Dynamic in-hand manipulation tasks can be challenging for soft robots. SWIFT is a system for learning these types of tasks using a soft robotic hand. As Carnegie Mellon University researchers explain, after 130 sampled actions, SWIFT achieved 100% success rate of different weights and weight distributions.
To pull this off, researchers used a soft robotic end effector with three tendon-driven soft robot fingers, each driven by two servos controlling four tendons. They pull on them to bend the fingers. You can find out more about it here.