The Columbia Grasp Database - Part I¶

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, , presents the database itself in much more detail, discusses some of its design choices, shows you how to extend it, etc. Finally, see the section for a complete discussion of theoretical aspects and research questions.

Formally, the database contains:

• the set of objects used is the Princeton Shape Benchmark (PSB). This set consists of approx. 1800 3D models, and is separate from GraspIt. However, it is freely available for download (see below for installation instructions).
• since grasping is inherently scale-dependant, in the CGDB, each of these 1,800 models is used at 4 different scales. Therefore, the CGDB can be considered to use approx. 1,800 * 4 = 7,200 individual 3D models.
• for each of these models, the CGDB contains a number of grasps. A grasp is defined as:
  • a hand position relative to the object
  • a set of hand joint angles
• quality metrics for all the grasps. Note that all of the
• hand-object contact points for all the grasps, for use in machine learning grasps from the CGDB have form-closure.
• grasps have been stored in the database for 3 hand models:
  • the Barrett hand assuming plastic finger covers (low friction coefficient)
  • the Barrett hand assuming rubber finger covers (high friction coefficient)
  • the Human Hand model
• in total, the CGDB contains more than 250,000 grasps.

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

• load any model from the PSB into GraspIt!, automatically rescaled to a size appropriate for grasping.
• see the other models in the PSB that are “geometrically similar” to the loaded model.
• load the grasps for that model that have been saved in the CGDB, using any of the three hand models above.
• use this information for data-driven grasp planning.

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.

Database installation¶

The CGDB comes as a separate download from GraspIt!, at the . 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.

GraspIt Interface Installation¶

Connecting to and browsing the database¶

Connecting to the CGDB:

• start GraspIt and load one of the hand models that are used in the CGDB (the Human hand or the Barrett hand). You don’t need to explicitly load any objects.
• The Database menu gives you access to all CGDB functionality from GraspIt. The first step is to establish a connection to the CGDB, using the Database-> Connect and Browse menu. After a connection is established, you can use the other functions in the Database menu as well. Click Database->Connect and Browse to bring up the CGDB-Browser dialog.
• set the connection parameters based on PostgreSQL server that you are connecting to. You must have access to a machine running a PostgreSQL server that serves the CGDB, as downloaded above. Most often, the machine running the server will be the same that GraspIt is running on; in this case, set the Host to localhost. The Port number is usually 5432. Set the User Name and Password based on the settings that you used when setting up your PostgreSQL server.
• note that it is also possible to connect to a remote machine running the PostgreSQL server. In the future, we might set up a machine in the Columbia Robotics lab that will offer a read-only version of the database for everybody. For now though, you have to set up your own PostgreSQL server.
• click Connect. If the connection is successful, the Browser group will become enabled, the Models drop-down list will be populated with a list of models and the currently selected model thumbnail will be shown in the dedicated space.
• if the connection fails, you will get an error message in the console. Usually, this error message will say that the Qt SQL driver is not properly loaded. Go back to the section for details.

Browsing the CGDB:

• select a model from the Models drop-down list. The model’s thumbnail will be shown. Note that the names of the models from the PSB are set as follows: psb_scale_modelnumber. Each model can be selected at one of four different scales (0.75, 1.0, 1.25 or 1.5, each compared to a reference size which has been determined empirically by us).
• click Load Model. The model will be loaded into GraspIt and set as the reference object for grasp quality computations.
• select the type of grasps that you want to load. There are three grasp types stored in the database:
  • EIGENGRASPS - this is the actual database, the grasps computed by our automatic planner. You will find this type of grasps for all the objects in the database. We recommend using only this type of grasps when using the CGDB.
  • HUMAN - grasps created by a human operator. This are being used for a current unfinished project in the Columbia Robotics Lab. For the moment, there are very few grasps of this type saved in the CGDB, approx 100 grasps over a set of 15 objects. In the future, the CGDB might include more grasps of this type, for comparison against the automatically generated grasps above.
  • HUMAN_REFINED - the same operator-created grasps, but further refined inside GraspIt by an operator. Again, very few of these are available.
• click Load Grasps and use the Grasps button group to browse the loaded grasps. Note that, for each grasp, you can look at either the final grasp posture, or the pre-grasp. See the CGDB Publications for the difference between these two.

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 section.

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:

• establish a connection to the CGDB as discussed above.
• load a model from the CGDB. This will serve as the target object, the one that we will plan grasps on.
• start the planner using the Database->Database Planner menu.

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

• Neighbor Generator
  • select the distance function used for finding geometric neighbors. Use SIFT_12_view_0 for our most up-to-date results on shape search. For more details, please see the CGDB publications.
  • choose the number of neighbors you want to use. Usually, a number between 3 and 5 does the job, but you can choose to use more of fewer neighbors based on your computational resources. Note that the more neighbors you use, the worse the “geometric matching” will get.
  • click Get Neighbors. The drop-down list of neighbors will be populated and you can browse through it and see the associated thumbnails.
• Alignment - Choose a method for aligning neighbors to the original model. SIFT_PI_ICP_FULL gives the best results.
• Grasp retrieval and ranking. Here we retrieve the grasps from the neighbors and rank them. Click Retrieve Grasps. Currently, the ranking methods we are using are still under development, so choose No ranking, then click Rank Grasps. After you have done this, the Selected Grasps group should become enabled, and the counter will show you the total number of grasps retrieved from neighbors.
• Grasp Browsing - you can browse through the list of retrieved grasps in the Selected Grasps group. Make sure that aligned original grasp is selected.
• Grasp execution - here we execute the retrieved grasps on the target object, and compute their quality. Make sure Static is selected in the grasp execution type drop down box (the other option, Dynamic execution, is still under development). You can test the grasps one at a time, by selecting Test current, or all at once, by selecting Test All. Note that this option will usually take up to a minute of computation, depending on the number of neighbors and the number of retrieved grasps.
• Solution inspection - after using the Test All option, you can browse the results of the planner, sorted by quality. Use the Selected Grasps group, but this time select tested grasps in the radio box.

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.