Deep Reinforcement Learning for Object Displacement with a Highly Maneuverable Robot Using TQC-HER Algorithm

Robots were designed to aid humans in tasks that were repetitive and/or dangerous. Classical robotic control methods (such as PIDs) show little adaptability in difficult tasks. Deep reinforcement learning is a machine learning approach for finding an optimized agent via trial and error. This research explores the application of deep reinforcement learning (DRL) algorithms to perform a pick and place task with a robotic arm attached to a moving platform. The study focuses on the use of state-of-the-art RL algorithms, including Truncated Quantile Critics (TQC) and Hindsight Experience Replay (HER), to train an agent in a simulated environment. The paper discusses the robotic environment, the task, the training agent, and presents the results obtained. The findings demonstrate the effectiveness of the RL algorithms in enabling the agent to learn and execute the manipulation task successfully. The research also highlights the importance of the chosen reward function in enhancing the sample efficiency of the training algorithm. The paper concludes with proposed future works, including the use of non-holonomic bases for the mobile platform and the exploration of agents with recurrent neural networks for improved performance.

کلیدواژه ها:

deep reinforcement learning ، robotics ، robotic manipulation ، AI in robotics