Training an Agent to Ground Commands with Reward and Punishment

摘要

As robots and autonomous assistants becomemore capable, there will be agreater need for humans to easilyconvey to agents the complex tasks they wantthem to carry out. Conveying tasks throughnatural language provides an intuitive interfacethat does not require any technical expertise,but implementing such an interface requires methods forthe agent to learn a grounding of natural language commands.In this work, we demonstrate how high-level task groundings can belearned from a human trainer providing online reward and punishment.Grounding language to high-level tasks for the agent to solveremoves the need for the human to specify low-level solution details intheir command.Using reward and punishment for trainingmakes the training procedure simple enough to be used by people withouttechnical expertise and also allows a human trainer to immediatelycorrect errors in interpretation that the agent has made. We present preliminary results from a single usertraining an agent in a simple simulated home environment and showthat the agent can quickly learn a grounding oflanguage such that the agent can successfully interpretnew commands and executethem in a variety of different environments.