Due to the deformability of garments, generating a large amount of diverse and high-quality data for robotic garment manipulation tasks is highly challenging. In this paper, we present FoldNet, a synthetic garment dataset that includes assets for four categories of clothing as well as high-quality closed-loop folding demonstrations. We begin by constructing geometric garment templates based on keypoints and applying generative models to generate realistic texture patterns. Leveraging these garment assets, we generate folding demonstrations in simulation and train folding policies via closed-loop imitation learning. To improve robustness, we introduce KG-DAgger, a keypoint-based strategy for generating recovery demonstrations after failures. KG-DAgger significantly improves the quality of generated demonstrations and the model performance, boosting the real-world success rate by 25%. After training with 15K trajectories (about 2M image-action pairs), the model achieves a 75% success rate in the real world. Experiments in both simulation and real-world settings validate the effectiveness of our proposed dataset.
Pipeline for garment mesh synthesis. By performing geometry generation, texture generation, combining-and-filtering, we can synthesize scalable, high-quality garment meshes.
Synthetic garment meshes. These static garment meshes can be used for subsequent physics simulation and policy learning.
Common method: generating perfect demonstration data.
Our method (KG-DAgger): generating demonstration data that includes recovery from errors.
We can generate large-scale and diverse demonstration data in simulation.
Our policy is robust to small perturbations in the real world.
Our policy is robust to small perturbations in the real world.
@misc{chen2025foldnetlearninggeneralizableclosedloop,
author={Chen, Yuxing and Xiao, Bowen and Wang, He},
journal={IEEE Robotics and Automation Letters},
title={FoldNet: Learning Generalizable Closed-Loop Policy for Garment Folding via Keypoint-Driven Asset and Demonstration Synthesis},
year={2026},
volume={},
number={},
pages={1-8},
keywords={Clothing;Geometry;Imitation learning;Annotations;Trajectory;Training;Synthetic data;Pipelines;Grasping;Filtering;Bimanual manipulation;deep learning for visual perception;deep learning in grasping and manipulation},
doi={10.1109/LRA.2026.3656770}}
}