Distributed multi-robot system for gas source localization: belief merging, region allocation, and informative path planning
This thesis presents the design, implementation, and evaluation of a distributed Multi-Robot System (MRS) for Gas Source Localization (GSL) in indoor environments using Khepera IV robots eqquiped with Metal-Oxide (MOX) gas sensors. The work extends an existing SingleRobot framework by incorporating...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/443236 |
| Acceso en línea: | https://hdl.handle.net/2117/443236 |
| Access Level: | acceso abierto |
| Palabra clave: | Mobile robots Distributed artificial intelligence Gas detectors Robots mòbils Intel·ligència artificial distribuïda Detectors de gasos Àrees temàtiques de la UPC::Informàtica |
| Sumario: | This thesis presents the design, implementation, and evaluation of a distributed Multi-Robot System (MRS) for Gas Source Localization (GSL) in indoor environments using Khepera IV robots eqquiped with Metal-Oxide (MOX) gas sensors. The work extends an existing SingleRobot framework by incorporating belief sharing and probabilistic map merging across multiple agents. A new distributed Finite State Machine (FSM) is developed to coordinate inter-robot communication and ensure synchronization among robots. A distributed control architecture is adopted in which robots share information with each other in the form of their individual belief maps. Each robot independently merges the received information using Bayes’ theorem, resulting in a belief map that becomes common across the entire team. As a consequence, all robots are considered equally reliable, and equal weight is assigned to each belief map during the fusion process. Several region allocation strategies are studied to efficiently divide the search space and avoid redundant exploration. The process began with Voronoi partitioning as a baseline approach due to its simplicity and speed. However, its tendency to create unbalanced regions motivated the use of Lloyd’s algorithm-both in its distance-based and information-based variants- as a refinement step. To further improve both spatial fairness and information distribution, a Region Growing strategy was finally developed. Within their assigned regions, robots plan trajectories using a combination of adapted and proposed path planning strategies, such as Greedy, Belief Clustering-based Informative Path Planning, Informative Center, and multiple variants of Cumulative Belief Informative Path Planning (CB-IPP) with different exploration-exploitation trade-offs. To optimize the efficiency of the MRS-specifically, to reduce overall idle time- more frequent belief map updates were enabled through an information gain-based approach. KullbackLeibler divergence is used to quantify changes in the belief distribution, allowing the system to trigger replanning only when significant updates are detected. Comprehensive simulations and performance evaluations under both calibrated and uncalibrated sensor conditions demonstrate that collaborative planning significantly reduces search time and minimizes travel distances while ensuring high success in GSL. The analysis is not only results-oriented but also investigates the emergent collective behavior of the MRS throughout the task. The results confirm the viability and benefits of employing the MRS for GSL tasks in complex environments with obstacles, setting the groundwork for future real-world implementations. |
|---|