FoldNet++

Official code for FoldNet++: a Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding.

• Project: https://pku-epic.github.io/FoldNetXX/
• Dataset: https://huggingface.co/datasets/GalaxyGeneralRobotics/FoldNetXX

Requirements

• A CUDA GPU with 48 GB of memory is recommended (24 GB cards may OOM).
• uv for dependency management.

Installation

# Install dependencies
uv sync

# Download the dataset (tar shards), then unpack into storage/release
HF_TOKEN=hf_your_token uv run hf download GalaxyGeneralRobotics/FoldNetXX --repo-type dataset --local-dir storage/release_tar
uv run python script/unpack_tar.py -i storage/release_tar -o storage/release

# Download pretrained backbones (VAE, PaliGemma, ResNet) into storage/pretrained
HF_TOKEN=hf_your_token uv run python script/prepare_ckpt.py

# Link the pi0 checkpoint used for pretraining
mkdir -p storage/pretrained/openpi
ln -s ../../release/ckpt/pi0.safetensors storage/pretrained/openpi/pi0.safetensors

HF_TOKEN is your Hugging Face access token; it is required for gated repos such as google/paligemma-3b-mix-224.

Usage

Dataset visualization

See example/visualize_data.ipynb.

Training

Pretraining:

SWANLAB_API_KEY=your_api_key bash example/pretrain.sh

Finetuning:

SWANLAB_API_KEY=your_api_key bash example/finetune.sh

Both scripts auto-detect the number of GPUs on the node. For multi-node training, set WORLD_SIZE, RANK, MASTER_ADDR, and MASTER_PORT before launching.

Repository structure

script/                 entry-point scripts
  train_vla.py          VLA training (launched with torchrun)
  unpack_tar.py         unpack the downloaded dataset tars
  prepare_ckpt.py       download pretrained backbones
example/                runnable examples
  pretrain.sh           pretraining launch script
  finetune.sh           finetuning launch script
  visualize_data.ipynb  dataset visualization notebook
src/foldnetxx_net/      main package
  config.py             model and data configs
  dataset/              data loading and preprocessing
  model/                VLA and diffusion-policy models
  server/               policy inference agents for deployment
  train/                training loop utilities
  utils/                args, logging, file management
storage/                datasets, checkpoints, pretrained weights (git-ignored)

Citation

@inproceedings{
chen2026foldnetxx,
title={FoldNet++: a Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding},
author={Yuxing Chen and Zhiyuan Wei and Bowen Xiao and Zhizheng Zhang and He Wang},
booktitle={10th Annual Conference on Robot Learning},
year={2026},
url={https://openreview.net/forum?id=JMcC5JoPKD}
}