Instance-based learning of affordances

Abstract

The discovery of possible interactions with objects is a vital part of an exploration task for robots. An important subset of these possible interactions are affordances. Affordances describe what a specific object can afford to a specific agent, based on the capabilities of the agent and the properties of the object in relation to the agent. For example, a chair affords a human to be sat-upon, if the sitting area of the chair is approximately knee-high. In this work, an instance-based learning approach is made to discover these affordances solely through different visual representations of point cloud data of an object. The point clouds are acquired with a Microsoft Kinect sensor. Different representations are tested and evaluated against a set of point cloud data of various objects found in a living room environment.