Learning Invariant Feature Spaces to Transfer Skills with Reinforcement Learning

People can learn a wide range of tasks from their own experience, but can learn from observing other creatures. This can accelerate acquisition of skills even when the observed agent differs substantially from the learning in terms of morphology. In this paper, we examine how reinforcement algorithms can transfer knowledge between morphologically different (e. g. , different robots). We introduce a problem formulation where two are tasked with learning multiple skills by sharing information. Our uses the skills that were learned by both agents to train invariant spaces that can then be used to transfer other skills from one agent to. The process of learning these invariant feature spaces can be viewed a kind of "analogy making", or implicit learning of partial correspondences two distinct domains. We evaluate our transfer learning algorithm in simulated robotic manipulation skills, and illustrate that we can transfer between simulated robotic arms with different numbers of links, as as simulated arms with different actuation mechanisms, where one robot is-driven while the other is tendon-driven.