ClothDojo NOCS generation teachers
This repository contains the NOCS generation teachers for garment flattening and folding. Code and installation scripts are maintained in the ClothDojo GitHub repository. Assets and initial states are provided separately in clothdojo/assets.
| Task | Weights | Matching normalization |
|---|---|---|
| Flatten | flatten/weights |
flatten/norm_stats |
| Fold | fold/weights |
fold/norm_stats |
Quickstart
Use Linux x86_64, Python 3.11, Git, patch, uv, and an NVIDIA GPU with native BF16 support. Isaac Sim is not required for NOCS generation.
git clone https://github.com/clothdojo/clothdojo.github.io.git ClothDojo
cd ClothDojo
bash scripts/download.sh flatten
bash scripts/setup.sh
bash scripts/quickstart.sh flatten
The run script generates one trajectory and prints the high and low camera MP4 paths. To try folding in the same environment:
bash scripts/download.sh fold
bash scripts/quickstart.sh fold
See the Quickstart guide for requirements and output details. This model repository supplies weights, normalization, and model metadata; it does not supply a second code runtime.
Identity and terms
Keep each task's weights, normalization, policy.json, and base_model.json together. Do not mix normalization sets. Exact checkpoint identities and provenance are retained in those files.
Both checkpoints are task fine-tuned modifications of pi0.5, stored in the native RLinf wrapper layout. Gemma terms, NOTICE, and use restrictions apply. The upstream tokenizer and caches are not included.
The Fold generation teacher differs from the historical Fold benchmark policy; do not attribute that policy's score to this checkpoint. These are NOCS generation teachers, not released RGB policy checkpoints.