Stochastic Graph Games-Inspired Reinforcement Learning for Self-Organizing Robot Design

This study investigates the reinforcement learning algorithm and structural design of modular self-organizing robots in robotics. It introduces a reinforcement learning-based goal decision-making and path planning algorithm, utilizing uniform grid partitioning to manage the state space efficiently, mitigating the curse of dimensionality. The algorithm, developed in Python, demonstrates its effectiveness through simulations and real-world decision-making via a QT host computer program. Additionally, the paper describes a stochastic game-based reinforcement learning algorithm with an enhanced reward function rooted in potential energy, which is integrated into a hierarchical decision-making framework. The structural design of the self-organizing robot is executed using SolidWorks, complemented by 3D printing. The design extends to the robot's printed circuit board (PCB) and visual positioning system. Empirical validation, involving ID recognition and visual positioning experiments conducted with OpenMV on five self-organizing robots, encompasses recognition, target searching, multi-module docking, release, and mobile navigation tasks. The successful outcomes validate the proposed method's effectiveness and the applicability of reinforcement learning in multi-body robot control, laying a foundational algorithm and hardware platform for future control research.