THE ROLE OF AI IN A VR WORLD

Transcription

1 In partnership with THE ROLE OF AI IN A VR WORLD Nirmal Mehta - Bayesian by Birth Drew Farris Grudgingly Bayesian Cameron Kruse Bayesian by Default OCTOBER 2018

2 INTRODUCTIONS Nirmal Mehta: ChiefTechnologist. He has 10 years ofdistributedapplicationarchitectures and emerging technology research, prototype development and implementation. He leads the firm s efforts in Immersive Machine Intelligence and emerging technology strategy. He focuses on bringing leading edge technologies to enterprise systems for commercial and public-sector clients. Cameron Kruse: Lead Technologist. He works within Booz Allen s Strategic Innovation Group working on projects at the intersection of AI and immersive technology. He likes working in this space as he sees technology as something that should help humansbetter explore and connect withthe world around us. Drew Farris: Chief Technologist. He is one of Booz Allen s machine learning subject matter experts. He mostly focused on Information Retrieval and Natural Language Processing, he has a latent love of Virtual Reality, with an undergrad degree in Computer Graphics. 1

3 IMMERSIVE TECHNOLOGY NEEDS AI AI NEEDS IMMERSIVE TECHNOLOGY ALL EMERGING TECHNOLOGY GAIN PREVALENCE AS AN ECOSYSTEM OF CONNECTED TECHNOLOGIES. TO CHART A PATH INTO THE FUTURE, IT IS BETTER TO LOOK AT THE CONNECTIONS BETWEEN TECHNOLOGIES. We see the key convergences creating the connections that will shape the future to be Cloud Computing, AI, and Immersive Technology. Both AI and Immersive are diverse sets of technologies. Today we will put most of our focus on the intersection of AIand VR. The convergence of AI and AR is also very interesting, but for the sake of clarity we ll leave itfor a future conversation. AI Cloud Immersive Immersive AI 3d Model Video Robots 2

4 SMART AGENTS: REINFORCEMENT LEARNING LEARNING STRATEGIES WITH REWARD FUNCTIONS David Busch Booz Allen Hamilton 3

5 SMART AGENTS: DATA-DRIVEN ANIMATION LEARNING COMPLEX KINEMATIC SYSTEMS CrowdAI, NIPS 2017 Learning to Run Challenge 4

6 SMART AGENTS: MULTI-AGENT LEARNING AGENTS DEVELOP STRATEGIES WHEN TRAINING AGAINST HUMANS IN MULTI-PLAYER GAMES OpenAI Five: DotaGameplay 5

7 POSE ESTIMATION FOR INTERACTIVITY 2-D TO 3-D POINT DETECTION OpenPose 6

8 POSE ESTIMATION FOR INTERACTIVITY MOTION TRANSFER Chan, et al. Everybody Dance Now 7

9 HUMAN MESH RECOVERY Kanazawa, et al Peng, et al. 8

10 POSE ESTIMATION TECHNIQUES Peng, et al. 9

11 LEARNING MOTION FROM VIDEO COMBINING POSE ESTIMATION TECHNIQUES AND DATA TO DEVELOP ANIMATION BEHAVIORS Peng, et. al. Reinforcement Learning of Physical Skills from Videos 10

12 USING GENERATIVE ADVERSARIAL NETWORKS TO CREATE 3D CONTENT CONTENT CREATION IS THE MOST EXPENSIVE AND TIME CONSUMING PART OF DEVELOPING IMMERSIVE EXPERIENCES. USING AI TO CREATE 3D CONTENT COULD CUT COSTS AND MAKE ARTISTS FASTER AT THEIR JOBS. Wu, Jiajun, et al. - Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling 11

13 USING GENERATIVE ADVERSARIAL NETWORKS TO CREATE 3D CONTENT CONTENT CREATION IS THE MOST EXPENSIVE AND TIME CONSUMING PART OF DEVELOPING IMMERSIVE EXPERIENCES. USING AI TO CREATE 3D CONTENT COULD CUT COSTS AND MAKE ARTISTS FASTER AT THEIR JOBS. Wu, Jiajun, et al. - Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling 12

14 USING GENERATIVE ADVERSARIAL NETWORKS TO CREATE 3D CONTENT CONTENT CREATION IS THE MOST EXPENSIVE AND TIME CONSUMING PART OF DEVELOPING IMMERSIVE EXPERIENCES. USING AI TO CREATE 3D CONTENT COULD CUT COSTS AND MAKE ARTISTS FASTER AT THEIR JOBS. Wu, Jiajun, et al. - Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling 13

15 TRAINING AI ALGORITHMS IN 3D ENVIRONMENTS THERE ARE MANY TIMES WHEN TRAINING DATA IS TOO COSTLY OR CANNOT BE OBTAINED. WE HAVE SEEN SUCCESS TRAINING AI IN 3D ENVIRONMENTS. Andre Nguyen Booz Allen Hamilton 14

16 TRAINING AI ALGORITHMS IN 3D ENVIRONMENTS THERE ARE MANY TIMES WHEN TRAINING DATA IS TOO COSTLY OR CANNOT BE OBTAINED. WE HAVE SEEN SUCCESS TRAINING AI IN 3D ENVIRONMENTS. Andre Nguyen Booz Allen Hamilton 15

17 TRAINING ROBOTS THROUGH IMITATION USING VIRTUAL REALITY AS THE INTERFACE TRAINING ROBOTS TO PERFORM HUMAN TASKS IS USUALLY COSTLY AND HARD. INSTEAD OF TRYING TO PROGRAM A ROBOT WE COULD TRAIN ROBOTS THROUGH IMITATION. OpenAI Teaching Robots to learn 16

18 TRAINING ROBOTS THROUGH IMITATION USING VIRTUAL REALITY AS THE INTERFACE TRAINING ROBOTS TO PERFORM HUMAN TASKS IS USUALLY COSTLY AND HARD. INSTEAD OF TRYING TO PROGRAM A ROBOT WE COULD TRAIN ROBOTS THROUGH IMITATION. OpenAI Teaching Robots to learn 17

19 TRAINING ROBOTS THROUGH IMITATION USING VIRTUAL REALITY AS THE INTERFACE TRAINING ROBOTS TO PERFORM HUMAN TASKS IS USUALLY COSTLY AND HARD. INSTEAD OF TRYING TO PROGRAM A ROBOT WE COULD TRAIN ROBOTS THROUGH IMITATION. Vision Network OpenAI Teaching Robots to learn 18

20 TRAINING ROBOTS THROUGH IMITATION USING VIRTUAL REALITY AS THE INTERFACE TRAINING ROBOTS TO PERFORM HUMAN TASKS IS USUALLY COSTLY AND HARD. INSTEAD OF TRYING TO PROGRAM A ROBOT WE COULD TRAIN ROBOTS THROUGH IMITATION. Imitation Network Vision Network OpenAI Teaching Robots to learn 19

21 TRAINING ROBOTS THROUGH IMITATION USING VIRTUAL REALITY AS THE INTERFACE TRAINING ROBOTS TO PERFORM HUMAN TASKS IS USUALLY COSTLY AND HARD. INSTEAD OF TRYING TO PROGRAM A ROBOT WE COULD TRAIN ROBOTS THROUGH IMITATION. OpenAI Teaching Robots to learn 20

22 In partnership with PANEL DISCUSSION Nirmal Mehta - Bayesian by Birth Drew Farris Grudgingly Bayesian Cameron Kruse Bayesian by Default OCTOBER 2018

23 CONTINUE THE CONVERSATION IF YOU HAVE A USE CASE OR ARE INTERESTED IN COLLABORATING COME TALK TO US! Find out more about BAH immersive studio at immersive.bah.com Game Engines: Generative Adversarial Networks: Pose Estimation, Human Mesh Recovery, Everybody Dance Open AI Five (DOTA): Open AI (Teaching Robots to learn): 22