| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - robotics |
| tags: |
| - tactile-sensing |
| - contact-fields |
| - manipulation |
| --- |
| |
| # SCFields Release Artifacts |
|
|
| This dataset repository contains assets, contact-field datasets, and contact-field checkpoints used by the SCFields release code. |
|
|
| **Paper:** [Semantic-Contact Fields for Category-Level Generalizable Tactile Tool Manipulation](https://huggingface.co/papers/2602.13833) |
|
|
| **Project page:** https://kevinskwk.github.io/SCFields |
| **Code:** https://github.com/Kevinskwk/SCFields |
|
|
| ## Layout |
|
|
| - `assets/tools`: generated TacSL tool assets. |
| - `assets/peeler_raw`: raw original peeler assets. |
| - `assets/peeler_combined`: combined peeler assets. |
| - `data/sim/tools`: simulated contact-field data for tools (`tools_mixed`). |
| - `data/sim/peelers`: simulated contact-field data for peelers (`peelers_combined_new`). |
| - `data/real/scraper`: converted real scraper contact-field data (`real_scraper_corrected_lambda1`). |
| - `checkpoints/contact_field/tools_sim`: tool contact-field checkpoint pretrained on simulated data. |
| - `checkpoints/contact_field/tools_real`: tool contact-field checkpoint finetuned on real data. |
| - `checkpoints/contact_field/peelers_sim`: peeler contact-field checkpoint pretrained on simulated data. |
| - `checkpoints/contact_field/peelers_real`: peeler contact-field checkpoint finetuned on real data. |
|
|
| Each checkpoint folder contains: |
|
|
| - `model.ckpt`: PyTorch Lightning checkpoint. |
| - `config.yaml`: the accompanying training configuration. |
|
|
| ## Usage |
|
|
| ### Download |
|
|
| You can use the Hugging Face CLI to download the artifacts: |
|
|
| ```bash |
| hf download Kevinskwk/scfields-release \ |
| --repo-type dataset \ |
| --include "assets/**" \ |
| --include "data/sim/**" \ |
| --include "data/real/scraper/**" \ |
| --include "checkpoints/contact_field/**" \ |
| --local-dir /path/to/scfields |
| ``` |
|
|
| The release code's `scripts/download_assets.sh` maps shortened hosted asset paths to the repo-local asset layout expected by IsaacGym: |
|
|
| ```text |
| assets/peeler_raw -> assets/peeler |
| assets/peeler_combined -> assets/peelers_combined |
| ``` |
|
|
| ### Training Example |
|
|
| To train the SCFields policy using the provided dataset and checkpoints, you can use the following command structure: |
|
|
| ```bash |
| python train.py \ |
| --config-dir=config/scraping_real \ |
| --config-name=contact_field_delta_ee.yaml \ |
| data_root=/path/to/scfields \ |
| task.dataset.dataset_dir=/path/to/scfields/data/real/real_scraper_corrected_lambda1 \ |
| task.dataset.contact_field_checkpoint_path=/path/to/scfields/checkpoints/contact_field/tools_real/model.ckpt |
| ``` |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{ma2026semanticcontactfieldscategorylevelgeneralizable, |
| title={Semantic-Contact Fields for Category-Level Generalizable Tactile Tool Manipulation}, |
| author={Kevin Yuchen Ma and Heng Zhang and Weisi Lin and Mike Zheng Shou and Yan Wu}, |
| year={2026}, |
| eprint={2602.13833}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.RO}, |
| url={https://arxiv.org/abs/2602.13833}, |
| } |
| ``` |
