CoGrasp 6-DoF Grasp Generation for Human-Robot Collaboration Abhinav K. Keshari Hanwen Ren and Ahmed H. Qureshi Abstract Robot grasping is an actively studied area in

2025-04-27 0 0 5.5MB 8 页 10玖币

侵权投诉

CoGrasp: 6-DoF Grasp Generation for Human-Robot Collaboration

Abhinav K. Keshari, Hanwen Ren, and Ahmed H. Qureshi

Abstract— Robot grasping is an actively studied area in

robotics, mainly focusing on the quality of generated grasps

for object manipulation. However, despite advancements, these

methods do not consider the human-robot collaboration settings

where robots and humans will have to grasp the same objects

concurrently. Therefore, generating robot grasps compatible

with human preferences of simultaneously holding an object

becomes necessary to ensure a safe and natural collaboration

experience. In this paper, we propose a novel, deep neural

network-based method called CoGrasp that generates human-

aware robot grasps by contextualizing human preference mod-

els of object grasping into the robot grasp selection pro-

cess. We validate our approach against existing state-of-the-

art robot grasping methods through simulated and real-robot

experiments and user studies. In real robot experiments, our

method achieves about 88% success rate in producing stable

grasps that also allow humans to interact and grasp objects

simultaneously in a socially compliant manner. Furthermore,

our user study with 10 independent participants indicated our

approach enables a safe, natural, and socially-aware human-

robot objects’ co-grasping experience compared to a standard

robot grasping technique.

I. INTRODUCTION

Co-grasping is an essential part of human-robot collabo-

ration tasks where a human and robot simultaneously grasp

an object during manipulation. The need for collaborative

robot systems has become evident from the lack of available

skilled workforce in hospitals, factory ﬂoors, and at home to

assist people in their daily lives [1]. For instance, at hospitals,

robots with co-grasping skills can assist in passing various

equipment to and from surgeons during surgery or passing

medicines to maintain a safe distance between healthcare

workers and patients concerning contagious diseases like

CoVID-19. Likewise, at factory ﬂoors, the tasks for assistive

robots could include fetching and handing over various

tools to and from their human collaborator in the loop or

performing complex assembly tasks through human-machine

teaming, which can signiﬁcantly improve the overall work

efﬁciency and throughput. Similarly, at home, our elderly

often struggle to fetch various objects. Therefore, robots with

co-grasping skills can assist them by bringing and handing

over different daily-life things, such as utensils, keys, tv

remotes, etc.

Although several methods for robot grasp generation exist

[2]–[10], ranging from geometric to data-driven strategies,

they are human-agnostic and provide robot-centric algo-

rithms, i.e., observing an object and selecting a gripper’s pose

to pick an object without considering collaborating human

partners. Generally, in human-robot collaboration, we would

A.K. Keshari, H. Ren, and A.H. Qureshi are with Purdue University,

{akeshari, ren221, ahqureshi}@purdue.edu

(a) Human-aware Grasping (b) Human-unaware grasping

Fig. 1. CoGrasp generates human-aware robot grasps (a) compared to

traditional methods (b) that do not consider humans in the loop.

expect our robot to grasp things that are also comfortably

graspable by their interacting partners during cooperation.

For instance, consider a scenario in Fig. 1. Both grasps

in Fig. 1a and Fig. 1b would be considered valid for the

robot by the existing approaches. However, the grasps in

Fig. 1b are inherently invalid as they point sharp ends toward

humans, which would be considered unsafe for collaboration.

Similarly, in other situations with no sharp objects, the robot

would be expected to leave sufﬁcient space for humans to

co-grasp objects simultaneously.

In this paper, we propose a human-aware robot grasp

generation pipeline called CoGrasp that considers both robot

grasp quality and human-in-the-loop for safe and com-

pliant collaboration. Our approach accomplishes human-

aware grasping from a raw partial 3D object point cloud

by optimizing robot grasp generation using a deep neural

network-based object shape completion network, a socially-

compliant human grasp prediction network, and a pruning

network. Our pruning network builds on our novel co-grasp

evaluation algorithm to select stable robot grasps compat-

ible with predicted human grasps for a given object. The

overview of our pipeline is shown in Fig. 2. Our approach

demonstrates producing grasps appropriate for robot-human

collaborative object manipulation, which also works in real-

world experiments (Fig. 1). The main contributions of our

paper are summarized as follows:

•A novel and, to the best of our knowledge, the ﬁrst

end-to-end human-aware 6-DoF robot grasp generation

method that works in both simulation and real-world

environments.

arXiv:2210.03173v1 [cs.RO] 6 Oct 2022

•An algorithm that computes grasp quality scores us-

ing the geometric information (approach direction and

spatial representation) from interactions between the

objects, human hand, and robot gripper.

•A neural model for fast and parallel evaluation of

various robot grasp candidates for human friendliness

and stability.

•A new set of metrics that evaluate the quality of grasps

based on their safety, human friendliness, and efﬁciency

for human-robot collaboration tasks.

•A validation of our CoGrasp approach through real-

robot experiments, demonstrating our method achieves

a 88% success rate in generating stable robot grasps

while leaving socially compliant space on the object

for humans to co-grasp concurrently.

•A validation of our approach through a user study with

10 participants, indicating our method achieves 22%

higher scores on various metrics of CoGrasp’s social

compliance and safety than a traditional robot-centric

method [6].

II. RELATED WORK

This section discusses various techniques that generate

collision-free, stable robot grasps for object manipulation.

We divide these methods into three categories, i.e., classical,

data-driven, and contextual, as described in the following.

A. Classical Method

The study of robot grasp generation goes very back,

starting from attempting to handle objects using a robot

hand with elastic ﬁngers [11]. It gave rise to geometric-based

approaches [12]–[14] for producing grasps using the contact

points’ classiﬁcation as frictionless, friction, or soft contact

to identify parts on the object for a successful grasping.

Another line of work, like [15], studied the number of contact

points needed for stable grasping. The contact points are

essential for grasp stability, but the current techniques only

look into these for identifying suitable regions from the

robot’s perspective. In a similar vein, [16], [17] demonstrated

that grasping an object results in a pull force and overcoming

that wrench is an essential aspect of stability. [18]–[20]

includes studying complex kinematics of the object and the

hand motion involved during an interaction, displaying the

movement of an approaching hand or the gripper to be

critical for grasping. Following the formulations of stable

grasping, geometry-based techniques [5], [21]–[24] were

proposed that rely directly on the object shape to generate a

suitable grasp. However, such methods do not generalize to

real-world scenarios where object models are often unknown.

Modern methods [3] tend to account for surface normals

to evaluate the quality of grasps and use them to compute

a safe distance for a stable grasp. [25] models the mean

axis of an object by running PCA and empirically choosing

a safe space from a normal plane to that axis. Despite

progress in stable robot grasping, these geometric approaches

do not consider human-in-the-loop and solely rely on having

one manipulator; thus do not apply to collaboration tasks

requiring the co-grasping of the objects.

B. Data-Driven Method

With the advancement in computational resources and

deep neural models, data-driven methods have emerged

signiﬁcantly for generating grasps. In emerging scenarios

where input is only available from visual sensors, the ex-

isting methods lean on computing the object models and

their 6DoF poses [26] before deploying traditional grasping

methods. Furthermore, when the complete 3D object models

are unavailable, learning-based shape reconstruction [27]

from inference or multiple views [28] are proposed to ﬁll

in the gap. Many reinforcement learning techniques have

also come up, which learn gripper poses using exploration

[29] and learn policies for manipulation [30]. They help

identify dynamic responses to disturbances while grasping.

Still, such perturbations are only based on gripper-object and

environment interaction. Nevertheless, 3D reconstruction and

exploring the large 6-DoF state space suffer from storing a

huge amount of data that consists exclusively of the gripper

and object relative poses.

Since contact points are crucial in grasping, methods

like PointNet++ [31] are used to learn patterns from point

clouds. For instance, [6], [32], [33] utilizes PointNet++ to

learn geometric forms between grippers and objects from

the contact points data available in grasping datasets like

ACRONYM [34]. To extend the point space that is not

limited to the contact points, learning-based methods like

Dexnet [35]–[37] directly learn the orientations and the

approach direction of the gripper. There are also methods

[38], [39] that tend to produce a grasp score for each point in

the space when contact information is not present. However,

the contact information and orientations that are looked upon

come only from the gripper and the object. Some other neural

sampling-based methods [40], [41] use different grasp quality

metrics as objective functions. These metrics depending on

the gripper orientations and surface areas, only relate to the

overall stability rather than human awareness.

C. Contextual Grasping

Contextual grasping refers to grasp generation with some

context about the objects and their underlying tasks. This

problem often involves encoding contextual information like

the semantic representation or the object properties into

the network inputs. For example, [9] encodes the target

candidate’s visual, tactile, and texture information while

performing tasks like picking, lifting, or pouring. [10] also

considers encoding the relationships of objects in the scene,

which allows reasoning about invisible points, enabling

collision-free grasp. However, the enhanced reliabilities of

grips come from the extensive cost when acquiring hand-

labeled training data. In addition, these works focus more on

producing human-like grasps or moving the object of interest

to the target position, which involves no human actions.

In summary, none of the abovementioned methods con-

siders a simultaneous human grasp while producing a stable

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CoGrasp:6-DoFGraspGenerationforHuman-RobotCollaborationAbhinavK.Keshari,HanwenRen,andAhmedH.QureshiAbstractRobotgraspingisanactivelystudiedareainrobotics,mainlyfocusingonthequalityofgeneratedgraspsforobjectmanipulation.However,despiteadvancements,thesemethodsdonotconsiderthehuman-robotcollaboration...

展开>> 收起<<

CoGrasp 6-DoF Grasp Generation for Human-Robot Collaboration Abhinav K. Keshari Hanwen Ren and Ahmed H. Qureshi Abstract Robot grasping is an actively studied area in.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

CoGrasp 6-DoF Grasp Generation for Human-Robot Collaboration Abhinav K. Keshari Hanwen Ren and Ahmed H. Qureshi Abstract Robot grasping is an actively studied area in

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: