Analysis of Affordance Detection Methods for Real-World Robotic Manipulation

摘要

Understanding the set of actions an object affords is the first step to granting robots the intelligence required to operate in everyday environments. These actions, called affordances, are intrinsic to the structure and properties of an object. In this paper, we present a novel summary of the state-of-the-art affordance detection approaches. Building on the object segmentation problem that is common in the computer vision community, affordance detection seeks to assign possible actions according to an object's class using convolutional neural networks that are trained on 2D or 3D image data. We compare the performance of recent approaches to the state-of-the-art and identify areas in need of further research.