Model-Free Dynamic Control of Five-bar Mechanism via Subchain Assembly and Reinforcement Learning

Parallel robots are versatile but challenging to control due to their complex dynamics and the need for precise modeling. This research presents a novel approach to modular control of parallel robots, utilizing reinforcement learning to dynamically assemble and control complex arbitrary systems. By treating each sub chain of a parallel robot as an individual agent, the proposed method offers a flexible and adaptable framework for various control tasks. The study compares the learning performance of different RL algorithms and proposed decision modeling on the tracking control of a five-bar robot with a 55 cm link length, aiming to achieve an error of less than 5 cm in the PyBullet simulator. In single-agent reinforcement learning, three algorithms (DDPG, PPO, and SAC) were employed, with PPO and SAC effectively solving the problem. In multi-agent reinforcement learning, two algorithms (FACMAC and MAPPO) were utilized. While MAPPO successfully learned the problem, FACMAC exhibited difficulties in grasping the problem. The results obtained from comparing the proposed RL algorithms indicated that single-agent RL learns the control problem better and faster than multi-agent reinforcement learning. In short, the paper focuses on formulating an RL solution which can effectively control parallel robots and can potentially be used for pre-training subchains.