A Task and Motion Planning Framework for Partially Observable Household Manipulation Scenes

Task and Motion Planning (TAMP) that integrates task planning with motion planning endows robots with the ability to directly achieve abstract objectives under specified initial and goal states. However, in more practical scenarios like partially observable household scenes, the clutter and uncertainty of object placements, coupled with the often nonspecific goal states provided by human users, pose significant challenges. Conventional TAMP planners struggle to efficiently navigate these spaces based on nonspecific goals and also find it difficult to manage environments cluttered with irrelevant objects. To address these difficulties, this article introduces a TAMP framework based on a broadly used TAMP planner PDDLStream, requiring only nonspecific goal states to complete tasks. This framework employs Task‐level Backward Search to identify relevant objects for task completion, thus balancing autonomous exploration with the use of known information. Additionally, it uses an Object Manipulation Constraint Graph to generate a safe, collision‐free sequence for handling cluttered objects. The experiments demonstrate the effectiveness of the proposed framework in addressing TAMP tasks within partially observable, cluttered household environments with nonspecific goal states.