Learning Visual Feedback Control for Dynamic Cloth Folding
Julius Hietala, David Blanco-Mulero, Gokhan Alcan, Ville Kyrki
School of Electrical Engineering, Aalto University, Finland
2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2022)
Abstract
Robotic manipulation of cloth is a challenging task due to the high dimensionality of the configuration space and the complexity of dynamics affected by various material properties. The effect of complex dynamics is even more pronounced in dynamic folding, for example, when a square piece of fabric is folded in two by a single manipulator. To account for the complexity and uncertainties, feedback of the cloth state using e.g. vision is typically needed. However, construction of visual feedback policies for dynamic cloth folding is an open problem.
In this paper, we present a solution that learns policies in simulation using Reinforcement Learning (RL) and transfers the learned policies directly to the real world. In addition, to learn a single policy that manipulates multiple materials, we randomize the material properties in simulation. We evaluate the contributions of visual feedback and material randomization in real-world experiments. The experimental results demonstrate that the proposed solution can fold successfully different fabric types using dynamic manipulation in the real world.
Video
Below is a video overview of the problem and the solution we propose in the paper.
Method Overview
The cloth folding policies are trained in simulation and transferred directly to the real world. The transfer is enabled by using Domain Randomization (DR) and an identical Cartesian controller both in Sim and Real.
Experiments
We train 3 different policies
Fixed
Trained in simulation without visual feedback (full states) and without cloth dynamics randomization
Tested in the real world without visual feedback (fixed trajectories from simulation success)
Visual
Trained in simulation with visual feedback, domain randomization, and no cloth dynamics randomization
Tested in the real world with visual feedback
Proposed Method
Trained in simulation with visual feedback, domain randomization, and cloth dynamics randomization
Tested in the real world with visual feedback
Fixed Policy
X Visual Feedback
X Cloth dynamics randomization
Distance = 5.5 cm
Distance = 4.2 cm
Distance = 2.6 cm
Visual Policy
✓ Visual Feedback
X Cloth dynamics randomization
Distance = 1.1 cm
Distance = 2.5 cm
Distance = 2.9 cm
Proposed Method
✓ Visual Feedback
✓ Cloth dynamics randomization
Distance = 0.3 cm
Distance = 0.7 cm
Distance = 1.4 cm
Results
The results indicate that using both visual feedback and dynamics domain randomization is essential for consistent performance across different materials.
Mean trajectories from the different policies also show that only the policy trained with dynamics domain randomization is able to react to the differences in behavior caused by the different dynamics.
Fixed Policy
Visual Policy
Proposed Method
Citation
To cite this work, please use the following BibTex entry:
@inproceedings{hietala2022closing,
author={Hietala, Julius and Blanco–Mulero, David and Alcan, Gokhan and Kyrki, Ville},
booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
title={Learning Visual Feedback Control for Dynamic Cloth Folding},
year={2022},
volume={},
number={},
pages={1455-1462},
doi={10.1109/IROS47612.2022.9981376}}