Sim-to-Real Transfer for Robotics Using Model-Free Curriculum Reinforcement Learning

In this paper we propose a novel approach for transfer of model-free Reinforcement Learning (RL) methods from simulation to a real-world model of a six-link robotic arm for industrial tasks. We develop an alternative approach to training robots for industrial tasks solely using a simulation of the system instead of the physical model itself as direct training on the physical robot models is a time and energy-consuming endeavor. The use of simulations not only ensures safe and controlled training implementation but also permits parallelization of the environment for improved training speed. As actions learned through the reinforcement learning approach in simulation often underperform in the real world because of the considerable gap in modeling real-world physical effects involved in robot actuation we therefore focuses on developing a iterative learning scheme using increasingly more high-fidelity models of the robot and the subsequent transfer. The experimental results prove the effectiveness of the proposed approach where combining the fine-tuned RL algorithm with the high-fidelity simulation of the robot resulted in a hundred percent task completion rate in simulation and subsequently in the real world.