Download files
Abstract
The ability to use tools is important for humans as it enables us to do complicated tasks that would not otherwise be possible. Robots can also benefit from having such an ability, which may be pre-programmed or learned by the robot. Various approaches to tool-use learning have been proposed, mostly using a feature-based representation. They assume that a tool, suitable for the given task, exists among a pool of available tools, otherwise their methods fail. A restrictive representation is preventing them from learning complex tool actions and being able to extend learning to creating a new tool.
A relational learning system is a potential framework for tool creation because of it's expressiveness. It can capture complex relationships between objects and tools, and it can accumulate learned concepts for future learning. We propose a novel relational learning mechanism that enables a robot to not only learn to use a tool, but also to create a new tool, if needed to complete a task. The tool creation system is based on earlier work by Brown and Sammut [21] on tool-use learning, which learns by observing a single example from a teacher and carrying out a series of experiments to refine an action model. We extend this by developing new mechanisms to generate tools by performing learning by experimentation based on Inductive Logic Programming (ILP) [101].
Unlike previous studies carried out in simulations, we validate our tool creation algorithm using a real robot, in this case a Rethink Robotics Baxter. Simulation is still used, initially, to minimise real-world experiments. Three different kinds of tools (hook, wedge, and lever) are used in empirical evaluations of our learning performance. New tools are fabricated by a 3D printer. Based on the experimental results, we find that our robot is able to learn how to use a tool and, if needed, create a novel tool to solve a problem using a relatively small amount of data.