A Decentralized Reinforcement Learning Approach for Modular Octoped Locomotion Control

With the increasing deployment of quadruped robots in complex environments, reinforcement learning (RL) has shown strong potential in tasks such as gait generation and dynamic balance. The modular octoped robot is composed of two quadruped subsystems connected front and rear. It features detachable and reconfigurable capabilities, enabling coordinated motion control. However, centralized reinforcement learning frameworks struggle with modular octoped robots due to limited coordination between subsystems, highdimensional action spaces, and poor scalability. To address these issues, this paper proposes a decentralized reinforcement learning control method that decouples the system into two quadruped subsystems, each with an independent PPO policy for local decision-making. By enabling asynchronous control and lightweight communication, the system achieves efficient and coordinated locomotion by enabling local decision-making and asynchronous coordination. Experiments on the IsaacGym platform demonstrate that the decentralized approach improves motion performance, and adaptability, offering a scalable and flexible solution for complex multi-robot systems.