برنامه‌ريزي حركت و كنترل بازوي ربات با استفاده از يادگيري تقويتي عميق به‌منظور عدم برخورد با موانع

Motion planning an‎d control of a robot manipulator using deep reinforcement learning to collision avoid obstacles

مهندسي برق

This research is designed an‎d conducted with the aim of developing an intelligent an‎d safe controller for path planning an‎d dynamic obstacle avoidance in robotic manipulators. In the first stage, a comprehensive research path is pursued, including the selec‎tion an‎d comparative eva‎luation of reinforcement learning algorithms. In this eva‎luation, the DDPG, TD3, an‎d SAC algorithms are compared, an‎d the results demonstrate that SAC outperforms the others in complex tasks such as simultaneous control of position an‎d orientation as well as dynamic obstacle avoidance. Furthermore, by employing the Hindsight Experience Replay (HER) technique, the efficiency an‎d success rate of SAC, along with its capability in challenging scenarios, are examined. Nevertheless, the inherent limitations of reinforcement learning policies in physical interactions are criticized, an‎d a hybrid control architecture based on an impedance control layer is proposed as a solution to enhance the safety an‎d robustness of the system. Simulation validation shows that this architecture can generate appropriate responses when facing unintended collisions while simultaneously accomplishing the primary tasks with high accuracy. Subsequently, the Franka Emika Pan‎da robot with 7 degrees of freedom is initialized an‎d set up from scratch to provide the foundation for the practical implementation of the algorithms. Moreover, due to existing constraints, an obstacle avoidance algorithm based on Quadratic Programming (QP) is selec‎ted an‎d eva‎luated in a real environment. Practical results indicate that the proposed controller not only maintains stability, safety, an‎d efficiency in simulation conditions but also preserves these properties when confronted with real-world hardware challenges.

6