MSU-Fold model checkpoints

Pre-trained checkpoints for MSU-Fold: Unified Multi-scale Heatmap Decoding and Language-grounded Skill Decomposition for Multi-step Robotic Cloth Folding.

File	Benchmark	Training data
unimanual_1000demos_last.pth	Single-arm SoftGym benchmark (CornerFold, TriangleFold, StraightFold, TshirtFold, TrousersFold)	1000 demonstrations per task, 30 epochs
bimanual_1000demos_best.pth	Dual-arm mixed-step benchmark (SquareHalfFold, SquareCornerFold, TshirtFold, TrousersFold)	150 demonstrations per task, 40 epochs

Both checkpoints share the same architecture: a frozen SigLIP-Base backbone with LoRA (r=8) and a token-wise multi-scale visual adapter (K=8), a Transformer fusion network, and a single shared pick/place heatmap head that serves both unimanual and bimanual steps (the number of extracted peaks is dictated by each step).

Usage

git clone https://github.com/hua-369/MSU-Fold
cd MSU-Fold
conda env create -f environment.yml && conda activate MSU-Fold
bash deps/prepare.sh && bash deps/compile.sh   # build PyFlex
bash run_skill_softgym.sh --skill unimanual-tshirt-fold --num-evals 1

Point the skill definitions (skills/*/SKILL.md) or --checkpoint to the downloaded file.

Datasets are available at https://huggingface.co/datasets/huangxuan111/MSU-Fold-datasets.