RoboFS-Data

Data of RoboFS: Scalable Data Generation and Text-to-Action Learning for Robot Fingerspelling (WACV 2027 Workshop HOUR). Code: https://github.com/JunukCha/RoboFS · Project page: https://junukcha.github.io/project/robofs/

Path	Content
word_lists/{train,validation,test,train_50k}_words.txt	Word lists of the experiments
generated/{train,validation,test}.npz	Human fingerspelling landmarks generated by OpenFS, palm aligned
targets/relation/{train,validation,test}.npz	Relation-preserving retargeted robot targets
targets/relation_net/{validation,test}.npz	Targets predicted by the retargeting network
robofs_data/part_00.npz ... part_09.npz	RoboFS-Data: executed robot actions for 50,000 words, 5,000 per part

Download with bash scripts/download.sh eval-data, train-data or robofs-data in the code repository. The file formats are described in the code repository's README (section "Data formats"). MANIFEST.json lists every file with its size and SHA-256.

Sources

The words come from the label sets of FSboard (CC BY 4.0), ChicagoFSWild, ChicagoFSWild+ and FSNeo. The FSNeo (neobench) words were collected by the OpenFS authors from Reddit, news articles, the New York Times and dictionaries. The human landmarks were generated with the OpenFS generator; no video or landmark data of the source data sets is redistributed. Users are responsible for complying with the terms of these data sets. Please also cite OpenFS.

License

RoboFS Non-Commercial Research License (see LICENSE): non-commercial research and educational use only.

Citation

@inproceedings{kim2027robofs,
  title     = {RoboFS: Scalable Data Generation and Text-to-Action Learning for Robot Fingerspelling},
  author    = {Kim, Jihyeon and Cha, Junuk},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
  year      = {2027}
}