Swarm Control with RRT-APF Planning and FNN Task Allocation Tested on Mobile Differential Platform

Efficient coordination of autonomous robotic swarms represents a fundamental challenge in modern automation and distributed systems engineering. Prior research has shown that maintaining global coherence while individual agents navigate complex, obstacle-laden environments requires high-level computational oversight. Conventional methodologies often struggle to balance the computational demands of real-time pathfinding with the logistical complexity of dynamic task distribution across multiple mobile units. Many existing frameworks fail to account for the intricate spatial constraints encountered when multiple differential-drive robots operate within the confined workspaces of industrial facilities. The lack of integrated systems that simultaneously optimize both movement trajectories and agent-to-task assignments hinders the deployment of flexible manufacturing solutions. This absence of evidence motivated the design of a centralized control architecture capable of harmonizing these disparate operational requirements into a single cohesive framework.

This investigation develops a unified centralized control architecture that integrates sophisticated path planning with real-time task allocation for robotic swarms. The primary objective involves utilizing a Feedforward Neural Network (FNN) to generate suitability scores that determine the most appropriate agent for any given assignment. The researchers aimed to merge the exploratory capabilities of the Rapidly Exploring Random Tree (RRT) algorithm with the local navigation precision of Artificial Potential Field (APF) methods. Addressing the persistent issue of simultaneous agent collisions at intersections through the implementation of a Multi-Agent Pathfinding (MAPF) solution was a central goal. The study sought to validate these theoretical algorithms using a physical experimental platform featuring differential-drive robots tracked by high-precision ArUco markers. This research specifically targets the operational needs of swarm reconfigurable manufacturing systems by testing the system on tasks that mirror industrial production workflows.

The experimental framework employed a Feedforward Neural Network (FNN) architecture to process agent data and output suitability scores for dynamic task assignment. For navigation, the team implemented a hybrid planning algorithm that synthesized the global search capabilities of Rapidly Exploring Random Tree (RRT) with the reactive steering of Artificial Potential Field (APF). The researchers integrated a Multi-Agent Pathfinding (MAPF) protocol to manage the spatial interactions of multiple agents and resolve potential deadlocks at narrow intersections. Physical validation was conducted using a custom-built mobile differential platform where each robotic unit was monitored via an ArUco-based optical tracking system. The control software was developed to manage the swarm centrally, ensuring that task allocation and path execution were synchronized across all active robotic agents. Performance evaluation involved subjecting the swarm to a series of simulated manufacturing tasks designed to test the robustness of the integrated control logic.

The centralized control system demonstrated the ability to manage a swarm of robotic agents effectively by successfully interleaving task allocation and path execution. The Feedforward Neural Network (FNN) provided reliable suitability scores in real time, which facilitated the optimal distribution of manufacturing tasks among the available agents. Utilizing the hybrid Rapidly Exploring Random Tree (RRT) and Artificial Potential Field (APF) planner resulted in efficient trajectories that avoided both static and dynamic obstacles. The Multi-Agent Pathfinding (MAPF) component successfully mitigated the risk of simultaneous collisions, particularly when multiple robots converged on shared intersections during high-traffic scenarios. Data gathered from the differential-drive hardware experiments confirmed that the proposed algorithms translate effectively from simulation to physical robotic platforms. Quantitative assessments showed that the integrated approach maintained high performance levels during tasks that closely resemble the operational demands of reconfigurable manufacturing systems.

The integration of neural network-based task allocation with hybrid path planning provides a robust foundation for the management of centralized robotic swarms. These findings suggest that combining Rapidly Exploring Random Tree (RRT) and Artificial Potential Field (APF) algorithms effectively overcomes the limitations of individual navigation methods. The successful resolution of agent collisions through Multi-Agent Pathfinding (MAPF) protocols highlights the importance of specialized intersection management in multi-robot environments. The researchers conclude that this control architecture is particularly well-suited for deployment in swarm reconfigurable manufacturing systems where flexibility and coordination are paramount. Future development of the differential-drive platform and ArUco tracking system may further enhance the precision and scalability of this centralized swarm solution. This study establishes that a unified approach to planning and allocation is essential for the next generation of autonomous industrial robotic swarms.