EVOLVING NEURAL NETWORK CONTROLLERS FOR MODULAR ORIGAMI ROBOTS USING EMBODIED ARTIFICIAL INTELLIGENCE

Modular Origami Robots (MORs) offer adaptable morphology through foldable units, enabling shape reconfiguration. Each structural variation demands new control strategies, making manual programming inefficient and impractical. The study outlines an automated framework to evolve neural network controllers using evolutionary algorithms in simulation. Controllers evolved within the PyBullet engine via Pyrosim, across two MOR morphologies: a Three-Cubed Robot and a quadruped. Implemented and compared three algorithms: Random Search (RS), Hill Climber (HC), and Parallel Hill Climber (PHC). Performance metrics included locomotion speed, energy efficiency, gait stability, and adaptability across terrains. The Parallel Hill Climber achieved the highest speed (0.3 m/s), lowest energy usage (15.0 AU/m), and fastest convergence, outperforming all other algorithms (RS and HC) across both robot morphologies (Three-Cubed Robot and Quadruped Robot). Quadrupeds evolved stable, natural-like trotting behaviors, while even the minimal three-cube robot achieved forward locomotion via learned coordination. The evolved neural controllers generalized to unseen environments without retraining, demonstrating transferability and robustness. No prior knowledge or manual tuning was required, proving that evolutionary optimization can autonomously generate effective, morphology-aware control policies. The findings suggest that an embodied AI approach produce adaptive behavior in MORs. The approach shows promise for developing robots that operate in unstructured settings (e.g., search-and-rescue or space exploration) without extensive manual reprogramming.