A Hybrid Reinforcement Learning Framework for Autonomous Robotic Navigation in Dynamic Environments

Abstract

Autonomous agents—systems that make independent decisions without human input—are foundational to modern robotics and enable intelligent responses in dynamic situations. In this work, a hybrid agent-based system that integrates software agents, or programs that represent users with multiple decision-making modules. The design integrates perception (gathering and interpreting sensory data), planning (scheduling a sequence of actions), and reinforcement learning, in which agents use feedback from their surroundings to improve their actions through try and error. Mathematical modelling and experimental evaluation reveal efficiency gains over rule-based systems that rely only on predefined instructions. Consequently, the framework ensures adaptability, scalability, and robustness in uncertain, unpredictable environments. Results show a 17% higher success rate and reduced execution time.