FoldNet++: a Cross-embodiment T-shirt Folding and Unfolding Model Pre-trained on Synthetic Data

Due to the deformable nature of garments, training a generalizable foundation model for T-shirt folding and unfolding that can transfer across different robotic embodiments is highly challenging. In this work, we propose FoldNet++, a cross-embodiment foundation model for T-shirt folding and unfolding trained entirely on large-scale synthetic data. For training data synthesis, we first adopt the approach of FoldNet to produce a large-scale set of physically simulatable T-shirts with diverse appearances and annotated semantic keypoints. Leveraging these semantic keypoints, we generate demonstrations across different robotic embodiments using a unified rule-based framework. The demonstrations covers T-shirt of arbitrary initial configurations. Then we use these demonstrations to train a vision-based model. Through large-scale synthesized cross-embodiment data, our model demonstrates strong zero-shot and few-shot capabilities across different robots. Our model can transfer to real-world environments without any real-world fine-tuning and can achieve a 93% success rate over the entire pipeline.