Robots Grip Better when They Grip Smarter
IEEE Spectrum
Charles Q. Choi, December 28, 2022
A robot has been trained by researchers from Carnegie Mellon University’s Robotics Institute to grasp items of various sizes, weights, shapes, and surfaces. The team had the robot attempt to grab items from an open bin that were initially oriented in ways that would not allow the robot to pick them up. Researchers used reinforcement to train a neural network; as the system attempted random actions to grasp an object, it was rewarded when those actions led to success. The system eventually adopted the most successful patterns of behavior. After first training the system in a physics simulator, researchers then tested it in a simple robot with a pincer-like grip.
No comments:
Post a Comment