From Natural Language Instructions to Structured Robot Plans.

Research into knowledge acquisition for robotic agents has looked at interpreting natural language instructions meant for humans into robot-executable programs; however, the ambiguities of natural language remain a challenge for such "translations". In this paper, we look at a particular sort of ambiguity: the control flow structure of the program described by the natural language instruction. It is not always clear, when more conditional statements appear in a natural language instruction, which of the conditions are to be thought of as alternative options in the same test, and which belong to a code branch triggered by a previous conditional. We augment a system which uses probabilistic reasoning to identify the meaning of the words in a sentence with reasoning about action preconditions and effects in order to filter out non-sensical code structures. We test our system with sample instruction sheets inspired from analytical chemistry.