Combining imitation and reinforcement learning to fold deformable planar objects

Research on robotic manipulation has primarily focused on grasping rigid objects using a single manipulator. It is however evident that in order to be truly pervasive, service robots will need to handle deformable objects, possibly with two arms. In this paper we tackle the problem of using cooperative manipulators to perform towel folding tasks. Differently from other approaches, our method executes what we call a momentum fold - a swinging motion that exploits the dynamics of the object being manipulated. We propose a new learning algorithm that combines imitation and reinforcement learning. Human demonstrations are used to reduce the search space of the reinforcement learning algorithm, which then quickly converges to its final solution. The strengths of the algorithm come from its efficient processing, fast learning capabilities, absence of a deformable object model, and applicability to other problems exhibiting temporally incoherent parameter spaces. A wide range of experiments were performed on a robotic platform, demonstrating the algorithm's capability and practicality.