Success-Guided Sampling picks robot-training tasks that aren't too easy or too hard

The SGS team says its method changes which tasks robots practice during reinforcement learning, favoring those neither too easy nor too hard. It reports that a UR5e trained only in simulation threaded nuts, inserted rods and meshed gears on a real task board using RGB images. The team also reports locomotion tests in simulation and improving success rates as training scaled to a million parallel environments.