Session 2B – Swarm Robotics
Date: Wednesday, 18th November 2026
Time: 14:00 – 15:30
Room: Aula Capella
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
14:00 - Reducing Stale-State Effects in Multi-Robot Planning through Runtime-Aware State Projection
Authors: Marco Moreira¹, Diogo Matos¹,², Pedro Costa¹,², E. J. Solteiro Pires¹,³, Rui Simão⁴ and P. B. Moura Oliveira¹,³
Affiliations: ¹ INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² FEUP – Faculty of Engineering of the University of Porto, Portugal; ³ UTAD – University of Trás-os-Montes and Alto Douro, Vila Real, Portugal; ⁴ Flowbotic Mobile Systems S.A., Mangualde, Portugal.
Abstract
In centralized multi-robot fleet coordination systems with a real-time supervisor, planning requests may be issued while robots are already executing previously assigned motions. Since planning requires non-negligible computation time, the fleet state used by the planner can become outdated before the computed plan is returned, leading to stale-state effects and redundant or poorly timed decisions.
This paper proposes a runtime-aware state projection mechanism that compensates for planning delays by predicting planner execution time online and projecting robot states to the estimated plan reception instant before planner invocation. Experimental evaluation demonstrates that a simple, yet accurate, runtime predictor is sufficient to compensate for stale-state effects, allowing robots to safely continue moving during replanning.
14:15 - Priority Ordering Heuristics for Sequential Prioritized Multi-Robot Path Planning
Authors: Afonso Mateus¹,², Marco Moreira¹, Diogo Matos¹,², Héber Sobreira¹,², Gonçalo Teixeira³ and Pedro Costa¹,²
Affiliations: ¹ INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² FEUP – Faculty of Engineering of the University of Porto, Portugal; ³ Flowbotic Mobile Systems, S.A., Viseu, Portugal.
Abstract
Sequential prioritized planning remains attractive for centralized multi-robot coordination because it is simple, scalable, and compatible with online execution constraints. However, its performance depends strongly on the order in which robots are planned. This paper evaluates interpretable priority ordering heuristics for the Time Enhanced A* algorithm using path length, roadmap topology, local density, and predicted conflicts. Experiments across five controlled scenarios show that no single ordering rule performs best across all evaluated scenarios. Density heuristics help resolve congestion around clustered starts or goals, topology-based heuristics account for narrow passages, and conflict-based heuristics address predicted path interactions. In each scenario, at least one heuristic produces a complete solution and reduces excess trajectory cost relative to the mean excess cost of successful random-ordering runs, with reductions up to 74%. These results highlight the role of instance structure in choosing effective priority orderings for sequential multi-robot path planning.
14:30 - Fleet Assembly Optimisation for Cooperative Ground Transporters in Constrained Environments
Authors: Gonçalo Teixeira and Alberto Vale
Affiliation: Instituto de Plasmas e Fusão Nuclear (IPFN), Instituto Superior Técnico (IST), Universidade de Lisboa, Lisbon, Portugal.
Abstract
Many warehouses move heavy loads with single-pallet transporters. Replacing them with smaller cooperative ground transporters (trolleys) reduces occupied space and maintenance costs, but requires assembling, for each request, a fleet from the trolleys available in the facility. This paper formulates this online fleet-assembly problem as a multi-objective optimisation problem with five objectives: distance to the goal, inter-path distance (shared corridor), distance to hazardous areas, fleet battery level and fleet trust level, subject to a battery feasibility constraint. Candidate paths are generated over a Delaunay triangulation of the environment, and a constraint-aware NSGA-II-based algorithm selects the trolleys and their paths. On two simulated maps, selecting fleets of four trolleys from twenty (about 1.9 × 10⁹ combinations), the method returns a set of feasible trade-off fleets in under 4.5 minutes per run, with consistent results over 30 independent runs. The results show that battery and trust trade-offs are driven by the choice of trolleys rather than paths, and that elitism is essential for consistent results. The framework suits constrained facilities such as ITER and DEMO, and motivates many-objective methods as future work.
14:45 - Maestro: Multi-Robot Fleet Coordination Framework Based on Action-Oriented Graphs and the TEA* Algorithm
Authors: Diogo Miguel Matos¹, Pedro Costa¹, Héber Sobreira¹ and Rui Simão²
Affiliations: ¹ CRIIS – Centre of Robotics in Industry and Intelligent Systems, INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² Flowbotic Mobile Systems, Coimbra, Portugal.
Abstract
The Maestro framework bridges the gap between optimal spatio-temporal planning and the kinematic uncertainties inherent to real-world Multi-Agent Path Finding (MAPF) execution. By integrating the Time-Enhanced A* (TEA*) algorithm into an action-oriented graph representation, the framework explicitly models both the sequential dependencies of actions and their temporal execution costs. To mitigate temporal de-synchronizations caused by mechanical delays or actuation variability, Maestro introduces an intelligent edge decomposition strategy that segments continuous actions, alongside the dynamic insertion of ghost decision vertices in a closed-loop supervision system. This architecture ensures robust fleet coordination without forcing computationally demanding route recalculations for the entire fleet during minor execution delays. Empirical evaluations conducted across industrial and hospital topologies demonstrate that the time-aware Maestro framework successfully sustains operational safety and prevents systemic deadlocks for large fleets, significantly outperforming reactive, non-temporal baseline coordinators
15:00 - Micro Aerial Vehicle Platform for Resilient Swarm Communications
Authors: Leandro Resende¹, Diogo Moutinho¹, Alexandre Vilhena Costa¹, Daniel F. Campos¹ and Andry Maykol Pinto²
Affiliations: ¹ Centre for Power and Energy Systems, INESC TEC and FEUP, Porto, Portugal; ² Department of Electrical and Computer Engineering, FEUP and INESC TEC, Porto, Portugal.
Abstract
Cooperative Micro Aerial Vehicles (MAVs) can improve inspection and monitoring missions by distributing sensing and decision-making tasks across multiple lightweight platforms. However, their practical deployment remains constrained by limited payload capacity, antenna size, onboard computation, and communication range.
This work presents a lightweight MAV platform designed to support Ground Control Station (GCS) connectivity and inter-MAV communication through a 5G mobile network. The system integrates a custom sub-400 g MAV with a single-board computer (SBC), enabling a Linux-based ROS 2 software stack, cellular connectivity, video streaming, and onboard processing for future decentralized swarm coordination. The architecture separates low-level flight communication from high-level coordination. MAVLink is used for low-level communication, while ROS 2 is used for high-level information exchanged through a Zenoh Bridge over a VPN/5G network. This design supports both centralized swarm logic, where coordination is computed at the GCS, and decentralized logic, where each MAV processes shared information onboard.
The system is evaluated as a proof of concept through static two-node bench experiments over a cellular/VPN connection. Results show one-way MAVLink telemetry latency of approximately 42 ms, end-to-end 1080p video latency below 200 ms, and MAV-to-MAV communication latency close to 50 ms. These results indicate that the proposed architecture is a practical foundation for compact MAV platforms requiring remote operation, video transmission, and cooperative swarm behaviors.
15:15 - Learn alone or learn together? A minimal model of social learning in robot swarms
Authors: Gabriel Madroñero Pachajoa¹, David Torres Torres¹, Andrés Tutivén Gálvez¹, José Santos Terán¹, Carlos Samaniego Torres², Pablo Sebastián Padrón Martínez³,⁴, Bryan Puruncajas¹, Christian Tutivén Gálvez¹, Carlos Saldarriaga¹,⁵ and David Garzón Ramos⁶
Affiliations: ¹ Escuela Superior Politécnica del Litoral (ESPOL), Guayaquil, Ecuador; ² Universidad de Guayaquil, Guayaquil, Ecuador; ³ Universidad del Azuay, Cuenca, Ecuador; ⁴ Instituto Nacional de Biodiversidad (INABIO), Quito, Ecuador; ⁵ Universitat Politècnica de Catalunya (UPC), Barcelona, Spain; ⁶ University College Dublin, Dublin, Ireland.
Abstract
Robot swarms are groups of relatively-simple robots that can perform tasks in parallel, coordinating through local interactions. This ability to act in parallel makes robot swarms well suited to social learning: the capacity of a group to accumulate and transmit knowledge across individuals. Yet, this learning approach has received little attention in swarm robotics, where robots are typically programmed in advance and do not learn and adapt while operating. We study social learning with a minimal model in which learning takes two forms: individual learning, by which a robot completes a task faster the more often it performs it; and social learning, by which it improves further by observing nearby peers that perform the same task. We evaluate the model with a swarm of eight robots in simulation and with physical robots. For these experiments, we adapted and extended an experimental protocol from the literature, and we developed the QUPA, a robot for swarm robotics research. Results show that enabling social learning improved the performance in comparison to an individual learning alternative. We argue that exploring minimal models is a promising way to identify when social learning can benefit a swarm deployment, before turning to sophisticated learning methods or more capable robots.