Plan Verbalisation for Robots Acting in Dynamic Environments

Automated planning provides the tools for intelligent be- haviours in robotic platforms deployed in real-world environ- ments. The complexity of these domains requires planning models that support the system’s dynamics. This results in AI planning approaches often generating plans where the reason- ing around the solution remains obscure for the operator/user. This lack of transparency can reduce trust, results in frequent interventions, and ultimately represents a barrier to adopting autonomous systems. Explanations of behaviour in an easy- to-understand manner, such as in natural language, can help the user comprehend the reasoning behind autonomous ac- tions and help build an accurate mental model. This paper presents an approach for a type of explanation, namely plan verbalisation, that considers the properties of the planning model and describes the system behaviour during plan execu- tion, including replanning and plan repair. We use natural lan- guage techniques to support the disambiguation of the robot decision-making process, considering the planning model en- capsulated using the Planning Domain Definition Language (PDDL). The system is evaluated using an Autonomous Un- derwater Vehicle (AUV) inspection use case.