Fast Towel-Shaper: Pick-and-Place Towel Manipulation using Transporter Skill Policies under Goal-condition Phase Prediction
Towel manipulation is a crucial stepstone towards more general cloth manipulation, but it remains a challenge to fold a towel from any crumple state and recover from a failed folding step. In this work, we develop a two-layer hierarchical agent that uses a goal-condition phase-prediction network to decide which manipulation phase the towel is under and employs the associated pre-trained pick-and-place transporter skill policy to manipulate the towel. We adopt behaviour cloning to improve operational efficiency, transporter nets to promise picking accuracy, and a phase-prediction network to allow the system to recover from its failing states.
Keywords
robotics, control systems, deep learning, computer vision, robot vision, imitation learning, Artificial Intelligence
Staff
Abudureyimu Halite,[Kasim Terzic]{kt54}