A Dataset for Grasping and Manipulation using ROS

In this study, we introduce a household object dataset for recognition and manipulation tasks, focusing on commonly available objects in order to facilitate sharing of applications and algorithms. The core information available for each object consists of a 3D surface model annotated with a large set of possible grasp points, pre-computed using a grasp simulator. The dataset is an integral part of a complete Robot Operating System (ROS) architecture for performing pick and place tasks. We present our current applications using this data, and discuss possible extensions and future directions for shared datasets for robot operation in unstructured settings. I. DATASETS FOR ROBOTICS RESEARCH Recent years have seen a growing consensus that one of the keys to robotic applications in unstructured environments lies in collaboration and reusable functionality. An immediate result has been the emergence of a number of platforms and frameworks for sharing operational “building blocks,” usually in the form of code modules, with functionality ranging from low-level hardware drivers to complex algorithms such as path or motion planners. By using a set of now well-established guidelines, such as stable documented interfaces and standardized communication protocols, this type of collaboration has accelerated development towards complex applications. However, a similar set of methods for sharing and reusing data has been slower to emerge. In this paper we describe our effort in producing and releasing to the community a complete architecture for performing pick-and-place tasks in unstructured (or semistructured) environments. There are two key components to this architecture: the algorithms themselves, developed using the Robot Operating System (ROS) framework, and the knowledge base that they operate on. In our case, the algorithms provide abilities such as object segmentation and recognition, motion planning with collision avoidance, grasp execution using tactile feedback, etc. The knowledge base, which is the main focus of this study, contains relevant information for object recognition and grasping for a large set of common household objects. Some of the key aspects of combining computational tools with the data that they operate on are: • other researchers will have the option of directly using our dataset over the Internet (in an open, read-only fashion), or downloading and customizing it for their own applications; • defining a stable interface to the dataset component of the release will allow other researchers to provide their own modified and/or extended versions of the data to †Willow Garage Inc., Menlo Park, CA. Email: {matei, bradski, hsiao, pbrook}@willowgarage.com ∗University of Washington, Seattle, WA. the community, knowing that it will be directly usable by anyone running the algorithmic component; • the data and algorithm components can evolve together, like any other components of a large software distribution, with well-defined and documented interfaces, version numbering and control, etc. In particular, our current dataset is available in the form of a relational database, using the SQL standard. This choice provides additional benefits, including optimized relational queries, both for using the data on-line and managing it off-line, and low-level serialization functionality for most major languages. We believe that these features can help foster collaboration as well as provide useful tools for benchmarking as we advance towards increasingly complex behavior in unstructured environments. There have been previous example of datasets released in the research community (as described for example in [3], [7], [13] to name only a few), used either for benchmarking or for data-driven algorithms. However, few of these have been accompanied by the relevant algorithms, or have offered a well-defined interface to be used for extensions. The database component of our architecture was directly inspired by the Columbia Grasp Database (CGDB) [5], [6], released together with processing software integrated with the GraspIt! simulator [9]. The CGDB contains object shape and grasp information for a very large (n = 7, 256) set of general shapes from the Princeton Shape Benchmark [12]. The dataset presented here is smaller in scope (n = 180), referring only to actual graspable objects from the real world, and is integrated with a complete manipulation pipeline on the PR2 robot. II. THE OBJECT AND GRASP DATABASE