost includes both path distance and number of network handovers, and is solved efficiently via LP column generation and cooperative integer pruning (Tatino et al., 2020).

4. Real-Time and Online Operation

Real-world applicability of Multi-CAP algorithms is achieved through several practical mechanisms:

• Hierarchical, Multirate Architectures: A high-level planner computes assignment and reference trajectories at low rates (1 Hz), while low-level onboard controllers execute these commands at much higher frequencies (up to 60 Hz), interpolating planned trajectories for robust real-time control (Caregnato-Neto et al., 2022).
• Incremental Graph and Map Maintenance: The adjacency graph—encoding subareas and their connectivity—is updated online as new sensor data arrives, with flood-fill and A* algorithms used to update connectivity in the presence of dynamic obstacles (Shen et al., 18 Sep 2025).
• Decentralization and Minimal Communication: Distributed implementations require only neighborhood communication (or local estimation of Fiedler eigenvectors), limiting communication overhead while achieving consensus via asynchronous or synchronous ADMM iterations (Mikkelsen et al., 2024, Shetty et al., 2020).
• Online Learning of Network Models: Gaussian Processes ingest streaming RSSI/SNR data, incrementally refining connectivity-barrier certificates without needing pre-characterized channel models (Yang et al., 2024).

5. Empirical Validation and Performance

Multi-CAP performance is evaluated against carefully constructed benchmarks in simulation and hardware:

Framework / Application	Typical Team Size	Solve Frequency	Reported Performance Gains
VRP-based Coverage (Shen et al., 18 Sep 2025)	3–10	Real-time	–15–29% path length, –12–29% time vs. SoTA
Distributed ADMM (Mikkelsen et al., 2024, Shetty et al., 2020)	10–20	10–100 Hz	15–25% lower cost with trading; $V$8 always above minimum
Real-time MPC-MIP (Caregnato-Neto et al., 2022)	5–8	1 Hz Planning, 60 Hz tracking	$V$93cm RMS tracking error, 100% connectivity
Dynamic Heterogeneous Deployment (Lin et al., 2020)	20–80	10s–100s Hz	$E$010x faster assignment vs. MINLP/GA, near-optimal fulfillment

Further, ablation studies confirm robust performance degradation when connectivity-awareness or global assignment is disabled (Shen et al., 18 Sep 2025). In mmWave AP–robot scenarios, handover-optimized association-path planning achieves $E$150% reduction in handovers at near-optimal path lengths (Tatino et al., 2020).

6. Theoretical Properties and Extensions

• Completeness: Connectivity-aware frameworks that update coverage graphs over bounded domains guarantee eventual coverage of all free cells (Shen et al., 18 Sep 2025).
• Robustness: Under mild assumptions, enforcing algebraic or barrier certificate constraints ensures forward invariance of the connectivity set; distributed planners converge under convexity and consensus (Shetty et al., 2020, Yang et al., 2024, Mikkelsen et al., 2024).
• Scalability limits: Purely centralized planners (e.g., manipulator-based or MI-MPC approaches) become computationally intractable as $E$2; distributed and decentralized variants enable scalability at the cost of (often modest) optimality loss.
• Extensions: Ongoing work proposes the use of control/barrier function blending for nonholonomic, hybrid, and mixed-modal teams, heterogeneous chain architectures (including 3D aerial/ground hybrids), and further robustification via online learning and adaptation (Yang et al., 2024, Santos et al., 2024).

7. Open Challenges and Future Directions

• Computational Tractability: For large teams, further advances in distributed optimization, efficient eigenvalue sensitivity estimation, and compressed map representation are needed.
• Learning-Based Models: Data-driven barrier certificates enable adaptation to realistic channels, but scaling sparse GP inference and guaranteeing safety with bounded measurement error remain open.
• Integration with Higher-level Mission Planning: Seamless embedding of Multi-CAP modules within human-in-the-loop task management, resource scheduling, and resilience/recovery layers is under active study.
• Guarantees under Non-Idealities: Incorporating lossy communication, controller delay, and adversarial interference (network slicing, DoS) into the Multi-CAP constraint stack requires further theoretical and practical investigation.

In total, the Multi-Robot Connectivity-Aware Planner paradigm synthesizes graph theory, distributed optimization, learning theory, and real-time control, providing a rigorous foundation for robust, scalable, effective robot team deployment in complex dynamic environments (Lin et al., 2020, Mikkelsen et al., 2024, Caregnato-Neto et al., 2022, Yang et al., 2024, Shen et al., 18 Sep 2025, Shetty et al., 2020, Marchukov et al., 24 Mar 2025, Tatino et al., 2020, Santos et al., 2024).