Object Goal Navigation Based on Semantics and RGB Ego View Snehasis Banerjee12 Brojeshwar Bhowmick1and Ruddra Dev Roychoudhury1 Abstract This paper presents an architecture and method-

2025-05-02 0 0 1.99MB 4 页 10玖币

侵权投诉

Object Goal Navigation Based on Semantics and RGB Ego View

Snehasis Banerjee1,2, Brojeshwar Bhowmick1and Ruddra Dev Roychoudhury1

Abstract— This paper presents an architecture and method-

ology to empower a service robot to navigate an indoor

environment with semantic decision making, given RGB ego

view. This method leverages the knowledge of robot’s actuation

capability and that of scenes, objects and their relations –

represented in a semantic form. The robot navigates based

on GeoSem map - a relational combination of geometric and

semantic map. The goal given to the robot is to ﬁnd an object

in a unknown environment with no navigational map and only

egocentric RGB camera perception. The approach is tested both

on a simulation environment and real life indoor settings. The

presented approach was found to outperform human users in

gamiﬁed evaluations with respect to average completion time.

I. INTRODUCTION

Recent advances in AI and Robotics has led to the emer-

gence of service robots capable of complex navigation tasks,

which has two well established paradigms – (a) geometric

reconstruction and path-planning based approaches (b) end-

to-end learning-based methods. To navigate an environment,

the robot needs to sense the scene in some way. In this tasks,

camera sensor has been the primary choice for perception in

navigation tasks along with relevant distance measure sensors

such as echo, infra-red, lidar. However, among the popular

sensors, RGB camera has a lower cost and higher availability.

In fact to make service robots more affordable and main-

stream, a system using only RGB perception and relying on

software intelligence to enable navigation capability is highly

desired. This paper is an effort in that line to use only RGB

perception in a wheeled robot to enable smart navigation

using fast semantic inferences.

Humans are generally very good at the task of navigation.

If a person is asked to ﬁnd an object in a unknown scene,

his (or her) decision making will be based on visual cues in

current scene. Inspired by aforementioned human intuition,

in this work, the robot takes navigation decision based on an

amalgation of current zone’s probability of having an object,

visible objects that are closely related to the target object,

visible occlusions that may hide the object along with other

factors like free space, obstacles and risky or restricted areas.

A. Problem Description

The task is to navigate in an unmapped indoor environment

from a starting point to a speciﬁed object location and the

task is marked as complete if the object is visible to a

practical extent. The task needs to be completed based on

RGB perception of egocentric view of onboard robot camera.

The main contributions of the paper are as follows:

(i) object goal navigation algorithm using ‘GeoSem’ map and

1TCS Research, Tata Consultancy Services, Kolkata Primary

Author - 2snehasis.banerjee@tcs.com

Fig. 1: System Architecture for Semantic Navigation

semantic knowledge, with RGB ego view image as input.

(ii) a system architecture to enable semantic navigation.

(iii) successful functional evaluation of proposed methodol-

ogy in both simulated and real world indoor environments.

II. RELATED WORK

Although there has been a considerable amount of work

on the general object goal navigation problem, there has been

limited work on searching an object type just based on ego-

centric RGB camera mounted on a robot in indoor settings

using semantics. A number of works [1] uses deep neural

network for object-centric navigation policies. The closest

work on Semantic Visual Navigation [2] uses scene priors

to navigate to known as well as unknown objects in limited

settings using Graph Convolutional Networks (GCNs). The

aforementioned work has the following limitations: (a) no

concrete decision model when two or more objects are in

same frame (b) actual motion planning is not formulated

Learning framework requires signiﬁcant amount of training

to learn action policies. Our approach addresses the afore-

mentioned limitations in the sense that it tackles multiple

objects in single frame by design; path planning methods are

integrated in the decision making process; it has been tested

in limited but real world scenario; does not require extensive

training; and is based on knowledge-based decision making.

III. SYSTEM ARCHITECTURE

Fig. 1 presents the system architecture. For a wheeled

robot, the Action Space is limited to the movements in

forward direction or backward; and rotations towards left or

right with respect to ego view; and stop. The system takes in

perception from RGB camera mounted in ﬁrst person view

of robot and gives out an action control command for the

robot to move. The next RGB view is the only feedback

post actuation that the robot can get. However, as the action

arXiv:2210.11543v1 [cs.RO] 20 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：

ObjectGoalNavigationBasedonSemanticsandRGBEgoViewSnehasisBanerjee1;2,BrojeshwarBhowmick1andRuddraDevRoychoudhury1AbstractThispaperpresentsanarchitectureandmethod-ologytoempoweraservicerobottonavigateanindoorenvironmentwithsemanticdecisionmaking,givenRGBegoview.Thismethodleveragestheknowledgeofrobot...

展开>> 收起<<

Object Goal Navigation Based on Semantics and RGB Ego View Snehasis Banerjee12 Brojeshwar Bhowmick1and Ruddra Dev Roychoudhury1 Abstract This paper presents an architecture and method-.pdf

共4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！

Object Goal Navigation Based on Semantics and RGB Ego View Snehasis Banerjee12 Brojeshwar Bhowmick1and Ruddra Dev Roychoudhury1 Abstract This paper presents an architecture and method-

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: