Science Robotics

ONLINE COVER LearningBeyond Imitation. Robot learning, rooted in machine learning and human-robot interaction, is moving beyond simple kinematic planning, environment interaction, and elementary behavioral imitation from human demonstrators. For example, universal picking, or reliable robot grasping of a diverse range of objects from heaps, is a challenging goal. Mahler et al. trained Dex-Net 4.0 on a synthetic dataset including images, grasps, and rewards. The resulting policy allowed a robot with two grippers to consistently and reliably clear bins containing up to 25 novel objects. The combination of learning from synthetic data and applying to real-world situations may improve e-commerce, manufacturing, inspection, and home service robots. [CREDIT: JEFF MAHLER, STEPHEN MCKINLEY, KEN GOLDBERG/UNIVERSITY OF CALIFORNIA, BERKELEY; PHOTO: ADRIEL OLMOS]

See the rest here:

Science Robotics

Related Post

Comments are closed.