Rapid Motor Adaptation – Exploring in-Hand Object Rotation With a Robotic hand

Rapid Motor Adaptation for In-Hand Rotation of Objects

Robotics has been unable to solve the problem of generalized in-hand manipulating for a long time. We demonstrate a small but important step in achieving this goal by demonstrating how to learn and design a simple adaptive control to rotate objects using just fingertips. The controller can be trained in simulation using only cylindrical objects. It then, without fine-tuning, can be deployed directly to a robot hand and rotate dozens objects of different sizes, shapes, weights, etc. over the z axis. The robot controller is quickly adapted online to the object properties by using the proprioception data. Moreover, the reinforcement learning technique is used to train the control policy. This results in a natural and stable gait.