Upload README.md with huggingface_hub

README.md

@@ -1,186 +1,61 @@
 ---
 pretty_name: DexBioLab
 license: cc-by-4.0
-language:
-  - en
-tags:
-  - robotics
-  - dexterous-manipulation
-  - imitation-learning
-  - simulation
-  - laboratory-automation
-  - lerobot
 ---
 
 # 🧪 DexBioLab
 
 **Benchmarking Dexterous Manipulation with Dynamic Experimental Phenomena in Realistic Laboratories**
 
-[Paper](https://arxiv.org/abs/2610.13643) · [Project Page](https://cumulusalpha.github.io/DexBioLabWebPage/) · [Code](https://github.com/CumulusAlpha/DexBioLab) · [Dataset](https://huggingface.co/datasets/22364yiqun/DexBioLab)
+DexBioLab is a benchmark for dexterous robotic manipulation in realistic laboratory environments. It provides a simulation-based platform for laboratory task execution, scripted demonstration collection, LeRobot dataset generation, and learned policy evaluation.
 
-DexBioLab provides simulated demonstrations for learning dexterous robotic manipulation in laboratory environments. This repository contains task-specific datasets in **LeRobot v3.0 format**, including robot observations, actions, multi-view RGB videos, task descriptions, and simulation joint-position arrays for replay.
+## Paper
 
-The accompanying benchmark studies tool manipulation and laboratory instrument operation under changing visual conditions. Its framework combines scripted demonstration collection, trajectory replay, dataset rendering, and closed-loop policy evaluation. See the [paper](https://arxiv.org/abs/2610.13643) for the benchmark methodology and experimental results.
+This dataset accompanies the following paper:
 
-## Dataset Overview
+- **Paper:** [DexBioLab: Benchmarking Dexterous Manipulation with Dynamic Experimental Phenomena in Realistic Laboratories](https://arxiv.org/abs/2610.13643)
+- **Hugging Face Paper Page:** https://huggingface.co/papers/2610.13643
 
-The current release contains **13 task datasets across four embodiment families: Aloha, Shadow, Wuji, and XHand**.
-
-| Property | Current release |
-| --- | --- |
-| Task directories | 13 |
-| Demonstration episodes | 1,201 |
-| Timesteps | 672,939 |
-| Sampling frequency | 20 Hz |
-| Trajectory duration | Approximately 9.35 hours, before counting camera views separately |
-| RGB cameras | Three views per task |
-| Image resolution | 640 × 480 pixels |
-| Video format | MP4 containers with AV1-encoded RGB video |
-| Data format | LeRobot v3.0: Parquet tables, metadata, and external videos |
-| Published splits | `train` only; episode ranges are specified in each task's `meta/info.json` |
-
-These statistics are derived from the published `meta/info.json` files at revision [`541fc035957deb449c3e64bac5c8dbd20347d63c`](https://huggingface.co/datasets/22364yiqun/DexBioLab/tree/541fc035957deb449c3e64bac5c8dbd20347d63c). A timestep contains multiple camera observations; the frame count above is not the sum of video frames across cameras.
-
-## Available Tasks
-
-Each directory is an independent LeRobot dataset. Observation and action dimensions vary by embodiment and task.
-
-| Task directory | Episodes | Timesteps | State dimension | Action dimension |
-| --- | ---: | ---: | ---: | ---: |
-| `ABPP-Aloha-Screw-Tighten` | 100 | 111,578 | 16 | 16 |
-| `ABPP-Shadow-Centrifuge-Insert` | 100 | 43,231 | 28 | 25 |
-| `ABPP-Shadow-Pipette-Insert` | 100 | 32,064 | 28 | 25 |
-| `ABPP-Shadow-Pipette-Manipulate` | 100 | 32,454 | 28 | 25 |
-| `ABPP-Shadow-Tube-Insert` | 100 | 38,336 | 28 | 25 |
-| `ABPP-Shadow-Tube-Rotate` | 1 | 828 | 56 | 50 |
-| `ABPP-Wuji-Centrifuge-Close` | 100 | 38,200 | 26 | 27 |
-| `ABPP-Wuji-HeatingPlate-Rotate` | 100 | 62,274 | 26 | 27 |
-| `ABPP-Wuji-Tweezers-Manipulate` | 100 | 54,849 | 26 | 27 |
-| `ABPP-XHand-Bunsen-Manipulate` | 100 | 73,850 | 18 | 19 |
-| `ABPP-XHand-MixerPanel-Manipulate` | 100 | 40,775 | 18 | 19 |
-| `ABPP-XHand-ThermalCycler-Close` | 100 | 74,300 | 18 | 19 |
-| `ABPP-XHand-ThermalCycler-Open` | 100 | 70,200 | 18 | 19 |
-
-`ABPP-Shadow-Tube-Rotate` currently contains one episode. `ABPP-XHand-Tube-Insert` is not included in this release. The published task coverage and episode counts should be distinguished from the full benchmark described in the paper.
-
-## Repository Structure
-
-```text
-DexBioLab/
-├── README.md
-├── ABPP-Shadow-Tube-Insert/
-│   ├── data/
-│   │   └── chunk-000/
-│   │       └── file-000.parquet
-│   ├── meta/
-│   │   ├── info.json
-│   │   ├── stats.json
-│   │   ├── tasks.parquet
-│   │   └── episodes/
-│   │       └── chunk-000/
-│   │           └── file-000.parquet
-│   ├── videos/
-│   │   ├── observation.images.front/
-│   │   ├── observation.images.left/
-│   │   └── observation.images.wrist1/
-│   └── replay/
-│       └── qpos.npy
-└── ...
-```
+## Dataset Overview
 
-Video files are stored in chunks beneath each camera directory. Camera names depend on the task: some tasks use `front`, `left`, and `wrist1`, while others use `front`, `wrist1`, and `wrist2`.
+DexBioLab supports research on laboratory manipulation, including tool use, precise object interactions, instrument operation, and dynamic experimental phenomena.
 
-| Component | Contents |
-| --- | --- |
-| `data/` | Per-timestep observations, actions, timestamps, and episode/task indices |
-| `meta/info.json` | Format version, sampling frequency, feature shapes, file patterns, and collection-specific metadata |
-| `meta/stats.json` | Feature statistics for normalization |
-| `meta/tasks.parquet` | Task indices and language descriptions |
-| `meta/episodes/` | Episode metadata, including the indexing information used to align tables and video segments |
-| `videos/` | RGB observations stored as external video files |
-| `replay/qpos.npy` | Simulation joint-position arrays; replay requires the matching simulator, assets, and task configuration |
+The benchmark and simulation framework support:
 
-The tables contain `observation.state` and `action` as `float32` vectors, along with `timestamp`, `frame_index`, `episode_index`, `index`, and `task_index`. Consult the corresponding `meta/info.json` and simulator configuration for coordinate conventions, action semantics, and feature dimensions. The `ABPP-*` identifiers are preserved from dataset generation; consult the code repository for current simulator task IDs.
+- **Simulation:** Execute laboratory manipulation tasks in realistic simulated environments.
+- **Demonstration Collection:** Generate scripted robot demonstrations.
+- **Dataset Generation:** Render simulation trajectories into LeRobot-compatible datasets.
+- **Policy Evaluation:** Evaluate learned manipulation policies and analyze task success rates.
 
-## Download and Usage
+This repository contains task-specific LeRobot datasets with robot observations, actions, multi-view RGB videos, and task metadata. For benchmark details and experimental results, please refer to the paper.
 
-### Download one task
+## Project Resources
 
-Install the Hub client:
+- **Project Page:** https://cumulusalpha.github.io/DexBioLabWebPage/
+- **GitHub Repository:** https://github.com/CumulusAlpha/DexBioLab
+- **Hugging Face Dataset:** https://huggingface.co/datasets/22364yiqun/DexBioLab
 
-```bash
-pip install -U huggingface_hub
-```
+## Usage
 
-Download only the task you need:
+Download the dataset using `huggingface_hub`:
 
 ```python
 from huggingface_hub import snapshot_download
 
-task = "ABPP-Shadow-Tube-Insert"
 snapshot_download(
     repo_id="22364yiqun/DexBioLab",
     repo_type="dataset",
-    allow_patterns=[f"{task}/**"],
     local_dir="./DexBioLab",
-    max_workers=1,
-)
-```
-
-For reproducible experiments, also pass a full commit hash using `revision`. Omit `allow_patterns` to download the complete repository. See the [Hub download guide](https://huggingface.co/docs/huggingface_hub/en/guides/download) for additional options.
-
-### Load a local task with LeRobot
-
-Use a LeRobot installation compatible with the v3.0 dataset format and an AV1-capable video decoder. After downloading a complete task directory:
-
-```python
-from pathlib import Path
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-task = "ABPP-Shadow-Tube-Insert"
-dataset = LeRobotDataset(
-    repo_id="22364yiqun/DexBioLab",
-    root=Path("./DexBioLab") / task,
 )
-
-sample = dataset[0]
-print(sample["observation.state"].shape)
-print(sample["action"].shape)
-```
-
-The local `root` must point to the selected task directory, where `meta/info.json` is located. The repository root contains multiple independent datasets with different feature shapes, so select a task rather than treating the entire repository as one uniform dataset.
-
-### Inspect numerical tables without decoding video
-
-```python
-from pathlib import Path
-from datasets import load_dataset
-
-root = Path("./DexBioLab/ABPP-Shadow-Tube-Insert")
-files = sorted(str(p) for p in (root / "data").glob("chunk-*/*.parquet"))
-table = load_dataset("parquet", data_files={"train": files}, split="train")
-print(table[0])
 ```
 
-This reads the timestep tables only. Use LeRobot to synchronize and decode the external video observations.
+Each task directory contains an independent LeRobot dataset. To download a single task, add `allow_patterns=["ABPP-Shadow-Tube-Insert/**"]` to the call above.
 
-## Intended Use and Limitations
-
-The dataset supports imitation learning, vision-language-action policy development, and analysis of laboratory manipulation trajectories. The [official code repository](https://github.com/CumulusAlpha/DexBioLab) provides simulation setup, demonstration collection, rendering, and policy evaluation workflows.
-
-These demonstrations are generated in simulation. Performance on this dataset alone does not establish real-world reliability or safe laboratory operation. The release provides no separate validation or test split; report any custom episode splits and evaluation protocol. The single-episode tube-rotation dataset should not be assumed to have the same coverage as the 100-episode task datasets. Closed-loop task success must be evaluated in the corresponding environment rather than inferred from offline training metrics.
-
-## Paper and Project Resources
-
-- **Paper:** [DexBioLab: Benchmarking Dexterous Manipulation with Dynamic Experimental Phenomena in Realistic Laboratories](https://arxiv.org/abs/2610.13643)
-- **Hugging Face Paper Page:** https://huggingface.co/papers/2610.13643
-- **Project Page:** https://cumulusalpha.github.io/DexBioLabWebPage/
-- **GitHub Repository:** https://github.com/CumulusAlpha/DexBioLab
-- **Dataset Repository:** https://huggingface.co/datasets/22364yiqun/DexBioLab
+For simulation setup, demonstration collection, dataset generation, and policy evaluation, see the [official GitHub repository](https://github.com/CumulusAlpha/DexBioLab).
 
 ## License
 
-This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. You may share and adapt the licensed material, including for commercial purposes, provided that you give appropriate credit, link to the license, and indicate any changes. The simulation code and third-party assets are governed by their respective licenses.
+This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. Please provide appropriate attribution when using the dataset.
 
 ## Citation
 
@@ -192,7 +67,6 @@ If you use DexBioLab in your research, please cite:
   author = {Wu, Xuanye and Shi, Yue and Yang, Yongcan and Wang, Yiqun and Wan, Zhaoqi and Xia, Dayu and Shao, Yanming and Cao, Jiezhang and Li, Xiangfei and Mu, Yao and Zhai, Guangtao},
   journal = {arXiv preprint arXiv:2610.13643},
   year = {2026},
-  doi = {10.48550/arXiv.2610.13643},
   url = {https://arxiv.org/abs/2610.13643}
 }
 ```