Enhance dataset card: Add paper, code, project page, task category, and sample usage

README.md

@@ -1,5 +1,116 @@
 ---
-license: mit
 language:
 - en
----
+license: mit
+task_categories:
+- robotics
+---
+
+# TacThru-UMI Tasks Dataset
+
+This repository contains the datasets used in the paper [Simultaneous Tactile-Visual Perception for Learning Multimodal Robot Manipulation](https://huggingface.co/papers/2512.09851).
+
+This work introduces **TacThru**, a novel see-through-skin (STS) sensor enabling simultaneous visual perception and robust tactile signal extraction, and **TacThru-UMI**, an imitation learning framework that leverages these multimodal signals for robotic manipulation. The datasets provided here are generated through this framework and are used to train and evaluate generalist policies for precise manipulation tasks on five challenging real-world tasks.
+
+*   **Paper:** [https://huggingface.co/papers/2512.09851](https://huggingface.co/papers/2512.09851)
+*   **Project Page:** [https://tacthru.yuyang.li/](https://tacthru.yuyang.li/)
+*   **Code/GitHub Repository:** [https://github.com/YuyangLee/TacThru](https://github.com/YuyangLee/TacThru)
+*   **Video:** [https://vimeo.com/1145307821](https://vimeo.com/1145307821)
+*   **Hardware Guide:** [https://docs.google.com/document/d/1fpZRiGoxWqLoFs-zxnG4d_d3hy0eHjlLA4nsuEKCvg/edit?usp=sharing](https://docs.google.com/document/d/1fpZRiGoxWqLoFs-zxnG4d_d3hy0eHjlLA4nsuEKCvg/edit?usp=sharing)
+
+## Dataset Tasks
+
+This dataset includes the following tasks used in our experiments:
+
+*   `PickBottle`
+*   `PullTissue`
+*   `SortBolt`
+*   `HangScissors`
+*   `InsertCap`
+
+## Dataset Structure
+
+The datasets are provided as Zarr files, following a structure similar to the example below. You can use `scripts/show_ds.py` from the associated codebase to inspect the Zarr files.
+
+```
+data/camera0_rgb               shape=(N, 224, 224, 3) dtype=uint8
+data/robot0_demo_end_pose      shape=(N, 6) dtype=float64
+data/robot0_demo_start_pose    shape=(N, 6) dtype=float64
+data/robot0_eef_pos            shape=(N, 3) dtype=float32
+data/robot0_eef_rot_axis_angle shape=(N, 3) dtype=float32
+data/robot0_gripper_width      shape=(N, 1) dtype=float32
+data/tacthru_l_marker          shape=(N, 64, 2) dtype=float32
+data/tacthru_l_rgb             shape=(N, 224, 224, 3) dtype=uint8
+data/tacthru_r_marker          shape=(N, 64, 2) dtype=float32
+data/tacthru_r_rgb             shape=(N, 224, 224, 3) dtype=uint8
+meta/episode_ends              shape=(M,) dtype=int64
+```
+where `N` is the total number of frames and `M` is the number of episodes.
+
+## Sample Usage
+
+To utilize these datasets, you will typically interact with the main [TacThru codebase](https://github.com/YuyangLee/TacThru).
+
+### Environment Setup
+
+First, clone the main repository and set up the environment:
+
+```shell
+git clone https://github.com/YuyangLee/TacThru
+cd TacThru
+```
+
+We use `uv` to manage the virtual environment. Install the basic dependencies:
+
+```shell
+uv sync
+```
+
+For policy training and validating robotic manipulation, include the optional `umi` dependencies:
+
+```shell
+uv sync --extra umi
+```
+
+### Download Datasets
+
+The datasets are provided as a Hugging Face Dataset and are set up as a submodule under `data/tasks/` in the main `TacThru` repository. To download them, ensure you have Git LFS installed and then run:
+
+```bash
+git submodule init
+git submodule update
+```
+
+You can also use [sparse checkout](https://git-scm.com/docs/git-sparse-checkout) to download partially.
+
+### Train Policy
+
+An example training script is provided in the main repository's `./train_tf.sh` file. Here's how you might run it for a specific task, such as `pick_bottle`:
+
+```shell
+task=pick_bottle
+
+# TacThru w/ marker deviations
+tac_active_keys="[tacthru_l_rgb,tacthru_l_markers]"
+obs_tag="tt_m"
+exp_tag="run"
+
+uv run scripts/train.py --config-name=train_tf exp_name=tf-$obs_tag-$exp_tag task=$task tac_active_keys=$tac_active_keys
+```
+
+The `task` can be one of: `pick_bottle`, `pull_tissue`, `sort_bolt`, `hang_scissors`, `insert_cap`.
+
+The `tac_active_keys` argument should include the observation keys defined in `train.task.shape_meta.obs`. In the provided datasets, `tacthru_l_*` refers to signals from the TacThru sensor (left finger), while `tacthru_r_*` refers to signals from the GelSight-type sensor (rectified).
+
+## Citation
+
+If you find our work helpful, please consider citing it:
+
+```bibtex
+@article{li2025simultaneous,
+  title={Simultaneous Tactile-Visual Perception for Learning Multimodal Robot Manipulation},
+  author={Yuyang Li and Yinghan Chen and Zihang Zhao and Puhao Li and Tengyu Liu and Siyuan Huang and Yixin Zhu},
+  journal={arXiv preprint arXiv:2512.09851},
+  year={2025}
+}
+```