BURG: Benchmarks for UndeRstanding Grasping

Lead Research Organisation: University of Birmingham
Department Name: School of Computer Science

Abstract

Grasping rigid objects has been reasonably studied under a wide variety of settings. The common measure of success is a check of the robot to hold an object for a few seconds. This is not enough. To obtain a deeper understanding of object manipulation, we propose (1) a task-oriented part-based modelling of grasping and (2) BURG - our castle of setups, tools and metrics for community building around an objective benchmark protocol.

The idea is to boost grasping research by focusing on complete tasks. This calls for attention on object parts since they are essential to know how and where the gripper can grasp given the manipulation constraints imposed by the task. Moreover, parts facilitate knowledge transfer to novel objects, across different sources (virtual/real data) and grippers, providing for a versatile and scalable system. The part-based approach naturally extends to deformable objects for which the recognition of relevant semantic parts, regardless of the object actual deformation, is essential to get a tractable manipulation problem. Finally, by focusing on parts we can deal easier with environmental constraints that are detected and used to facilitate grasping.

Regarding benchmarking of manipulation, so far robotics suffered from incomparable grasping and manipulation work. Datasets cover only the object detection aspect. Object sets are difficult to get, not extendible, and neither scenes nor manipulation tasks are replicable. There are no common tools to solve the basic needs of setting up replicable scenes or reliably estimate object pose.

Hence, with the BURG benchmark we propose to focus on community building through enabling and sharing tools for reproducible performance evaluation, including collecting data and feedback from different laboratories for studying manipulation across different robot embodiments. We will develop a set of repeatable scenarios spanning different levels of quantifiable complexity that involve the choice of the objects, tasks and environments. Examples include fully quantified settings with layers of objects, adding deformable objects and environmental constraints. The benchmark will include metrics defined to assess the performance of both low-level primitives (object pose, grasp point and type, collision-free motion) as well as manipulation tasks (stacking, aligning, assembling, packing, handover, folding) requiring ordering as well as common sense knowledge for semantic reasoning.

Planned Impact

N/A
 
Title Grasping dataset YCB-76 
Description This is a dataset for evaluation of grasping from object point clouds. It contains 76 objects from the renowned YCB object set, which are arranged in 259 distinct grasping scenarios. For each scenario, we provide point clouds from synthetically created depth images. 
Type Of Material Database/Collection of data 
Year Produced 2022 
Provided To Others? Yes  
Impact This dataset has been used throughout our collaboration "end-to-end learning to grasp from object point clouds". It is suitable for evaluation purposes and specifically for analysing the capabilities of learning-based grasping methods to generalise to novel, unseen shapes. The performance is evaluated using simulation-based success rates. 
URL https://github.com/antoalli/L2G
 
Title Grasping dataset YCB-8 
Description This dataset is a synthetic dataset for robotic grasping from point clouds. It is made using 8 objects from the renowned YCB object set and contains 15 grasping scenarios with each object being in a certain resting pose. For each scenario, we provide point clouds from 300 different, synthetically rendered depth images as well as 100k grasp annotations. 
Type Of Material Database/Collection of data 
Year Produced 2022 
Provided To Others? Yes  
Impact This dataset has been used throughout our collaboration "end-to-end learning to grasp from object point clouds". It is suitable for evaluation purposes and specifically for analysing the capabilities of learning-based grasping methods to generalise to novel, unseen shapes. We can produce simulation-based success measures and, using the grasp annotations, also measures like coverage, i.e. which proportion of the annotated grasps could be reproduced by the method. 
URL https://github.com/antoalli/L2G
 
Description End-to-end learning to grasp from point clouds 
Organisation Italian Institute of Technology (Istituto Italiano di Tecnologia IIT)
Country Italy 
Sector Academic/University 
PI Contribution In this collaboration we are investigating the problem of grasping objects with a 2-finger parallel-jaw gripper. More specifically, the core of the problem is to identify suitable 6-DOF grasp poses based on a partial point cloud of the object obtained from a single-view depth image, for which we develop novel Deep-Learning-based methods. We contributed to the conceptual development of the method (e.g. used representations, network architecture) with our expertise in robotic grasping and 6-DOF object pose estimation, which is a related problem that often utilises similar Deep Learning methods based on point cloud representations. We created datasets that can be used for training as well as for simulation-based evaluation of the proposed grasps. In connection with the datasets, we also contributed tools to evaluate individual grasps in a physics-based simulation environment as well as tools and metrics for thorough evaluation of sets of grasp predictions. Furthermore, we implemented an experimental setup for real robot grasping trials and conducted the required experiments. In this collaboration, we also made use of equipment (e.g. robot, sensors, test objects, GPUs) in our robotics lab which has been acquired from earlier funding (not within this grant).
Collaborator Contribution Our partners have expertise in Deep Learning on point clouds, specifically encoding local as well as global object features, which has proven to be valuable for tasks such as shape completion (i.e. estimation of the unobserved part of the object). We consider this very valuable also in the context of robotic grasping. Besides the conceptual development of the method, our partners also contributed by implementing the model, designing experiments and ablation studies, and conducting the majority of simulation-based experiments using their high-performance computing equipment.
Impact The outcomes are partially published. Already published and described in the corresponding sections are: - Datasets: YCB-8, YCB-76 - Software/Tools: Grasp Evaluator Published by our collaborators: - Software: L2G (Learning to Grasp) method, central entry point to this work, also references all relevant tools and datasets; see https://github.com/antoalli/L2G Submitted but not published: - Paper "End-to-End Learning to Grasp from Object Point Clouds" submitted to RA-L/IROS This collaboration is not multi-disciplinary.
Start Year 2020
 
Description End-to-end learning to grasp from point clouds 
Organisation Polytechnic University of Turin
Country Italy 
Sector Academic/University 
PI Contribution In this collaboration we are investigating the problem of grasping objects with a 2-finger parallel-jaw gripper. More specifically, the core of the problem is to identify suitable 6-DOF grasp poses based on a partial point cloud of the object obtained from a single-view depth image, for which we develop novel Deep-Learning-based methods. We contributed to the conceptual development of the method (e.g. used representations, network architecture) with our expertise in robotic grasping and 6-DOF object pose estimation, which is a related problem that often utilises similar Deep Learning methods based on point cloud representations. We created datasets that can be used for training as well as for simulation-based evaluation of the proposed grasps. In connection with the datasets, we also contributed tools to evaluate individual grasps in a physics-based simulation environment as well as tools and metrics for thorough evaluation of sets of grasp predictions. Furthermore, we implemented an experimental setup for real robot grasping trials and conducted the required experiments. In this collaboration, we also made use of equipment (e.g. robot, sensors, test objects, GPUs) in our robotics lab which has been acquired from earlier funding (not within this grant).
Collaborator Contribution Our partners have expertise in Deep Learning on point clouds, specifically encoding local as well as global object features, which has proven to be valuable for tasks such as shape completion (i.e. estimation of the unobserved part of the object). We consider this very valuable also in the context of robotic grasping. Besides the conceptual development of the method, our partners also contributed by implementing the model, designing experiments and ablation studies, and conducting the majority of simulation-based experiments using their high-performance computing equipment.
Impact The outcomes are partially published. Already published and described in the corresponding sections are: - Datasets: YCB-8, YCB-76 - Software/Tools: Grasp Evaluator Published by our collaborators: - Software: L2G (Learning to Grasp) method, central entry point to this work, also references all relevant tools and datasets; see https://github.com/antoalli/L2G Submitted but not published: - Paper "End-to-End Learning to Grasp from Object Point Clouds" submitted to RA-L/IROS This collaboration is not multi-disciplinary.
Start Year 2020
 
Description SetupTool for benchmarking robotic grasping 
Organisation Spanish National Research Council (CSIC)
Country Spain 
Sector Public 
PI Contribution In this collaboration, we created a tool for arranging scenes for both simulated and real experiments. An intuitive GUI allows the arrangement of the 3d object models, while a physical simulation engine ensures physical plausibility of the scene. This can be used both in grasp simulations and to create printouts indicating the poses of the objects for arranging the real objects. We built upon our expertise in robotic grasping of rigid objects and worked on the design and implementation of the back-end of the software, which provides the core functionalities in the form of a Python package.
Collaborator Contribution Our partners at TU Vienna brought in their experience in building visual tools for e.g. labeling of object poses. They designed and implemented the front-end of the software, which handles all user interaction and allows intuitive arrangement of the objects. Our partners at CSIC Barcelona have expertise in grasping deformables, in particular cloth-like objects. They provided intellectual input on how to integrate such deformable objects into the SetupTool and provided valuable feedback throughout the development and testing stages.
Impact Outcomes that are published by us and described in the corresponding sections are: - Software: BURG Toolkit as back-end of the SetupTool Outcomes that are published by our partners: - Software: SetupTool GUI as front-end of the SetupTool This collaboration is not multi-disciplinary.
Start Year 2020
 
Description SetupTool for benchmarking robotic grasping 
Organisation Vienna University of Technology
Country Austria 
Sector Academic/University 
PI Contribution In this collaboration, we created a tool for arranging scenes for both simulated and real experiments. An intuitive GUI allows the arrangement of the 3d object models, while a physical simulation engine ensures physical plausibility of the scene. This can be used both in grasp simulations and to create printouts indicating the poses of the objects for arranging the real objects. We built upon our expertise in robotic grasping of rigid objects and worked on the design and implementation of the back-end of the software, which provides the core functionalities in the form of a Python package.
Collaborator Contribution Our partners at TU Vienna brought in their experience in building visual tools for e.g. labeling of object poses. They designed and implemented the front-end of the software, which handles all user interaction and allows intuitive arrangement of the objects. Our partners at CSIC Barcelona have expertise in grasping deformables, in particular cloth-like objects. They provided intellectual input on how to integrate such deformable objects into the SetupTool and provided valuable feedback throughout the development and testing stages.
Impact Outcomes that are published by us and described in the corresponding sections are: - Software: BURG Toolkit as back-end of the SetupTool Outcomes that are published by our partners: - Software: SetupTool GUI as front-end of the SetupTool This collaboration is not multi-disciplinary.
Start Year 2020
 
Title BURG Toolkit 
Description The BURG Toolkit is a Python package for Benchmarking and Understanding Robotic Grasping. The main features supported by the toolkit are: (i) core data structures for object libraries, scenes, grippers, grasps, and other fundamental constructs related to grasping; (ii) physical grasp simulation using the PyBullet physics engine; (iii) ability to provide printouts that can be used to arrange objects in the physical world in configurations that match those in the simulated environment (for experimental evaluation); (iv) ability to create datasets, especially sampling grasps based on 3D object model and rendering of (depth) images of scenes; and (v) visualisation of scenes and grasps for experimental evaluation. The toolkit can be used as a stand-alone Python package or with the BURG Toolkit GUI as an User Interface, which has been developed by project collaborators from TU Vienna. 
Type Of Technology Software 
Year Produced 2022 
Open Source License? Yes  
Impact The software has just been released so there are no impacts to report yet. We plan to promote the software at academic conferences and in our research networks. 
URL https://github.com/mrudorfer/burg-toolkit
 
Title Grasp Evaluator 
Description The grasp evaluator is a publicly available Python package for evaluating the grasps provided by different methods for estimating grasps. It is designed to work with multiple existing grasping datasets, including our own YCB-8 and YCB-76 datasets. It provides a variety of metrics and measures for evaluation that are based on simulation as well as comparison (if grasp annotations are available in the target dataset). 
Type Of Technology Software 
Year Produced 2022 
Impact The grasp evaluator supports a detailed analysis of grasp predictions, which helps provide a better understanding of the strengths and weaknesses of state of the art grasping methods. It also helps improve existing methods as well as develop new methods with the desired capabilities. This software has been used for the duration of our collaborative effort: "end-to-end learning to grasp from object point clouds", and it played an important role in enabling the related scientific contributions. It is publicly available on GitHub to encourage other researchers to use it as well. 
URL https://github.com/mrudorfer/grasp-evaluator