Accelerating AV Productizationwith AI

Danny Atsmon - CEO, CognataSimon Berard – CATIA Strategy Senior Manager, Dassault Systemes

Agenda

● Vision

● Challenges

○ AI Solutions in Design

○ AI Solutions in Validation

● Solution Convergence

AI-driven Design and Validation Solutions

Vision

Design Challenges

3x more System

Requirements

Challenge: Mission Driven Engineering

Autonomous

SystemsTraditional Vehicles

Taxi DriverDeliveryPeople

Mission left to the User Mission engineered within the System

Costs vs Complexity – AI is making it possible

Functional Requirements

AUTONOMOUSVEHICLES

__________

Leverage Patrimony

MASS-CUSTOMIZED

CARS__________

Explore Variants

MOBILITYEXPERIENCE

__________

Business Driven Innovation

COMPONENTS__________

Structure Legacy

AI is bringing Quality

Functional Requirements

Challenge 2: Physically Exact

Multidiscipline, Multiphysics, Multiscale

Consistent System Experience ValidationSensors optimization

AI Solutions for Design

Learning from Patrimony

Context Sensitive

Automated Assembly

Parameters space

Exploration

Function Driven

Generative Design

Model Based System

Engineering

Performance

Tradeoff

Validation Challenges

AV Tech Has Had Some Unplanned Setbacks

Challenge #1: Scale

Functional Requirements

Le

vel o

f A

uto

no

my

L0:INFORM

__________

Blind Spot

Lane Warning

Park Assist

L1: ASSIST__________

Adaptive Cruise Control

L2: ASSIST__________

Adaptive Cruise Control

+Lane

Centering

L2+: ASSIST__________

Highway Chauffeur

L4: DRIVE__________

Driverless Taxi

Challenge #1: Scale - Need Vs. Actual

The Need: 11B*

*Rand corporation

Actually driven: 16M (0.15%)

Challenge #2: Realism

REAL LIFE SIMULATIONVS.

Realism

100%

We need a realistic simulationfor a meaningful coverage

Scalable Realism

??%Scalable

Realism - Uncanny Valley (Masahiro Mori, 1978)

16

Completely machine like

EM

OT

ION

AL

RE

SP

ON

SE

+

-

Fam

iliar

ity

50% 100%

UNCANNYVALLEY

Human likeness

Moving

StillFully human

Industrial robotStuffed animal

Bunarkupuppet

Polar Express

Humanoid robot

Zombie Prosthetic handSimulation today

Realism - Uncanny Valley

17

Completely machine like

EM

OT

ION

AL

RE

SP

ON

SE

+

-

Fam

iliar

ity

50% 100%

UNCANNYVALLEY

Human likeness

Moving

StillFully human

Industrial robotStuffed animal

Bunarkupuppet

Polar Express

Humanoid robot

Zombie Prosthetic hand

Simulation needs tobe here

The animation solution

18

- Walt Disney

“An animator cannot capture all

of reality . Instead he picks 3 or

4 distinct elements and

exaggerate them.“

AI based methods for Realism

Nvidia & MIT - Video (Labels) to Video Synthesis – Wang et. al. 2018

END TO END

Procedural Modeling of a Building from a Single Image – Nishida et al, 2018

LAYERED APPROACH LAYERED APPROACH

Learning from Synthetic Humans – Varol et. al 2017

Realistic, not consistent Consistent, Not scalable (Manual)

Consistent, Not scalable (Variations)

Challenge #2: Realism conclusions

REAL LIFE SIMULATIONVS.

• DNN transfer functions ~= Realism

• Isolated layers brings better results than End to End

• Data sources should be wide (crowd sourced)

Cognata - 4 Technology layers

21

STATIC DYNAMIC SENSING CLOUD & ANALYTICS

A Combined Solution

Design and Validation Converge

Better Together

● Tests and designs together

● Validation base directly from

requirements

● Smart coverage

Takeaways

● Autonomous vehicles brings○ New designs and use cases

○ Large scale validation challenge

● AI is a key to get Autonomous vehicles in a safe and cost effective way

● A platform solution to manage, design and validate is the needed solution

Thank You!

End to end limitations

26

The Problem: Not consistent, overfeat

Buildings reconstruction conclusions

27

The Good: Consistent, Procedural

The Bad: Not practical

Learning pose - Conclusions

28

This the most advanced way of learning moving objects.

The Good: Consistent

The Bad: Not scalable