東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Semantic Robot Programming for Taska...

Zeng, Zhen.

FindBook

Google Book

Amazon

博客來

Semantic Robot Programming for Taskable Goal-Directed Manipulation.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Semantic Robot Programming for Taskable Goal-Directed Manipulation./
作者:	Zeng, Zhen.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2020,
面頁冊數:	119 p.
附註:	Source: Dissertations Abstracts International, Volume: 81-11, Section: B.
Contained By:	Dissertations Abstracts International81-11B.
標題:	Computer science. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28006610
ISBN:	9798643184553

Semantic Robot Programming for Taskable Goal-Directed Manipulation.
Zeng, Zhen.

Semantic Robot Programming for Taskable Goal-Directed Manipulation. - Ann Arbor : ProQuest Dissertations & Theses, 2020 - 119 p.

Source: Dissertations Abstracts International, Volume: 81-11, Section: B.

Thesis (Ph.D.)--University of Michigan, 2020.

This item must not be sold to any third party vendors.

Autonomous robots have the potential to assist people to be more productive in factories, homes, hospitals, and similar environments. Unlike traditional industrial robots that are pre-programmed for particular tasks in controlled environments, modern autonomous robots should be able to perform arbitrary user-desired tasks. Thus, it is beneficial to provide pathways to enable users to program an arbitrary robot to perform an arbitrary task in an arbitrary world. Advances in robot Programming by Demonstration (PbD) has made it possible for end-users to program robot behavior for performing desired tasks through demonstrations. However, it still remains a challenge for users to program robot behavior in a generalizable, performant, scalable, and intuitive manner.In this dissertation, we address the problem of robot programming by demonstration in a declarative manner by introducing the concept of Semantic Robot Programming (SRP). In SRP, we focus on addressing the following challenges for robot PbD: 1) generalization across robots, tasks, and worlds, 2) robustness under partial observations of cluttered scenes, 3) efficiency in task performance as the workspace scales up, and 4) feasibly intuitive modalities of interaction for end-users to demonstrate tasks to robots.Through SRP, our objective is to enable an end-user to intuitively program a mobile manipulator by providing a workspace demonstration of the desired goal scene. We use a scene graph to semantically represent conditions on the current and goal states of the world. To estimate the scene graph given raw sensor observations, we bring together discriminative object detection and generative state estimation for the inference of object classes and poses. The proposed scene estimation method outperformed the state of the art in cluttered scenes. With SRP, we successfully enabled users to program a Fetch robot to set up a kitchen tray on a cluttered tabletop in 10 different start and goal settings.In order to scale up SRP from tabletop to large scale, we propose Contextual-Temporal Mapping (CT-Map) for semantic mapping of large scale scenes given streaming sensor observations. We model the semantic mapping problem via a Conditional Random Field (CRF), which accounts for spatial dependencies between objects. Over time, object poses and inter-object spatial relations can vary due to human activities. To deal with such dynamics, CT-Map maintains the belief over object classes and poses across an observed environment. We present CT-Map semantically mapping cluttered rooms with robustness to perceptual ambiguities, demonstrating higher accuracy on object detection and 6 DoF pose estimation compared to state-of-the-art neural network-based object detector and commonly adopted 3D registration methods.Towards SRP at the building scale, we explore notions of Generalized Object Permanence (GOP) for robots to search for objects efficiently. We state the GOP problem as the prediction of where an object can be located when it is not being directly observed by a robot. We model object permanence via a factor graph inference model, with factors representing long-term memory, short-term memory, and common sense knowledge over inter-object spatial relations. We propose the Semantic Linking Maps (SLiM) model to maintain the belief over object locations while accounting for object permanence through a CRF. Based on the belief maintained by SLiM, we present a hybrid object search strategy that enables the Fetch robot to actively search for objects on a large scale, with a higher search success rate and less search time compared to state-of-the-art search methods.

ISBN: 9798643184553Subjects--Topical Terms:

523869
Computer science.
Subjects--Index Terms:

Robot programming by demonstration

Semantic Robot Programming for Taskable Goal-Directed Manipulation.
LDR:04928nmm a2200373 4500 001 2271457
005 20201007134729.5
008 220629s2020 ||||||||||||||||| ||eng d
020 $a 9798643184553
035 $a (MiAaPQ)AAI28006610
035 $a (MiAaPQ)umichrackham003062
035 $a AAI28006610
040 $a MiAaPQ $c MiAaPQ
100 1 $a Zeng, Zhen. $3 1920990
245 1 0 $a Semantic Robot Programming for Taskable Goal-Directed Manipulation.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2020
300 $a 119 p.
500 $a Source: Dissertations Abstracts International, Volume: 81-11, Section: B.
500 $a Advisor: Jenkins, Odest Chadwicke.
502 $a Thesis (Ph.D.)--University of Michigan, 2020.
506 $a This item must not be sold to any third party vendors.
506 $a This item must not be added to any third party search indexes.
520 $a Autonomous robots have the potential to assist people to be more productive in factories, homes, hospitals, and similar environments. Unlike traditional industrial robots that are pre-programmed for particular tasks in controlled environments, modern autonomous robots should be able to perform arbitrary user-desired tasks. Thus, it is beneficial to provide pathways to enable users to program an arbitrary robot to perform an arbitrary task in an arbitrary world. Advances in robot Programming by Demonstration (PbD) has made it possible for end-users to program robot behavior for performing desired tasks through demonstrations. However, it still remains a challenge for users to program robot behavior in a generalizable, performant, scalable, and intuitive manner.In this dissertation, we address the problem of robot programming by demonstration in a declarative manner by introducing the concept of Semantic Robot Programming (SRP). In SRP, we focus on addressing the following challenges for robot PbD: 1) generalization across robots, tasks, and worlds, 2) robustness under partial observations of cluttered scenes, 3) efficiency in task performance as the workspace scales up, and 4) feasibly intuitive modalities of interaction for end-users to demonstrate tasks to robots.Through SRP, our objective is to enable an end-user to intuitively program a mobile manipulator by providing a workspace demonstration of the desired goal scene. We use a scene graph to semantically represent conditions on the current and goal states of the world. To estimate the scene graph given raw sensor observations, we bring together discriminative object detection and generative state estimation for the inference of object classes and poses. The proposed scene estimation method outperformed the state of the art in cluttered scenes. With SRP, we successfully enabled users to program a Fetch robot to set up a kitchen tray on a cluttered tabletop in 10 different start and goal settings.In order to scale up SRP from tabletop to large scale, we propose Contextual-Temporal Mapping (CT-Map) for semantic mapping of large scale scenes given streaming sensor observations. We model the semantic mapping problem via a Conditional Random Field (CRF), which accounts for spatial dependencies between objects. Over time, object poses and inter-object spatial relations can vary due to human activities. To deal with such dynamics, CT-Map maintains the belief over object classes and poses across an observed environment. We present CT-Map semantically mapping cluttered rooms with robustness to perceptual ambiguities, demonstrating higher accuracy on object detection and 6 DoF pose estimation compared to state-of-the-art neural network-based object detector and commonly adopted 3D registration methods.Towards SRP at the building scale, we explore notions of Generalized Object Permanence (GOP) for robots to search for objects efficiently. We state the GOP problem as the prediction of where an object can be located when it is not being directly observed by a robot. We model object permanence via a factor graph inference model, with factors representing long-term memory, short-term memory, and common sense knowledge over inter-object spatial relations. We propose the Semantic Linking Maps (SLiM) model to maintain the belief over object locations while accounting for object permanence through a CRF. Based on the belief maintained by SLiM, we present a hybrid object search strategy that enables the Fetch robot to actively search for objects on a large scale, with a higher search success rate and less search time compared to state-of-the-art search methods.
590 $a School code: 0127.
650 4 $a Computer science. $3 523869
650 4 $a Robotics. $3 519753
653 $a Robot programming by demonstration
653 $a Robot perception for manipulation
653 $a Semantic mapping
653 $a Scene perception
690 $a 0771
690 $a 0984
710 2 $a University of Michigan. $b Electrical and Computer Engineering. $3 3284714
773 0 $t Dissertations Abstracts International $g 81-11B.
790 $a 0127
791 $a Ph.D.
792 $a 2020
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28006610