Learning Deformable Object Manipulation From Expert Demonstrations

We present a novel Learning from Demonstration (LfD) method, Deformable Manipulation Demonstrations (DMfD), to solve deformable manipulation tasks using states or images as inputs, given expert demonstrations. Our method uses demonstrations in three different ways, and balances the trade-off between exploring environment online guidance experts explore high dimensional spaces effectively. test DMfD on set of representative for 1-dimensional rope 2-dimensional cloth SoftGym suite tasks, each with state image observations. exceeds baseline performance by up 12.9% state-based 33.44% image-based comparable better robustness randomness. Additionally, we create two challenging environments folding 2D observations, benchmark them. deploy real robot minimal loss normalized during real-world execution compared simulation (~6%). Source code is github.com/uscresl/dmfd