Manipulating deformable objects is a difficult problem in the field of robotics. A recent paper on arXiv.org proposes the Learn from Demonstration – Manipulate Deformable from Demonstration (DMfD) as a solution. It absorbs expert guidance while learning online to tackle challenging deformable manipulative tasks like folding fabric.
The researchers add an exploration term to the weight reduction advantage to encourage widespread discovery. Instead of always resetting the agent to the states seen by the expert, the reference state initialization is called probabilistically. That promotes discovery and learning in hard-to-reach states.
DMfD is implemented on a real robot with minimal sim2real distance, thus showing that it can work in real world settings. This method outperforms baselines on both state-based and image-based environments.
We introduce a new method of Learning from Demonstration (LfD), Deformable Manipulation from Demonstration (DMfD), to solve deformable manipulation tasks using states or images as input, expert evidence. Our method uses representations in three different ways and strikes a balance between exploring the online environment and using guidance from experts to explore dimensional space efficiently. We test DMfD on a set of manipulation tasks representing 1-dimensional rope and 2-dimensional fabric from the SoftGym task set, each with state and image observations. Our method exceeds baseline performance by up to 12.9% for state-based tasks and up to 33.44% for image-based tasks, with equal or better strength than randomness. In addition, we created two challenging environments to fold a 2D fabric using image-based observations and set a performance benchmark for them. We implement DMfD on a real robot with minimal loss of normalized performance during real-world execution compared to simulation (~6%). The source code is on This http URL
Research articles: Salhotra, G., Liu, I.-CA, Dominguez-Kuhne, M., and Sukhatme, GS, “Learning to manipulate deformed objects from expert demonstrations”, 2022. Links: https://arxiv.org/abs/2207.110148
