RoboFS models

Checkpoints 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/

File	Model
tracker/model_1000.pt, tracker/params/{env,agent}.yaml	RoboFS-Tracker, the PPO tracking policy (Isaac Lab, Shadow Hand with a fixed wrist)
retarget_net/model.pt	Retargeting network that predicts relation-preserving reference commands
t2a/model.pt	RoboFS-T2A, trained on RoboFS-Data (50k words) to predict robot actions from input words

Download with bash scripts/download.sh models in the code repository. MANIFEST.json lists every file with its size and SHA-256.

The OpenFS generator and recognizer are not included; they are downloaded from 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}
}