Design and Implementation of a Cloud-Based Robot Swarm Control System with Integrated Simulation Environment
Csiszár, István (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060319552
https://urn.fi/URN:NBN:fi:amk-2025060319552
Tiivistelmä
This thesis investigates the development and empirical evaluation of a cloud-based robot swarm control system that seamlessly integrates physical and simulated robots through a unified web interface. The research focuses on a system designed to control multiple small, ESP32-S3-based robots, enabling collective navigation behaviors through on-policy reinforcement learning while providing a mirrored simulation environment for rapid prototyping and testing.
The objective of this research is twofold: first, to examine the effectiveness of a unified control architecture that bridges the gap between simulated and physical robot swarms; and second, to evaluate the performance of on-policy reinforcement learning algorithms in enabling autonomous navigation and coordination behaviors across the swarm. This study employed a methodological approach involving the design of a distributed MQTT-based communication infrastructure, the development of browser-based simulation tools, and the implementation of collective learning algorithms suitable for resource-constrained embedded systems.
Results from experimental trials demonstrate the efficacy of the integrated approach, highlighting both the advantages of cloud-based control architectures and the challenges of implementing reinforcement learning on small-scale robotic platforms. The system shows promising capabilities in terms of scalability, with consistent performance across varying swarm sizes and operational scenarios. Furthermore, the research identifies key considerations when transferring learned behaviors from simulated to physical environments, particularly regarding sensor limitations and real-world dynamics.
In conclusion, this thesis establishes that a cloud-based control system with an integrated simulation environment provides significant advantages for the development and deployment of robot swarms, offering a foundation for future research in distributed robotics and multi-agent reinforcement learning applications.
The objective of this research is twofold: first, to examine the effectiveness of a unified control architecture that bridges the gap between simulated and physical robot swarms; and second, to evaluate the performance of on-policy reinforcement learning algorithms in enabling autonomous navigation and coordination behaviors across the swarm. This study employed a methodological approach involving the design of a distributed MQTT-based communication infrastructure, the development of browser-based simulation tools, and the implementation of collective learning algorithms suitable for resource-constrained embedded systems.
Results from experimental trials demonstrate the efficacy of the integrated approach, highlighting both the advantages of cloud-based control architectures and the challenges of implementing reinforcement learning on small-scale robotic platforms. The system shows promising capabilities in terms of scalability, with consistent performance across varying swarm sizes and operational scenarios. Furthermore, the research identifies key considerations when transferring learned behaviors from simulated to physical environments, particularly regarding sensor limitations and real-world dynamics.
In conclusion, this thesis establishes that a cloud-based control system with an integrated simulation environment provides significant advantages for the development and deployment of robot swarms, offering a foundation for future research in distributed robotics and multi-agent reinforcement learning applications.