Transformer-based Localization from Embodied Dialog with Large-scale Pre-training Meera Hahn

2025-05-06 0 0 2.65MB 7 页 10玖币

侵权投诉

Transformer-based Localization from Embodied Dialog with Large-scale

Pre-training

Meera Hahn∗

Google Research

meerahahn@google.com

James M. Rehg

Georgia Institute of Technology

rehg@gatech.edu

Abstract

We address the challenging task of Localiza-

tion via Embodied Dialog (LED). Given a di-

alog from two agents, an Observer navigat-

ing through an unknown environment and a

Locator who is attempting to identify the Ob-

server’s location, the goal is to predict the Ob-

server’s ﬁnal location in a map. We develop

a novel LED-Bert architecture and present an

effective pretraining strategy. We show that a

graph-based scene representation is more ef-

fective than the top-down 2D maps used in

prior works. Our approach outperforms pre-

vious baselines.

1 Introduction

A key goal in AI is to develop embodied agents that

can accurately perceive and navigate an environ-

ment as well as communicate about their surround-

ings in natural language. The recently-introduced

Where Are You? (WAY) dataset (Hahn et al.,2020)

provides a setting for developing such a multi-

modal and multi-agent paradigm. This dataset (col-

lected via AMT) contains episodes of a localization

scenario in which two agents communicate via turn-

taking natural language dialog: An Observer agent

moves through an unknown environment, while a

Locator agent attempts to identify the Observer’s

location in a map.

The Observer produces descriptions such as ‘I’m

in a living room with a gray couch and blue arm-

chairs. Behind me there is a door.’ and can respond

to instructions and questions provided by the Lo-

cator:‘If you walk straight past the seating area,

do you see a bathroom on your right?’ Via this

dialog (and without access to the Observer’s view

of the scene), the Locator attempts to identify the

Observer’s location on a map (which is not avail-

able to the Observer). This is a complex task for

which a successful localization requires accurate

∗Work done in part at Georgia Institute of Technology.

Localization Error

Predicted Location

True Location

Can you describe

where you are?

I am in a room with

an eating area and

white chairs

Observer

Locator

Figure 1: WAY Dataset Localization Scenario: The Lo-

cator has a map of the building and is trying to localize

the Observer by asking questions and giving instruc-

tions. The Observer has a ﬁrst person view and may

navigate while responding to the Locator. The turn-

taking dialog ends when the Locator predicts the Ob-

server’s position.

situational grounding and the production of rele-

vant questions and instructions.

One of the benchmark tasks supported by WAY

is ‘Localization via Embodied Dialog (LED)’. In

this task a model takes the dialog and a represen-

tation of the map as inputs, and must output a pre-

diction of the ﬁnal location of the Observer agent.

The model’s performance is based on error distance

between the predicted location of the Observer and

its true location. LED is a ﬁrst step towards devel-

oping a Locator agent. One challenge of the task

is to identify an effective map representation. The

LED baseline from (Hahn et al.,2020) uses 2D

images of top down (birds-eye view) ﬂoor maps to

represent the environment and an (x,y) location for

the Observer.

This paper provides a new solution to the LED

task with two key components. First, we propose to

model the environment using the ﬁrst person view

(FPV) panoramic navigation graph from Matter-

port (Anderson et al.,2018a), as an alternative to

top-down maps. Second, we introduce a novel vi-

siolinguistic transformer model, LED-Bert, which

scores the alignment between navigation graph

arXiv:2210.04864v1 [cs.CV] 10 Oct 2022

nodes and dialogs. LED-Bert is an adaption of

ViLBERT (Lu et al.,2019) for the LED task, and

we show that it outperforms all prior baselines. A

key challenge is the small size of the WAY dataset

(approximately 6K episodes), which makes it chal-

lenging to use transformer-based models given their

reliance on large-scale training data. We address

this challenge by developing a pretraining approach

- based on (Majumdar et al.,2020) - that yields an

effective visiolinguistic representation.

Contributions: To summarize:

We demonstrate an LED approach using navi-

gation graphs to represent the environment.

We present LED-Bert, a visiolinguistic trans-

former model which scores alignment be-

tween graph nodes and dialogs. We develop

an effective pretraining strategy that leverages

large-scale disembodied web data and similar

embodied datasets to pretrain LED-Bert.

We show that LED-Bert outperforms all base-

lines, increasing accuracy at 0m by 8.21 abso-

lute percent on the test split.

2 Related Work

BERT

Bidirectional Encoder Representations from

Transformers (BERT) is a transformer based en-

coder used for language modeling. BERT is trained

on massive amounts of unlabeled text data, and

takes as input sentences of tokenized words and

corresponding positional embeddings per tokens.

BERT is trained using the masked language model-

ing and next sentence prediction training objectives.

In the masked language modeling schema, 15% of

the input tokens are replaced with a [MASK] token.

The model is then trained to predict the true value

of the input tokens which are masked using the

other tokens as context. In the next sentence pre-

diction schema, the model is trained to predicted

if the two input sentences follow each other or

not. BERT is speciﬁcally trained on Wikipedia and

BooksCorpus (Zhu et al.,2015).

ViLBERT

ViLBERT (Lu et al.,2019) is a multi-

modal transformer that extends the BERT archi-

tecture (Devlin et al.,2018) to learn joint visio-

linguistic representations. Similar multi-modal

transformer models exist (Li et al.,2020,2019;

Su et al.,2020;Tan and Bansal,2019;Zhou et al.,

2020). ViLBERT is constructed of two transformer

encoding streams, one for visual inputs and one

for text inputs. Both of these streams use the stan-

dard BERT-BASE (Devlin et al.,2018) backbone.

The input tokens for the text stream are text tokens,

identical to BERT. The input tokens for the visual

stream are a sequence of image regions which are

generated by an object detector pretrained on Vi-

sual Genome (Krishna et al.,2017). The input to

ViLBERT is then a sequence of visual and textual

tokens which are not concatenated and only en-

ter their respective streams. The two streams then

interact using co-attention layers which are imple-

mented by swapping the key and value matrices

between the visual and textual encoder streams for

certain layers. Co-attention layers are used to at-

tend to one modality via a conditioning on the other

modality, allowing for attention over image regions

given the corresponding text input and vise versa.

Vision-and-Language Pre-training

Prior work

has experimented with utilizing dual-stream trans-

former based models that have been pretrained with

self-supervised objectives and transferring them to

downstream multi-modal tasks with large success.

This has been seen for tasks such as Visual Ques-

tion Answering (Antol et al.,2015), Commonsense

Reasoning (Zellers et al.,2019), Natural Language

Visual Reasoning (Suhr et al.,2018), Image-Text

Retrieval (Lee et al.,2018), Visual-Dialog (Mura-

hari et al.,2020) and Vision Language Navigation

(Majumdar et al.,2020). Speciﬁcally VLN-Bert

and VisDial + BERT adapt the ViLBERT architec-

ture and utilize a pretraining scheme which inspired

our approach to train LED-Bert.

3 Approach

3.1 Environment Representation

A key challenge in the LED task is that environ-

ments often have multiple rooms with numerous

similar attributes, i.e. multiple bedrooms with the

same furniture. Therefore a successful model must

be able to visually ground ﬁne-grained attributes.

Strong generalizability is also required in order to

generalize to unseen test environments. The LED

baseline in (Hahn et al.,2020) approaches localiza-

tion as a language-conditioned pixel-to-pixel pre-

diction task – producing a probability distribution

over positions in a top-down view of the environ-

ment, illustrated in Part A, in the Supplementary,

Figure 3. This choice is justiﬁed by the fact that

it mirrors the observations that the human Locator

had access to during data collection, allowing for

a straightforward comparison. However, this does

not address the question of what representation is

optimal for localization.

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：

Transformer-basedLocalizationfromEmbodiedDialogwithLarge-scalePre-trainingMeeraHahnGoogleResearchmeerahahn@google.comJamesM.RehgGeorgiaInstituteofTechnologyrehg@gatech.eduAbstractWeaddressthechallengingtaskofLocaliza-tionviaEmbodiedDialog(LED).Givenadi-alogfromtwoagents,anObservernavigat-ingthrough...

展开>> 收起<<

Transformer-based Localization from Embodied Dialog with Large-scale Pre-training Meera Hahn.pdf

共7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！

Transformer-based Localization from Embodied Dialog with Large-scale Pre-training Meera Hahn

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: