Robotic Fabric Manipulation with Deep Imitation Learning and Reinforcement Learning in Simulation

As the ongoing COVID-19 pandemic spread around the world, endangering the lives of our loved ones and overwhelming many of our institutions, it exposed a need for automation. From our supply chain to our hospitals, automation can step in where it’s far too dangerous for humans. Unfortunately, while great progress has been made in robotics and machine learning research, many open problems remain to be solved before we can safely deploy reliable and intelligent robots in the real world. One such problem is manipulation of highly deformable structures such as fabric, which, for example, could be helpful in making hospital beds [50] and caring for seniors in nursing homes [19]. Fabric manipulation also has applications in sewing [48], ironing [29], laundry folding [36, 30, 68, 52], manufacturing upholstery [60], handling surgical gauze [57], and more. In Chapter 3, we consider learning a policy for the task of fabric smoothing: sequentially maximizing coverage of fabric in highly crumpled initial configurations. Fabric smoothing is both a starting point for further fabric manipulation and an interesting problem in its own right, as it standardizes the setup of fabric for subsequent tasks like folding. This chapter covers [49], work done by myself, Daniel Seita, Adi Ganapathi, Professor Goldberg and others. The paper contributes an open source fabric simulator, a smoothing policy learned via imitation of an algorithmic supervisor, and experimental validation on a physical robotic system with comparison of color and depth modalities. In Chapter 4, we generalize the problem to goal-conditioned fabric manipulation: sequentially manipulating fabric toward …