Large-Scale Self-supervised Robot Learning with GPU-enabled - PowerPoint PPT Presentation

Large-Scale Self-supervised Robot Learning with GPU-enabled Video-Prediction Models Frederik Ebert, Chelsea Finn, Alex Lee, Sergey Levine NVIDIA GTC 2018 1

Typical Bar in 20?? 1969 Stanford Arm 2015 DARPA Robotics Challenge

Humans have excellent mental models of physical objects 3

How can robots acquire general models and skills using large amounts of autonomously collected data? 4

Related work on self-supervised learning Levine et al. 2016 Pinto & Gupta, 2015 Predict raw sensory inputs instead of binary events . Gandhi et al. 2017 5

Related work on video-prediction Visual Model-Predictive Control Oh et al. 2015 Mathieu et al. 2016 Finn & Levine 2017 Byravan et al. 2017 6

Visual Model-Predictive Control User Input Planning Module Cost Function apply action to Robot Video Prediction Model Designated Pixel Goal Point [Finn et al. 2017] 7

Random Data Collection Collected 45,000 trajectories, recording camera images and actions 8

Action-Conditioned Video Prediction Real Action 2 Action 1 Action 0 state state Recurrent NN Recurrent NN Recurrent NN Generated 9

Skip Connection Neural Advection (SNA) DNA (Finn et al.) 10

Action-Conditioned Video Prediction Temporal Skip Connections Real Action 2 Action 1 Action 0 state state Recurrent NN Recurrent NN Recurrent NN Generated 11

Skip Connection Neural Advection (SNA) DNA (Finn et al.) SNA (Ours) 12

Skip Connection Neural Advection (SNA) 11 compositing masks Image of current time step 64x64x16 64 skip 32x32x32 32x32x32 16 16 skip 16x16x16 32 16x16x64 32 64 8x8x64 33 32 64 gen. Image of next time step conv conv conv conv conv conv conv de conv de conv de conv de conv de conv de conv 3×3 5x5 and channe l 3×3 3×3 3×3 3×3 1×1 3×3 3×3 3×3 3×3 3×3 3×3 softmax stride 2 stride 2 stride 2 stride 2 stride 2 stride 2 10 CDNA ke rne ls 9 maske d 9 compositing conv 9×9 Convolutional LSTM Conv-LSTM Masks CDNA Kernels Image from first time step, temporal skip connection 13

Prediction of Pixel Positions (Test Time) Action 2 Action 0 Action 1 state state Recurrent NN Recurrent NN Recurrent NN Generated 14

Effects of using temporal skip connections SNA (Ours) DNA (Finn et al.) Designated Pixel 15

Planning with Visual-MPC Designated Pixel Goal Pixel 16

Planning: Expected Distance to Goal Cost Predicted Designated Pixel Distance to Goal Distribution for Goal Point designated Pixel 17

Action Selection using Cross-Entropy Method Designated Pixel Goal Pixel Iteration 1 Iteration 2 Iteration 3 18

Results 19

Generalization to objects not seen during training 20

Collision Avoidance Task, involving Occlusion Designated Pixel Goal Pixel Static Pixel 21

Finn et al.

Ours 23

Multi-Goal Pushing Benchmark 24

Takeaways • Temporal skip connections significantly improve the ability to deal with occlusions . • Video-prediction models can be reused across many tasks . • Self-supervised learning on large scale data enables generalizable skills . 31

Q&A Chelsea Finn Alex X. Lee Sergey Levine Code and Data: https://sites.google.com/view/sna-visual-mpc

Large-Scale Self-supervised Robot Learning with GPU-enabled - PowerPoint PPT Presentation

Large-Scale Self-supervised Robot Learning with GPU-enabled Video-Prediction Models Frederik Ebert, Chelsea Finn, Alex Lee, Sergey Levine NVIDIA GTC 2018 1 Typical Bar in 20?? 1969 Stanford Arm 2015 DARPA Robotics Challenge Humans have

Large-Scale Self-Supervised Robotic Learning Chelsea Finn In collaboration with Sergey Levine

Augmenting Supervised Neural Networks with Unsupervised Objectives for Large-scale Image

Sibyl: A system for large scale supervised machine learning Kevin Canini, Tushar Chandra, Eugene

Parallelizing Semi- ReDAS Lab Supervised Learning Algorithms with MapReduce Nick Gauthier

Large-Scale Machine Learning at Twitter 2 Large-Scale Machine Learning at Twitter Jimmy Lin and

TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension Mandar

Semi-geostrophic equations A large-scale model in meteorology Stefania Lisai Supervised by B.

Ethics in Techniques for large-scale data Graham J.L. Kemp TECHNIQUES FOR LARGE-SCALE DATA

Learning frameworks Self-supervised learning: (Auto)encoder networks Supervised learning Network

Simulating Energy Aware Networks in Large Scale Distributed Systems Betsegaw Lemma Amersho

Weak Supervised Learning on Ray using Snorkel British Sentiment Analysis with LSTM Using Noisy

Testing Large Data Sets with Supervised Machine Learning

Learning with Large Datasets L eon Bottou NEC Laboratories America Why Large-scale Datasets?

FINANCING LARGE SCALE SOLAR Large Scale Solar Conference - Sydney Gloria Chan Director, Large

Self-Supervised Feature Learning by Learning to Spot Artifacts Wonbin Kim Self-Supervised

DSSTNE: Deep Learning At Scale For Large Sparse Datasets

Comparison of Normalization Methods for cDNA Microarrays Liling Warren, Ben Hui Liu

molecular mechanisms in Heavy Chain Deposition Disease Presented by Sbastien BENDER Disclosure

COVID-19 Employer Guidance: Workplace Safety-OSHA, CDC and Virus Testing Gregory R. Begg Partner

Care Among Hard-to-Reach Populations : The Northern New Jersey and New York City Hepatitis B

Changes Are Afoot: The Prosecution and Litigation Landscape Post KSR and the New Continuation

UKF project presentation: Elucidation of the physiological roles of human dipeptidyl peptidase

Introduction of Research Center for Allergy and Immunology (RCAI) -What is Immunogenomics

Thursday, January 25, 2018 10:00 a.m. Chief Executive Officer, CEA More Californians need