Data-Efficient Robotic Manipulation of Deformable One-dimensional Objects with Unreliable Dynamics

Abstract

If I were to ask you to fold a towel, you would probably do it without much thought or difficulty -- but how? How did you know where the towel was? Did you think about how the towel would move or deform, and how you'd react if the towel didn't move as you predicted? Or what if I asked you to fold it using your feet, could you do it?

These questions hopefully help explain why a robot, which lacks the intuition and experience of an adult human, might find this task challenging. A robot may need specific computational answers to all of these questions and more. In more technical terms, a robot needs to perceive, predict, and plan. Many methods for perceiving, predicting, and planning have been proposed, and they can be effective for some tasks in specific environments, but we have not yet achieved sufficiently fast and accurate methods to perform complex tasks like pruning a plant, sewing sutures, or installing cable harnesses -- at least not outside carefully controlled settings.

This dissertation takes a step to increase the abilities of robots to manipulate deformable objects. More specifically, deformable objects like ropes, hoses, and cables. I focus on two problems, our inability to accurately predict in all scenarios how these objects will move or deform (called unreliable dynamics), and the cost of collecting training data in robotic manipulation. I present a system that can quickly learn despite limited data and imperfect predictions, as well as plan grasps and regrasps when the robot gets stuck.