Work published in Cognitive Systems Research - Elsevier. This paper evaluates two RL-based drive optimization strategies within the framework of cognitive architectures for autonomous robots.
Dual or unified: optimizing drive-based reinforcement learning for cognitive autonomous robots / 2026 / Cognitive Systems Research (CSR)
Reinforcement learning (RL) methods inspired by cognitive architectures are crucial for empowering autonomous agents to tackle complex, dynamic tasks. This study evaluates two RL-based drive optimization strategies – 1-LDO and 2-LDO – within the framework of cognitive architectures for autonomous robots. 1-LDO integrates both motivational drives into a single learning model, whereas 2-LDO separates them into distinct models, allowing for modular learning. Grounded in Hull’s Drive Theory, we explore early versus late selection mechanisms to optimize drive reduction through RL, particularly in agents driven by curiosity and survival imperatives. Through reward and stress analyses, we demonstrate that Deep Q-Network (DQN) agents outperform traditional Q-Learning approaches in fine-grained environments, with the 2-LDO configuration showing marked advantages due to its modular design. In contrast, in coarser environments, 2-LDO combined with Q-Learning achieves superior efficiency, offering faster drive regulation at reduced computational cost. These results suggest that early selection mechanisms, aligned with Hull’s theoretical principles, may provide the most effective strategy for optimizing drive-based behaviors in autonomous agents.
Leonardo L. Rossi, Letícia Berto, Paula P. Costa, Ricardo Gudwin, Esther Colombini, Alexandre Simões. “Dual or unified: optimizing drive-based reinforcement learning for cognitive autonomous robots”, Cognitive Systems Research, Volume 95, 101422, ISSN 1389-0417, https://doi.org/10.1016/j.cogsys.2025.101422. 2026.