Keywords: Object-level planning, Abstraction, Task and motion planning
TL;DR: Blue Sky Paper submission entitled "Object-Level Planning and Abstraction"
Abstract: Task and motion planning (TAMP) aims to integrate higher-level symbolic task planning with lower-level motion planning. However, task-level representations must include detailed information expressing the robot’s own constraints, mixing logical object-level requirements (e.g., a bottle must be open to be poured) with robot constraints (e.g., a robot’s gripper must be empty before it can pick up an object). We propose an additional level of planning that will rest above the current TAMP pipeline, which we refer to as object-level planning (OLP). OLP is concerned with objects, but not with the robot and its constraints. It allows the robot to generate plan sketches that describe how objects in the environment must be changed to realize a plan, leaving the details of how these are achieved to task-level planning. Object-level plans have the benefit of being interpretable and portable while supporting generalization across objects and the details of any specific task. Using functional object-oriented networks (FOONs), a representation we introduced in prior work, we will briefly outline how object-level plans can be used to model object-level manipulation planning, resulting in plan sketches that can be extracted from natural language, inform a task-level PDDL planner, and generalize across object types.