Accelerated Robot Skill Acquisition by Reinforcement Learning-Aided Sim-to-Real Domain Adaptation

Since robot learning takes a lot of time and might be too time-consuming to perform on real robots, the learning process is often done in simulation. Bridging the sim-to-real gap, however, still requires real-world effort that might impose severe limitations due the still large number of the required iterations. In this paper we explore how to accelerate real-world skill learning with sim-to-real transfer learning. We propose a new methodology based on dimensionality reduction with deep autoencoders, followed by domain adaptation. The latter combines re-training of the complete deep neural network with real data and adaptation of a part of the network parameters with reinforcement learning. The results of our experiments demonstrate that a considerable acceleration of real-world learning is achieved by approach when learning complex robot skills.