Humans can learn from teachers by observing gestures, reinforcements, and words, which we collectively call signals. Often a new signal will support a different interpretation of an earlier signal, resulting in a different belief about the task being learned. If robots are to learn from these signals, they must perform similar inferences. We propose the use of Bayesian inference to allow robots to learn tasks from human teachers. We review Bayesian inference, describe its application to the problem of learning high-level tasks from a human teacher, and work through a specific implementation on a robot. Bayesian inference is shown to quickly converge to the correct task specification.
展开▼
