Bridging the Reality Gap Between Virtual and Physical Environments Through Reinforcement Learning

Creating Reinforcement learning(RL) agents that can perform tasks in the real-world robotic systems remains a challenging task due to inconsistencies between virtual-and real-world. This is known as “reality-gap” which hinders performance of RL agent trained virtual environment. The research describes techniques used train models, generate randomized environments, reward function, and utilized transfer model physical environment for evaluation. For this investigation, low-cost 3-degrees-of-freedom (DOF) Steward platform was 3D modeled created virtually physically. goal 3D-Stewart guide balance marble towards center. Custom end-to-end APIs were developed interact with Godot game engine, manipulate physics dynamics, in-game lighting randomizations. Two algorithms: Q-learning Actor-Critic, implemented evaluate by using domain randomization induced noise bridge reality gap. Q-learning, raw frames make predictions while Actor-Critic position, velocity vector relative position pre-processing captured frames. experimental results show effectiveness introduction during training.