A Unified Framework Integrating Object-oriented Task Learning and Knowledge-based Task Planning for Long-horizon Cooking Tasks

Abstract

Many tasks for service robots are complex and require lengthy processes. Task planning methods are widely used to address such challenges, but search-based planners are often inflexible, while learning-based planners do not guarantee feasibility. To overcome these limitations, we propose a hierarchical framework that integrates a knowledge base, a learning-based object-oriented task planner, and a symbolic robot task planner. The object-oriented task planner predicts subgoals, defined as changes in object states, from only a recipe name and a list of ingredients. The symbolic robot task planner then generates a feasible sequence of high-level robot actions using the proposed object knowledge base. Our framework focuses on high-level symbolic task planning and demonstrates generalization and feasibility across diverse recipes and action sets. We focus on cooking as a representative long-horizon domain, where sequential dependencies and embodiment-specific constraints naturally arise. Experimental validation was conducted on 20 representative recipes with 20,000 generated task samples, demonstrating robust performance across diverse cooking scenarios.