Abstract

In this minor project, together with Univerity of Borås we made a pre-study looking into using reinforcement learning in Isaac SIM to teach a robot to handle realisticly simulated cloth. I was responsible for the reinforcement learning aspect. The control and learning loop consist of a Deep Deterministic Policy Gradient together with a Hindsight Experience Replay buffer. The task is also split into subtasks such as approach, grasp and retract. We teach the robot each task separately, starting with approach and then grasp. Each subtask has its own agent that controls the robot. The agent hand over control to the next agent when a specific goal has been reached.