Communications of the ACM
Robots Have a Hard Time Grasping "Adversarial Objects"
University of California, Berkeley researchers developed physical adversarial objects to be tricky for conventional robot grippers to grasp.
Credit: University of California, Berkeley
Researchers at the University of California, Berkeley have extended the concept of adversarial images—which are images of things that have been modified to be especially difficult for computer vision algorithms to accurately identify—to robot grasping, using physical adversarial objects that are designed to be tricky for conventional robot grippers to pick up.
The key to creating adversarial objects is that they must look easy to grasp. For example, a cube with some shallow pyramids on three of the six sides causes significant problems for two-finger grippers.
In preliminary testing, a parallel jaw gripper with point contact fingers tried to pick up some adversarial objects. In each case, the computed grasp was predicted to succeed 100% of the time, but the actual success rates on the adversarial objects was just 13%.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2019 SmithBucklin, Washington, DC, USA
No entries found