I. Introduction
Machine learning methods can be used to train effective models for visual perception [1]–[3], natural language processing [4], [5], and numerous other applications [6], [7]. However, broadly generalizable models typically rely on large and highly diverse datasets, which are usually collected once and then reused repeatedly for many different models and methods. In robotics, this presents a major challenge: every robot might have a different physical configuration, such that end-to-end learning of control policies usually requires specialized data collection for each robotic platform. This calls for developing techniques that can enable learning from experience collected across different robots and sensors.