First Order Logic for Command Understanding

When given requests in natural language, ideally, the robot should have a single interpretation of that command. However, the more generic the request, the more potential execution pathways it may entail. To address this challenge, we introduce a two-step method designed to diminish the multitude of conceivable interpretations for natural language commands directed at a robot. In the first step, the request is translated into First Order Logic (FOL), subsequently generating interpretations that align with the agent's existing knowledge base. The interpretations are derived using MACE4, a tool for identifying finite models of FOL theories. In the second step, the agent systematically reduces the pool of interpretations hrough the formulation of informative queries directed towards the human agent. The effectiveness of these queries is assessed based on the information gain metric, adapted to suit finite interpretation models associated with FOL theories. We present a practical demonstration within the kitchen domain, wherein our approach is exemplified through the utilization of AbeSim, a robot simulation tool. Our research contribution primarily entails the utilization of the information gain metric to prioritize and rank queries based on their effectiveness in minimizing the multitude of interpretation models. This innovative methodology offers a promising solution for mitigating ambiguity within natural language commands issued to robotic systems.