Dr.Yong

Chen吉利汽車

GEELY

AUTO馭模有道 智勝未來G

u

i

d

e t

h

e m

o

d

e

l wi

t

h w

i

s

d

o

m s

u

r

p

a

s

s

i

n

g t

h

e f

u

t

u

r

e大 模F

o

u

n

d

a

t

i

o

n型 助 力 智 能 化 變 革mod

e

l f

a

c

i

l

i

t

a

t

e t

h

e i

n

t

e

l

l

i

g

e

n

t t

r

a

n

s

f

o

r

m

a

t

i

o

n基于用戶體驗(yàn)驅(qū)動(dòng)技術(shù)價(jià)值創(chuàng)造，使智能化設(shè)計(jì)回歸理性Createtechnologicalvaluedrivenbyuserexperience,bringingintelligentdesignbackto

rationality.由粗獷的硬件驅(qū)動(dòng)體驗(yàn)轉(zhuǎn)向數(shù)據(jù)算法驅(qū)動(dòng)體驗(yàn)Shiftingfromarawhardware-drivenexperiencetoadataalgorithm-driven

experience.技術(shù)驅(qū)動(dòng)創(chuàng)新Technology-driven

innovation系統(tǒng)集成化System

integrationAI算法迭代AIalgorithmiteration數(shù)據(jù)閉環(huán)DataClosure

Loop大模型Foundation

model……1R1V3R1V5R5V5R9V5R10V5R11V+1L/3L5R12V+2L5R13V+3L100-200T200-500T 500-800T1000T+1000+TOPSputational

30TPowerHighBeginnerMiddleBEV去硬件Hardware

reduction輕地圖Reduce

high-precisionmapsE2EODD

increaseMore…大模型發(fā)展核心四要素:3+1The

four

core

elements

of

the

development

of

foundation

models:

3+1涌 現(xiàn) 式 + 繼承 式人 工 智Artificia

l能 時(shí) 代 ：Intelligenc

eEr

a

： Emerge

n

t + Inherite

d大數(shù)據(jù)平臺(tái)BigData

Platform星睿智算中心XingruiIntelligenceputingCenterLLM/Multimodal

ModelAI-DRIVE智能駕駛大模型AI-DRIVEIntelligentDrivingFoundation

Model汽車行業(yè)的知識(shí)積累KnowledgeAccumulationintheAutomotive

Industry數(shù)據(jù)Data算 力putational

Power算 法Algorithm先驗(yàn)知

識(shí)Previous

Knowledge大模型助力智能化變革Foundationmodelsboosttheintelligent

transformation大語言模型

LLM

多模態(tài)模型

MultimodalModelGUI+MUIVUI+NUI代碼生成

Code

Generation摘要生成/知識(shí)問答

Abstractgeneration/KnowledgeQ&

A

+

LMLM

+豐富的生成內(nèi)容RichGenerative

Content全新的交互體驗(yàn)pletelynewinteractive

experience先進(jìn)的生產(chǎn)力工具Advancedproductivity

tools新的開發(fā)范式Newdevelopment

paradigm百模大

戰(zhàn)BattleofaHundred

Molds產(chǎn)品某省市場和價(jià)值定位Theproductneedstohaveamarketandvalue

positioning新技術(shù)是來解決問題的或創(chuàng)造價(jià)值增量的Newtechnologiesaredevelopedtosolveproblemsorto

createincremental

value.用戶場景決定技術(shù)價(jià)值Userscenariosdeterminethevalueof

technologyWe a

l

l need

f

o

u

n

d

a

t

i

o

nm

o

d

e

l

sD

o w

e a

l

l ne

ed foundati

o

n models

?智能駕駛核心要素KeyElementsofIntelligent

Driving在復(fù)雜道路和擁堵的交通流條件下，接管率高Hightakeoverrateinplexroadandcongested

trafficconditions.智能駕駛體驗(yàn)未實(shí)現(xiàn)全駕駛場景覆蓋，體驗(yàn)不連貫Ipletecoverageofdrivingscenariosinintelligentdrivingexperience,leadingtoinconsistent

experiences.大量冗余傳感器及技術(shù)，系統(tǒng)成本居高不下Highsystemcostsduetoredundantsensorsand

technologies.大規(guī)模的數(shù)據(jù)采集標(biāo)注、軟硬件設(shè)計(jì)開發(fā)Large-scaledatacollection,annotation,software,and

hardwaredesignand

development.安全：安全≠安全感Safety:Safety≠Senseof

Safety體驗(yàn)：有沒有≠好不好Experience:Presence≠

Quality成本：去冗余≠去體驗(yàn)Cost:Redundancy≠

Experience智能駕駛長尾效應(yīng)帶來的安全困境SafetyDilemmaCausedbytheLongTailEffect

ofIntelligent

Driving.感知大多數(shù)還停留在標(biāo)注階段，缺少認(rèn)識(shí)能力Mostperceptionremainsintheannotationstage,

lackingcognitive

ability.數(shù) 據(jù) 驅(qū)D

a

t

a

-

d

r

i

v

e

n動(dòng) 模 型 迭 代 和 體m

o

d

e

l i

t

e

r

a

t

i

o

n and e

x

p

e

r

i

e

n

c

e驗(yàn) 升 級(jí)e

n

h

a

n

c

e

m

e

n

t智能駕駛大模型應(yīng)用Applicationsofthefoundationmodelonautonomous

driving解

決

關(guān)

鍵

核

心

問

題

并

創(chuàng)

造

價(jià)

值A(chǔ)ddressingKeyCoreIssuesandCreating

Value數(shù)據(jù)量不足Insufficientdata

volume數(shù)據(jù)采集、標(biāo)注成本高Highcostsofdatacollectionand

annotation局部優(yōu)化、無認(rèn)識(shí)博弈Localoptimization,nocognitive

game數(shù)據(jù)合成技術(shù)Datasynthesis

technologyAIGC風(fēng)格遷移AIGCstyle

transfer虛擬資產(chǎn)開發(fā)Virtualasset

developmentFree

Space數(shù)據(jù)標(biāo)簽Data

labeling語義分割

Semanticsegmentation同類聚合

Similarityaggregation視頻理解Video

understanding自動(dòng)標(biāo)注

AutomaticannotationE2E大模型技術(shù)E2Efoundationmodel

technology多感知融合技術(shù)（車內(nèi)外）Multi-perceptionfusiontechnology(insideandoutsidethe

vehicle)解決安全性和可解釋性等問Addressingissuessuchassafetyand

interpretability智能駕駛大模型應(yīng)用Applicationsofthefoundationmodelonautonomous

driving感知行車場景

Perception

of

driving

scenarios泊車·場景

Parking

scenes智能駕駛CornerCase

Intelligent

driving

CornerCase大模型賦能數(shù)據(jù)合成技術(shù)：兩手都要抓

[質(zhì)+量]EmpoweringDataSynthesisTechnologywithFoundationModel:GraspingBothQualityand

Quantity智能駕駛大模型應(yīng)用Applicationsofthefoundationmodelonautonomous

driving數(shù)字孿生Digital

Twin場景重建SceneReconstruction合作伙伴Partnerships平臺(tái)功能Platform

FunctionalityLLM/GAIL/SUMO第三方標(biāo)準(zhǔn)庫Third-partystandard

library路采數(shù)據(jù)轉(zhuǎn)換Roaddata

conversion3D場景3D

scenes高精地圖High-precision

maps交通信號(hào)控制Trafficsignal

control環(huán)境控制Environment

controlAI交通流AItraffic

flow標(biāo)準(zhǔn)法規(guī)案例場景Standard

regulatoryscenario

cases路采數(shù)據(jù)驅(qū)動(dòng)Roadside

data-driven智駛功能設(shè)計(jì)案例Intelligentdrivingfunctiondesign

cases案例泛化Case

generalization仿真控制內(nèi)核Simulation

controlkernel仿真數(shù)據(jù)協(xié)議層Simulation

dataprotocol

layer車輛動(dòng)力學(xué)Vehicle

dynamics物理傳感器(視覺)Physicalsensors(vision)物理傳感器(GPS

、IMU、Lidar、Radar)Physical

sensors數(shù)據(jù)合成內(nèi)核Datasynthesiskernel云服務(wù)器部署Cloudserverdeployment標(biāo)注功能模塊Annotationfunctionmodule場景建設(shè)Scene

Construction靜態(tài)場景動(dòng)態(tài)場景LLMSim2Real視覺圖像AIGC風(fēng)格遷移VisualimageAIGC

styletransfer引擎渲染技術(shù)Engine

renderingtechnology物理傳感器模型Physicalsensor

model標(biāo)注數(shù)據(jù)集Annotated

datasetSim2Real圖像數(shù)據(jù)Sim2Realimage

data非標(biāo)注真值數(shù)據(jù)Unlabeledgroundtruth

data激光點(diǎn)云數(shù)據(jù)Laserpointcloud

data其它傳感器數(shù)Othersensor

data標(biāo)注數(shù)據(jù)

Annotated

data傳感器數(shù)據(jù)Sensor

dataMainvehicle

control主車控制Static

scenesDynamic

scenes合成數(shù)據(jù)集Synthetic

dataset仿真內(nèi)核Simulation

kernelApplicationsofthefoundationmodelonautonomous

driving將數(shù)字孿生技術(shù)應(yīng)用在自動(dòng)駕駛研發(fā)測試上，在虛擬空間某著名企業(yè)立物理世界模型，還原真實(shí)世界道路場景、交通流，構(gòu)建元宇宙智駕仿真技術(shù)平臺(tái)，應(yīng)用車輛動(dòng)力學(xué)建模和物理級(jí)傳感器建模關(guān)鍵技術(shù)和自動(dòng)標(biāo)注功能模塊，高效合成標(biāo)注數(shù)據(jù)，實(shí)現(xiàn)自動(dòng)駕駛算法數(shù)據(jù)訓(xùn)練，讓數(shù)據(jù)驅(qū)動(dòng)更安全的自動(dòng)駕駛。Applyingdigitaltwintechnologytoautonomousdrivingresearchandtesting,establishingphysicalworldmodelsinvirtualspace,reconstructingreal-worldroadscenesandtrafficflow,buildingameta-universeintelligentdrivingsimulationtechnologyplatform,andapplyingkeytechnologiessuchasvehicledynamicsmodelingandphysical-levelsensormodelingandautomaticlabelingfunctionmodules,efficientlysynthesizingannotateddata,achievingdata-drivensaferautonomous

driving.虛擬數(shù)據(jù)Virtual

data遷移數(shù)據(jù)Sim2Real

data實(shí)驗(yàn)g相對(duì)a提升3.17%；加入遷移數(shù)據(jù)后，16個(gè)類別中的15個(gè)類別得到提升；Experimentgimprovedby3.17%paredtoa;afteraddingtheSim2Realdata,15outof16categorieswere

improved.Applicationsofthefoundationmodelonautonomous

drivingSim2Real風(fēng)格遷移效果PerformanceofSim2Realstyletransfer

technology真實(shí)數(shù)據(jù)real

data預(yù)標(biāo)注大模型技術(shù)框架Pre-labelingfoundationmodel

technologyBackbone選用圖文多模態(tài)模型，大大增強(qiáng)了模型的理解和泛化能力；The

backbone

adopts

the

multimodal

model

of

text

and

images,

which

greatly

enhances

the

understanding

and

generalization

ability;結(jié)合多方數(shù)據(jù)源，及數(shù)據(jù)強(qiáng)化策略使模型更好地泛化業(yè)務(wù)場景;bining

multiple

data

sources

and

data

enhancement

strategies

allowsthe

model

to

fit

better

in

different

scenarios;最先進(jìn)的多尺度特征、去噪訓(xùn)練等策略的引入使得Transformer架構(gòu)性能優(yōu)越;The

introduction

of

multi-scale

features,

denoising

training,

and

other

strategies

makes

the

Transformer

superior

in

performance;可同時(shí)處理語義分割、物體檢測等2D圖像感知任務(wù)，可實(shí)現(xiàn)標(biāo)注數(shù)據(jù)互通。It

can

processing

2D

image

perception

tasks

such

as

semantic

segmentation

and

object

detection

simultaneously,

and

realize

annotation

datainteroperability.智能駕駛大模型應(yīng)用Applicationsofthefoundationmodelonautonomous

driving國際數(shù)據(jù)集Internationaldata

set吉利某車型productionreturn

data以上是在模型在國際數(shù)據(jù)集Cityscapes、ACDC以及量產(chǎn)回傳數(shù)據(jù)上的推理效果圖；TheaboveshowstheinferenceresultsoninternationaldatasetssuchasCityscapes,ACDC,andproductionreturn

data.Applicationsofthefoundationmodelonautonomous

driving預(yù)標(biāo)注大模型效果PerformanceofPre-labelingfoundation

model基于AI大模型與虛擬數(shù)據(jù)合成技術(shù)，吉利汽車在國際知名數(shù)據(jù)集CityScapes（語義風(fēng)格的標(biāo)桿數(shù)據(jù)集）及ACDC數(shù)據(jù)集（極端天氣場景數(shù)據(jù)集）上，取得了實(shí)時(shí)榜單全球第一的成績。BasedonAIfoundationmodelandvirtualdatasynthesistechnology,GeelyAutomobileink.rankedthefirstplace

inthereal-timelistontheinternationallyrenowneddatasetCityScapesand

ACDC.CityScapes：語義風(fēng)格的標(biāo)桿數(shù)據(jù)集CityScapes：Thebenchmarkdatasetforsemantic

segmentationACDC：極端天氣場景數(shù)據(jù)集ACDC：AdverseConditionsDataset

withCorrespondences智能駕駛大模型應(yīng)用Applicationsofthefoundationmodelonautonomous

driving大模型賦能數(shù)據(jù)合成技術(shù)Foundationmodelempowersdatasynthesis

technology銀河E8大模型音樂律動(dòng)GEELYGalaxyE8FoundationModel-Musical

Rhythm自 然 語 言 對(duì) 話 + 更 多 …Natur

a

l langua

g

e dialog

u

e + more..

.智能座艙大模型應(yīng)用Foundationmodelsappliedtointelligent

cockpit音

樂

不

僅

可

以

被

聽

見

，

還

可

以

被

看

見M

u

s

i

c c

a

n not on

l

y be h

e

a

r

d but al

s

o s

e

e

n大模型研發(fā)底座：吉利星睿智算中心FoundationmodelR&Dbase：GeelyXingruilntelligentputing

Center基于英偉達(dá)算力工具，

構(gòu)建全球車企首個(gè)“

云

-

數(shù)

-

智”

一體化超級(jí)云計(jì)算平臺(tái)Buildingtheworld'sfirst"clouddigitalintelligence"integratedsupercloudputingplatformforautomotivepaniesbasedonNVIDIA'sputingpowerand

toolchain八大創(chuàng)新 締造極致智算基座Eightmajorinnovationscreatetheultimateintelligentputing

base大模型研發(fā)底座：吉利星睿智算中心FoundationmodelR&Dbase：GeelyXingruilntelligentputing

Center基于英偉達(dá)算力工具，

構(gòu)建全球車企首個(gè)“

云

-

數(shù)

-

智”

一體化超級(jí)云計(jì)算平臺(tái)Buildingtheworld'sfirst"clouddigital