SLIDE 1
NVIDIA GTC 2018
Large-Scale Self-supervised Robot Learning with GPU-enabled Video-Prediction Models
Frederik Ebert, Chelsea Finn, Alex Lee, Sergey Levine
1
1969 Stanford Arm 2015 DARPA Robotics Challenge Typical Bar in 20??
Large-Scale Self-supervised Robot Learning with GPU-enabled - - PowerPoint PPT Presentation
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
SLIDE 2
SLIDE 3
Humans have excellent mental models
- f physical objects
3
SLIDE 4
How can robots acquire general models and skills using large amounts of autonomously collected data?
4
SLIDE 5
Related work on self-supervised learning
Gandhi et al. 2017 Levine et al. 2016 Pinto & Gupta, 2015 5
Predict raw sensory inputs instead of binary events.
SLIDE 6
Related work on video-prediction
Finn & Levine 2017 Visual Model-Predictive Control
6
Mathieu et al. 2016 Oh et al. 2015 Byravan et al. 2017
SLIDE 7
7
Visual Model-Predictive Control
Designated Pixel Goal Point User Input Video Prediction Model Cost Function Planning Module [Finn et al. 2017] apply action to Robot
SLIDE 8
Random Data Collection
Collected 45,000 trajectories, recording camera images and actions
8
SLIDE 9
Action-Conditioned Video Prediction
Action 0 Recurrent NN
Real
Action 2 state Recurrent NN Action 1 state Recurrent NN
Generated
9
SLIDE 10
Skip Connection Neural Advection (SNA)
DNA (Finn et al.)
10
SLIDE 11
Action-Conditioned Video Prediction
Action 0 Recurrent NN
Real
Action 2 state Recurrent NN Action 1 state Recurrent NN
Generated
11
Temporal Skip Connections
SLIDE 12
Skip Connection Neural Advection (SNA)
DNA (Finn et al.) SNA (Ours)
12
SLIDE 13
conv conv 3×3 64 64 de conv 3×3 stride 2 conv 3×3 stride 2 de conv 3×3 conv 3×3 conv 1×1 conv 3×3 stride 2 de conv 3×3 stride 2 de conv 3×3 stride 2 de conv 3×3 conv 3×3 10 CDNA ke rne ls 9 9 conv 9×9 de conv 3×3 and channe l softmax 11 compositing masks 32x32x32 16 32 16x16x16 32 32 16 64x64x16 33 maske d compositing 64 skip skip 16x16x64 8x8x64 32x32x32
Conv-LSTM Masks
5x5 stride 2
13
Skip Connection Neural Advection (SNA)
Convolutional LSTM Image of current time step CDNA Kernels
- gen. Image of next time step
Image from first time step, temporal skip connection
SLIDE 14
Action 2 Recurrent NN state
Prediction of Pixel Positions (Test Time)
Action 0 Recurrent NN Action 1 state Recurrent NN
Generated
14
SLIDE 15
Effects of using temporal skip connections
DNA (Finn et al.) SNA (Ours) Designated Pixel
15
SLIDE 16
Planning with Visual-MPC
Designated Pixel Goal Pixel
16
SLIDE 17
Planning: Expected Distance to Goal Cost
Predicted Distribution for designated Pixel Designated Pixel Goal Point Distance to Goal
17
SLIDE 18
Action Selection using Cross-Entropy Method
Iteration 1 Iteration 2 Iteration 3 Designated Pixel Goal Pixel
18
SLIDE 19
19
Results
SLIDE 20
20
Generalization to objects not seen during training
SLIDE 21
Designated Pixel Goal Pixel Static Pixel
Collision Avoidance Task, involving Occlusion
21
SLIDE 22
Finn et al.
SLIDE 23
23
Ours
SLIDE 24
Multi-Goal Pushing Benchmark
24
SLIDE 25
25
SLIDE 26
26
SLIDE 27
27
SLIDE 28
28
SLIDE 29
29
SLIDE 30
30
SLIDE 31
- 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
Takeaways
SLIDE 32
Q&A
Code and Data: https://sites.google.com/view/sna-visual-mpc
Chelsea Finn Alex X. Lee Sergey Levine