Learning task goals interactively with visual demonstrations

Abstract

Humans are extremely good at quickly teaching and learning new tasks through situated instructions; tasks such as learning a novel game or household chore. From studying such instructional interactions, we have observed that humans excel at communicating information through multiple modalities, including visual, linguistic, and physical ones. Rosie is a tabletop robot implemented in the Soar architecture that learns new tasks from online interactive language instruction. In the past, the features of each task’s goal were explicitly described by a human instructor through language. In this work, we develop and study additional techniques for learning representations of goals. For game tasks, the agent can be given visual demonstrations of goal states, refined by human instructions. For procedural tasks, the agent uses information derived from task execution to determine which state features must be included in its goal representations. Using both approaches, Rosie learns correct goal representations from a single goal example or task execution across multiple games, puzzles, and procedural tasks. As expected, in most cases, the number of words required to teach the task is reduced when visual goal demonstrations are used. We also identify shortcomings of our approach and outline future research.