Cloud Native Robotic Applications with GPU Sharing on Kubernetes Giovanni Toffetti1 Leonardo Militano1 Sean Murphy2 Remo Maurer1 and Mark Straub1 Abstract In this paper we discuss our experience in teach-

2025-04-27 0 0 267.84KB 5 页 10玖币

侵权投诉

Cloud Native Robotic Applications with GPU Sharing on Kubernetes

Giovanni Toffetti1, Leonardo Militano1, Se´

an Murphy2, Remo Maurer1, and Mark Straub1

Abstract— In this paper we discuss our experience in teach-

ing the Robotic Applications Programming course at ZHAW

combining the use of a Kubernetes (k8s) cluster and real, het-

erogeneous, robotic hardware. We discuss the main advantages

of our solutions in terms of seamless “simulation to real” expe-

rience for students and the main shortcomings we encountered

with networking and sharing GPUs to support deep learning

workloads. We describe the current and foreseen alternatives

to avoid these drawbacks in future course editions and propose

a more cloud-native approach to deploying multiple robotics

applications on a k8s cluster.

I. INTRODUCTION

The Robotic Applications Programming course (RAP) has

been taught at the School of Engineering of the Zurich Uni-

versity of Applied Sciences since Spring 2021. The course

is offered to bachelor students in IT in their last semesters

with the aim of learning how to program robotic applications

using the ROS framework, as well as combining knowledge

from other courses (e.g., computer vision, artiﬁcial intelli-

gence, cloud computing) to make robots autonomous.

The course is organized in three main sections: (1) robotics

(e.g., basic robotic HW, URDF/XACRO, rviz, poses, co-

ordinate frames and transformations, controllers) and ROS

fundamentals (communication primitives, ROS packages);

(2) base capabilities (e.g., SLAM, navigation, perception,

arm motion planning and control); and (3) distributed robotic

applications culminating with a yearly challenge. This year’s

challenge was inspired by the DARPA Subterranean Chal-

lenge1: student groups had to program a mobile manipulator

(a Robotnik Summit XL with a UR5 arm) to autonomously

explore an unknown indoor area, detect coke cans, and bring

them back to the starting point.

We designed the theoretical modules and practical labs of

the course for students to minimize the conﬁguration and

set-up effort to work collaboratively in order to focus on

software development and, at the same time, preserve the

excitement of seeing one’s own code running on real robotic

hardware. In order to do this we ﬁrst investigated available

solutions we could use for our scenario. We report on related

work in the next section.

II. RELATED WORK

As the complexity of robotic applications is growing

steadily, with the adoption of advanced solutions such as

This work was accepted at the IROS’22 Cloud Robotics Workshop:

https://sites.google.com/view/iros22-cloud-robotics

1Zurich University of Applied Sciences (ZHAW), Switzerland

[toff|milt|murm|stmr]@zhaw.ch

2ETH Z¨

urich, Swiss Data Science Center (SDSC), Switzerland

sean.murphy@sdsc.ethz.ch

1https://www.subtchallenge.com

AI for semantic mapping or grasp generation, new needs

appeared in terms of computation, networking and storage

resources. To cope with them, Fog/Cloud-Robotic solutions

are gaining traction in several domains. The possibility for

remotely controlling robotic systems further reduces costs

for deployment, monitoring, diagnostic and orchestration

of any robotic application. This, in turn, allows for build-

ing lightweight, low cost and smarter robots as the main

computation and communication burden is brought to the

cloud. Since 2010, when the ”cloud robotics” term was

ﬁrst used, several projects (e.g., RoboEarth [8] DAVinci

[9]) investigated the ﬁeld pushing forward both research

and products to appear on the market. Companies started

investing in the ﬁeld as they recognized the huge potential

of cloud robotics. This led to open-source cloud robotics

frameworks appearing in recent years. An example of this is

the solution from Rapyuta Robotics2. Similarly, commercial

solutions for developers have seen the light with the big

players in the Cloud ﬁeld joining the run (e.g., Amazon

Robomaker3) or startups (e.g., Formant.io).

A. Robotic Applications Development in Education

Depending on the educational level and the requirements

for professional knowledge of robotic application develop-

ment, different teaching and learning approaches can be

identiﬁed as a combination of simulation- and hardware-

based solutions.

Simulation-based learning leverage software tools and

programming languages to simulate the behavior of robots

without direct interaction with a physical robot. Under this

category we include web robotics as a way of learning online

using a web-based platforms for simulating robots, as e.g., in

[2]. This latter is gaining momentum with offerings such as

The Construct4or AWS RoboMaker5which are cloud-based

simulation services that enable robotics developers to run,

scale, and automate simulation without managing any infras-

tructure. Simulation based solutions are clearly useful and

serve some important educational needs; however, the models

on which they are based always have some limitations

which can become apparent in a real world context. Further,

adopting a simulation only approach does not give students

experience with some of the more practical considerations

associated with working with physical devices.

Hardware-based learning allows for direct interaction

and programming of physical robots. In some simple do-

2https://www.rapyuta-robotics.com/

3https://cloud.google.com/cloud-robotics/

4https://www.theconstructsim.com/

5https://aws.amazon.com/robomaker/

arXiv:2210.03936v2 [cs.RO] 31 Oct 2022

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：

CloudNativeRoboticApplicationswithGPUSharingonKubernetesGiovanniToffetti1,LeonardoMilitano1,Se´anMurphy2,RemoMaurer1,andMarkStraub1AbstractInthispaperwediscussourexperienceinteach-ingtheRoboticApplicationsProgrammingcourseatZHAWcombiningtheuseofaKubernetes(k8s)clusterandreal,het-erogeneous,robotich...

展开>> 收起<<

Cloud Native Robotic Applications with GPU Sharing on Kubernetes Giovanni Toffetti1 Leonardo Militano1 Sean Murphy2 Remo Maurer1 and Mark Straub1 Abstract In this paper we discuss our experience in teach-.pdf

共5页,预览1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！

Cloud Native Robotic Applications with GPU Sharing on Kubernetes Giovanni Toffetti1 Leonardo Militano1 Sean Murphy2 Remo Maurer1 and Mark Straub1 Abstract In this paper we discuss our experience in teach-

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: