Planning Robotic Cutting Operations

Abstract

Classical planning and most PDDL variants operate on the assumption that the number and types of objects present in the environment are known at the time of initialization and neither can nor do change during plan execution. However, there are many domains in which it is helpful and necessary to be able to capture action (or environment) effects that are able to change the existence of objects rather than just facts about these objects. PDDLStream already provides a framework for "certifying" new facts about the environment as necessary throughout plan execution; I propose using PDDLStream to construct a principled way to reason over not just added facts, but also added or removed objects in the environment. In order to do this, I will work within the domain of cutting operations in the kitchen, as this is a domain that both necessitates a lot of object change as objects are cut and often requires chains of these generated objects to be fully reasoned over. Additionally, I will lay the groundwork to use this principled way to reason over new objects to implement different types of cutting operations in the kitchen, with the eventual goal of a robot planner being able to sequence different provided actions to more efficiently work with knives in the kitchen in a human-like manner.