While ML researchers can easily acquire robotic interaction data through random exploration techniques, it is more challenging to get high-quality data, which usually requires human supervision to obtain. Although task-agnostic exploration may be useful in simple domains, this approach doesn’t work well in more challenging settings (like real, high-dimensional scenes). In this paper, Chen et al. present a framework, Batch Exploration with Examples (BEE), which uses weak human supervision to guide an agent’s exploration towards more relevant states (i.e. so the agent doesn’t need to explore everything). A human provides a few relevant examples of the task at hand which the agent uses to guide their exploration. They find that BEE enables faster data collection and better downstream task performance.