【DL輪読会】Implicit Behavioral Cloning
871 Views
May 13, 22
スライド概要
2022/05/13
Deep Learning JP:
http://deeplearning.jp/seminar-2/
DL輪読会資料
関連スライド
各ページのテキスト
DEEP LEARNING JP [DL Papers] Implicit Behavioral Cloning (CoRL 2021) Koki Yamane, University of Tsukuba http://deeplearning.jp/ 1
◼ Implicit Behavioral Cloning ◼ CoRL 2021 ◼ Robotics at Google Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson ◼ https://implicitbc.github.io/ 2022/5/13 2
(BC) 2022/5/13 EBM 3
Behavior Cloning ෝ = 𝑭𝜽 𝒐 𝒂 ◼ ◼ ◼ 2022/5/13 4
BC Explicit Model ( ෝ = 𝑭𝜽 𝒐 𝒂 2022/5/13 ) Implicit Model ( ) ෝ = argmin 𝑬𝜽 𝒐, 𝒂 𝒂 5
ෝ = argmin 𝑬𝜽 𝒐, 𝒂 𝒂 𝑁 𝐿InfoNCE = −log 𝑝෦𝜃 𝑦𝑖 |𝑥, 𝑖=0 𝑝෦𝜃 𝑦𝑖 |𝑥, 2022/5/13 𝑗 𝑁𝑛𝑒𝑔 𝑦𝑖 𝑗=1 𝑒 −𝐸𝜃 = 𝑒 −𝐸𝜃 𝑥𝑖,𝑦𝑖 + 𝑗 𝑁𝑛𝑒𝑔 𝑦𝑖 𝑗=1 𝑥𝑖 ,𝑦𝑖 𝑁𝑛𝑒𝑔 −𝐸𝜃 𝑥𝑖 ,𝑦 𝑗 𝑖 σ𝑗=1 𝑒 6
(CNN+) MLP 2022/5/13 7
Implicit Model 2022/5/13 8
Implicit Model 2022/5/13 9
Implicit Model 2022/5/13 10
2022/5/13 11
Bi-Manual Sweeping Task Explicit Model ( 2022/5/13 ) Implicit Model ( ) 12
Insertion Task Explicit Model ( 2022/5/13 ) Implicit Model ( ) 13
Sorting Task Explicit Model ( 2022/5/13 ) Implicit Model ( ) 14
◼ BC EBM ◼ ◼ ◼ RNN 2022/5/13 15