Analysis of the CPG-DRL framework for Quadruped Robot Locomotion

Abstract

Autonomous robots are expected to be crucial in
mitigating the lack of precise border surveillance in complex
terrains, especially within forests, mountains, and industrial
settings, focused on predicting the entry of terrorists and illegal
refugees. These robots must demonstrate navigational proficiency
in both controlled areas, such as border control gates and
borders with plane surfaces, and unplanned country borders in
daily surveillance tasks, including mountains, forests, and mud.
In this regard, quadruped robots are attracting considerable
interest due to their ability to carry substantial surveillance
equipment while navigating inclines and obstacles. Given the
environment-dependent characteristics of optimal gait patterns,
it becomes a necessity for mechanisms capable of independently
and effectively optimizing gait patterns for adaptation to different
scenarios. To address these challenges, we explored a model for
quadruped locomotion that combines central pattern generators
(CPGs) with deep reinforcement learning (DRL). Using both
external and internal sensing, the agent learns to coordinate
rhythmic movement among multiple oscillators to follow speed
instructions, while also adjusting these instructions to avoid
bumping into things around it. This research helps to use
DRL to study important questions in neurology, such as how
certain pathways, connections between oscillators, and sensory
information affect walking patterns. Moreover, this paper will
present fundamental knowledge, research trends, and challenges
regarding the implementation of the CPG-DRL framework in
robotic perception.