Abstract: Hierarchical plan validation gets a sequence of actions and determines if it is a valid hierarchical plan, that is, it is executable and can be obtained by decomposing some task. If the plan is valid, then a proof in the form of the task decomposition structure can be released. However, if the plan is not valid, plan validators simply state this. This is where plan correction may help by suggesting modifications to invalid plans, such as inserting or deleting actions, to make them valid hierarchical plans. Although the original motivation for plan correction was to enhance plan validation, the concept proved to apply to various other problems, including plan recognition, plan repair, and even planning itself.

The talk illustrates a journey from plan validation to plan correction, advocating plan correction as a general concept that encompasses various hierarchical planning tasks. As a future research direction, it suggests that a similar approach can be applied to planning domain models to obtain truly autonomous planning agents capable of learning and maintaining models of own behavior.

Abstract: Goal Recognition is the task of inferring an agent’s goal from a sequence of observations. It has been studied in a wide range of real-world domains, including human–robot interaction, smart homes, autonomous driving, and surveillance. In this talk, we will explore how Goal Recognition has been addressed within the Automated Planning community, highlighting the different assumptions and methodologies that have shaped the field. Building on these foundations, we will then discuss the emergence of reinforcement learning and deep learning approaches. Finally, we will look at possible future directions for the field, particularly in light of the recent breakthrough of Large Language Models and their growing impact across Artificial Intelligence.