This page contains demos of our attacks. For the patch attack, we show the rollout video. For global attacks, we show videos of the attacked frames. Each video has two images because the condition image encoder takes the past two image as input. The inference uses receding horizon control so there are motion gaps between frames of the video.
Embedded Files
Physical Patched Attacks (Offline - Untargeted)
Lift
Can
Square
Toolhang
Local Attacks (Online - Targeted - Target=1's)
PushT
Can
Lift
Square
Transport
Toolhang
Local Attacks (Online - Untargeted)
PushT
Can
Lift
Square
Transport
Toolhang
Global Attacks (Offline - Targeted - Target = 1's)
PushT
Can
Lift
Square
Transport
Toolhang
Global Attacks (Offline - Untargeted)
PushT
Can
Lift
Square
Transport
Toolhang
With larger attack budget, we can fully control the robots action. The following uses attack parameters: eps=0.05, N=50, sigma=0.002. We first show the attacked frames.
Attacked frames:
Target = normalized action of -1's (top left corner).
Target = normalized action of -1's (top left corner).
Target = normalized action of 1's (bottom right corner).
Target = normalized action of 1's (bottom right corner).
Rollouts:
Target = normalized action of -1's (top left corner).
Target = normalized action of 1's (bottom right corner).
Target = normalized action of -1's (top left corner).
Target = normalized action of 1's (bottom right corner).
Page updated
Google Sites
Report abuse