A Relational Approach to Tool-Use Learning in Robots

Abstract

We present a robot agent that learns to exploit objects in its environment as tools, allowing it to solve problems that would otherwise be impossible to achieve. The agent learns by watching a single demonstration of tool use by a teacher and then experiments in the world with a variety of available tools. A variation of explanation-based learning (EBL) first identifies the most important sub-goals the teacher achieved using the tool. The action model constructed from this explanation is then refined by trial-and-error learning with a novel Inductive Logic Programming (ILP) algorithm that generates informative experiments while containing the search space to a practical number of experiments. Relational learning generalises across objects and tasks to learn the spatial and structural constraints that describe useful tools and how they should be employed. The system is evaluated in a simulated robot environment.