AI Planning from Natural-Language Instructions for Trustworthy Human-Robot Communication

Having deterministic communication between humans and robots is essential for a safe, reliable, and trustworthy workspace. Despite the extensive efforts in training robots to comprehend human instructions, the predominant focus has been on improving the generative aspects of models rather than the determinism. This paper presents a frame-based method using planning language and Controlled Robot Language to construct a reliable and deterministic linguistic channel. The model takes multiple instructions as input and generates an appropriate syntactic and formal representation. Core information is extracted from formal representation using bottom-up visitors. The obtained information is used to generate a planning script in Planning Domain Definition Language (PDDL), which can be used directly to control the robot system. The experiment demonstrated great performance of the proposed method on many text-processing tasks and promising results in deterministic communication with robots on a manually created new dataset focusing on the robotic domain.