huangxuan111/MSU-Fold-datasets

MSU-Fold multi-step folding datasets

File	Tasks	Demonstrations	Size
single_data_sequential/All_100.pkl	CornerFold, TriangleFold, StraightFold, TshirtFold, TrousersFold (unimanual benchmark)	100 per task	160 MB
single_data_sequential/All_1000.pkl	same tasks	1000 per task	1.6 GB
dual_data_sequential/All_150.pkl	MixedSquareHalfFold, MixedSquareCornerFold, MixedTshirtFold, MixedTrousersFold	150 per task	230 MB

Each episode contains, for every atomic step:

• rgbs / depth: RGB-D observation (224×224);
• pick / place: pixel-level keypoint annotations (one row per arm);
• primitive: execution form of the step (single, dual, or multi = form not specified in the source demonstrations);
• instruction: the natural-language instruction of the step;
• success: whether the expert step achieved its intended fold.

Loading

import pickle

with open("dual_data_sequential/All_150.pkl", "rb") as f:
    episodes = pickle.load(f)["episodes"]

ep = episodes[0]                 # one demonstration (2 steps)
print(ep["instruction"][0])      # per-step instruction
print(ep["pick"][0])             # pixel keypoints (K x 2)
print(ep["rgbs"][0].shape)       # (224, 224, 3)

Code: https://github.com/hua-369/MSU-Fold