UseAppliance training split (JSON + images)

Training annotations and images associated with UseAppliance, built via MAGE as described in:

Scaling Manual-Grounded Appliance Manipulation with Data Synthesis and Unified Planning (ACM MM ’26)
https://doi.org/10.1145/3767308.3836196

Layout

Train_data/*.json   # ShareGPT-style multimodal samples
images.tar          # unpack to get images/ (paths match JSON refs)
stats.json

Setup / paths

# from this dataset root (HF: NikolaKang/UseAppliance)
tar -xf images.tar
# yields: images/...  (JSON fields like images/<...> are relative to IMAGE_PATH)

export IMAGE_PATH=/path/to/NikolaKang/UseAppliance   # this folder
export ROOT_PATH=$IMAGE_PATH
# annotations: $IMAGE_PATH/Train_data/*.json
# images:      $IMAGE_PATH/images/...
# model:       export MODEL_PATH=/path/to/NikolaKang/AppliancePlan
# merge JSON:  tools/merge_train_data.py in the code repo

Code	https://github.com/Niko-kang/AppliancePlan
Model	https://huggingface.co/NikolaKang/AppliancePlan

Covers plan / bbox / keypage / closed-loop style supervision used to train AppliancePlan.

Citation

@inproceedings{long2026applianceplan,
  title     = {Scaling Manual-Grounded Appliance Manipulation with Data Synthesis and Unified Planning},
  author    = {Long, Yuxing and Kang, Lei and Yu, Ziyan and Gao, Yuzheng and Cheng, Bin
               and Zhang, Jiyao and Li, Xiaoqi and Yang, Haolin and Li, Dongjiang
               and Shen, Hui and Dong, Hao},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia (MM '26)},
  year      = {2026},
  doi       = {10.1145/3767308.3836196}
}