GTC DC 2017 Isaac Presentation Final Speakernotes

27
Gavriel State | NVIDIA | November 1 st 2017 ISAAC LAB SIMULATING VIRTUAL WORLDS FOR ROBOT INTELLIGENCE

Transcript of GTC DC 2017 Isaac Presentation Final Speakernotes

Gavriel State | NVIDIA | November 1st 2017

ISAAC LABSIMULATING VIRTUAL WORLDS FOR ROBOT INTELLIGENCE

2

AI: A NEW ERA OF COMPUTING

PC INTERNET

AI & INTELLIGENT DEVICES

MOBILE-CLOUD

3

AI PROCESSOR FOR AUTONOMOUS MACHINES

XAVIER30 TOPS DL

30WCustom ARM64 CPU512 Core Volta GPU

10 TOPS DL Accelerator

GeneralPurposeArchitectures

DomainSpecificAccelerators

Energy Efficiency

CPU

CUDAGPU

DLA

Volta

+

4

ROBOTS

LEARN& PLAN

SEE, HEAR,TOUCH

ACT

5

REINFORCEMENT LEARNING

Robot Environment

Critic

ActorAction

State + Reward

6

REINFORCEMENT LEARNING - EXAMPLE

Credit: Yevgen Chebotar, Karol Hausman, Marvin Zhang, Sergey Levine

7

ISAAC ROBOT SIMULATOR

NVIDIA GPU Computer

ISAACRobot Simulator

OpenAIGYM

Robot &Environment

Definition

Virtual Jetson

LEARN& PLAN

SEE, HEAR,TOUCH

ACT

8

ISAAC HOCKEY

9

ISAAC ROBOT SIMULATOR

Train robots in virtual environments to perform tasks in the real world

Use state-of-the-art rendering (based on Unreal Engine) and realistic physics (PhysX, FleX) to create virtual learning environments

Faster, safer, and cheaper than reinforcement learning with real robots!

Leverage the compute power of NVIDIA GPUs to train with multiple episodes and/or agents simultaneously

10

PHYSICS SIMULATION

11

ISAAC SIMULATION MODEL

GAME ENGINE

Rendering

Environment

Physics

VIRTUAL ROBOT

Virtual Sensors

Intelligence

Virtual Effectors

12

RL Environment

SIMULATION + RL

ISAAC

Rendering Engine

Robot Definition

Environment

Gym Interface

ActionsPhysics Engine

Virtual Sensors

VirtualEffectors

Observations OpenAIBaselines

TensorFlow

13

LETS DO THE TIME WARPSpeed up training using multiple agents

14

ISAAC GOLF EXAMPLE

15

ISAAC DOMINOES

16

ISAAC DOMINOES IN VR

17

ISAAC DOMINOES DETAILS

Four software components:

• Vision system

• Gameplay

• Physical movement

• Human Interaction

How all the pieces fit together

AI server

Client

18

VISION – SUPERVISED LEARNINGRequires labeled data

19

VISION TRAINING - SYNTHETIC DATA

20

VISION RESULTS – REAL WORLD

21

VISION RESULTS – REAL WORLD

22

REINFORCEMENT LEARNINGContinuous Action Space

23

REINFORCEMENT LEARNINGLimited Action Space

24

HUMAN / ROBOT INTERACTION

25

NVIDIA POWERING THE AI REVOLUTION

NVIDIA GPU Cloud

ISAAC

NVIDIA GPU in Every Cloud

Xavier DLA Open Source

DGX-1 and DGX StationTesla V100

TensorRT

Tensor Core

NVIDIAGPU CLOUD

CSPs

26

Fundamentals

Autonomous Vehicles Media & Entertainment

Finance

NVIDIA DEEP LEARNING INSTITUTE

Training available as online self-paced labs and instructor-led workshops

Take self-paced labs at www.nvidia.com/dlilabs

Find or request an instructor-led workshop at www.nvidia.com/dli

Educators: download the Teaching Kit at developer.nvidia.com/teaching-kit and contact [email protected] for info on the University Ambassador Program

Machine Vision - IVA

Healthcare

…and more