Reinforcement Learning for Robotic Applications with Vision Feedback

The presence of robots in our daily life is becoming more common, where robots start carrying out more complex tasks. This increase in the complexity of tasks makes conventional control system insufficient. Therefore, a plausible approach is required for robots to learn how to perform these tasks. Reinforcement learning enables robots to perform complex tasks without highly engineered control systems. However, using reinforcement learning in robotic applications is challenged by several problems such as high dimensionality. Thus, in this paper, we study the performance of the Hindsight Experience Replay (HER) algorithm which addresses the high dimensionality problem. In this paper, we analyze the algorithm performance using a simulated robotic arm to pick and place different objects. Then, we propose the use of vision feedback which is used to control the gripper of the robotic arm. The results and analysis highlights some of HER limitations when dealing with objects that have limited grasping points. Our proposed method allows the robotic arm to pick objects using the same trained policy without the need to retrain the agent for new objects. Finally, we prove that using our method the robotic arm can pick the objects with higher success rate compared to the one without vision feedback.