17 The Columbia Grasp Database: Part IITop15 Hardware connections16 The Columbia Grasp Database - Part I

Contents

16 The Columbia Grasp Database - Part I

16.1 Overview

The Columbia Grasp Database (CGDB) is a repository of more than 250,000 grasps over a set of approximately 8,000 3D models. The interface discussed here allows you to load these models into GraspIt, load and inspect the grasps for these objects stored in the database and use them for data-drive, machine-learning based algorithms inside GraspIt.

This chapter is dedicated to the interface between the CGDB and GraspIt. It will show you how to install and use this interface. The next chapter, The Columbia Grasp Database - Part II, presents the database itself in much more detail, discusses some of its design choices, shows you how to extend it, etc. Finally, see the Publications section for a complete discussion of theoretical aspects and research questions.

Formally, the database contains:

The interface between GraspIt and the CGDB will enable you to:

16.2 Caveats

An aspect of particular importance concerns the quality of the information available in the CGDB. Many of the objects in the PSB are not intuitively graspable. You will find extensive sets of ants and spiders, battleships, furniture etc. For the purposes of the CGDB, all objects have been rescaled to what we empirically considered to be "graspable" size. Informally, this is approximately the size of a toy model of the object.

Since the majority of the grasps in the database were found using an automated planner, not all of the grasps are truly humanlike or reliable. There can be cases where a grasp satisfies our quality metrics, but would require a degree of precision that cannot be obtained in real-life execution. Aside from the intrinsic limitations of grasp quality metrics, for which there is as of yet no firm consensus on which to use, our approach to grasp planning is purely geometric. This presents problems for objects that do not match our assumptions. For example, our assumption that all objects are rigid plastic results in geometrically correct but unrealistic grasps on objects such as flowers or leaves. Furthermore, the lack of domain-specific knowledge means that some of our grasps are semantically incorrect, such as a mug grasped by placing the fingers inside.

Finally, our automatically computed grasps were obtained from pre-grasps that sample a low-dimensional subspace of the hand DOF space. This is for the moment a necessary simplification, without which the planning problem for dexterous hands is intractable at this scale. While our choice of subspace is theoretically justified and shown to be effective [3], we cannot reasonably claim that the database covers the entire space of possible grasps.

16.3 Database installation

The CGDB comes as a separate download from GraspIt!, at the Columbia Robotics website. It is available for download as a PostgreSQL database backup file (70M) and requires a PostgreSQL installation to use. In order to use the interface presented here, you will need to install a PostgreSQL server on your machine and load the provided backup file. PostgreSQL is open source, and easy to install. We recommend getting the binary package.

16.4 GraspIt Interface Installation

Windows

Linux

16.5 Connecting to and browsing the database

Connecting to the CGDB:

Browsing the CGDB:

16.6 Geometric similarity

A very important aspect of the CGDB concerns geometric similarity between objects. Our group has been implementing existing tools, and also developing new methods for this area. These tools are not included with GraspIt! - this means that, if you have a new 3D model, this interface will not be able to find its geometric "neighbors" in the CGDB. However, we have precomputed this information for all the models that are already part of the CGDB, and included this information in the CGDB. This means that, for any model in the CGDB, you can see which other models are "geometrically similar" based on our set of tools.

Geometric similarity is a vibrant research area, far exceeding the scope of this user manual. For more details, please see the Publications section.

16.7 Database-backed grasp planning

Database-backed grasp planning works by finding geometrically similar "neighbor" objects to the target object in the CGDB. Due to the reasons explained above, the version of the planner included with this distribution only works for models that are already in the CGDB and as such have pre-computed neighbor information. However, we are providing this code in the hope that it will serve as a blueprint for developing your own CGDB-backed algorithms.

Please note that this is a fairly complex machinery, and all the details of its execution exceed the scope of this chapter. More information is provided in the next chapter, which discusses advanced concepts pertaining to the CGDB. You might need to peruse the code itself, and its documentation for more details.

To start, use the following steps:

The CGDB Planner goes through a few steps, each with its own dedicated button group in the Planner dialog.

This is just a very high-level overview and walk-through for the CGDB-backed planner. For more details, please see the next chapter of this manual, the source code documentation, the Publications chapter, or contact us.


Copyright (C) 2002-2009 Columbia University


17 The Columbia Grasp Database: Part IITop15 Hardware connections16 The Columbia Grasp Database - Part IContents