README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
configs:
- config_name: default
  data_files: data/*/*.parquet
---

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=zfff/sroiv2_strawberry_picking_lab_validation">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
</a>

# SROI v2 — Strawberry Picking (Lab) — Validation Set

![Example frames from four episodes](examples/montage.jpg)

Held-out **validation set** for the SROI v2 strawberry-picking data ([project page](https://agroboticsresearch.github.io/sroi_v2/), Zhejiang University): 100 human strawberry-picking demonstrations recorded with the **SROI V2 handheld data-acquisition device** — a UMI-style gripper with an integrated Intel RealSense D405 stereo camera — on live plants in a laboratory setup. **No robot arm is involved during collection**: the 7-DoF end-effector actions are recovered off-line (ORB-SLAM3 stereo SLAM with gripper mask for the device trajectory; AprilTags for the gripper opening), and the learned policy is deployed on a robot arm carrying the same end effector, whose camera viewpoint is identical to the one in these recordings.

- **Task:** `pick the strawberry` (single task, all 100 episodes)
- **Collection:** human demonstrations with the SROI V2 handheld device (UMI-style — not recorded on a robot arm)
- **Camera:** Intel RealSense D405 mounted on the device (18 mm stereo baseline, zero distortion), 480×640 RGB, 30 fps, first-person view
- **Project:** [agroboticsresearch.github.io/sroi_v2](https://agroboticsresearch.github.io/sroi_v2/) · [all SROI datasets](https://agroboticsresearch.github.io/sroi_datasets/)
- **Training counterpart:** [`zfff/sroiv2_strawberry_picking_lab_1459_occlusion`](https://huggingface.co/datasets/zfff/sroiv2_strawberry_picking_lab_1459_occlusion)

This collection was **recorded on a different day (2026-07-14)** than the training recordings (2026-07-09 onward), so there is **no episode leakage** between train and validation. Task, fps, and schema are identical, so it drops in as a validation holdout.

## Quickstart

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("zfff/sroiv2_strawberry_picking_lab_validation")
episode = dataset[0]  # dict with "observation.images.camera", "action", ...
```

## Dataset Summary

| | |
|---|---|
| Episodes | 100 |
| Frames | 9,274 |
| fps | 30 |
| Format | LeRobot v3.0 (Parquet + AV1 video) |
| Observation | `observation.images.camera` — 480×640×3 RGB video |
| Action | `action` — 7-D float32: `ee.x, ee.y, ee.z, ee.wx, ee.wy, ee.wz, ee.gripper_pos` |
| Total size | ~155 MB |

## Processing

Recorded MP4s → frame decode → ORB-SLAM3 stereo trajectory estimation (**with gripper mask**) → trajectory transform → AprilTag-based gripper pose estimation (median filter 3) → visual QC → LeRobot conversion.

- All 100 recorded episodes passed visual QC (`ok` rating, 100 kept / 0 dropped).
- The masked-SLAM processing matches the training pipeline, so train and validation trajectories are directly comparable.
- The gripper position channel is normalized to `[0, 1]` with one robust pooled range across these 100 episodes.
- Per-episode camera intrinsics are preserved under `meta/camera_info/`.

## Dataset Structure

[meta/info.json](meta/info.json):
```json
{
    "codebase_version": "v3.0",
    "fps": 30,
    "features": {
        "observation.images.camera": {
            "dtype": "video",
            "shape": [480, 640, 3],
            "names": ["height", "width", "channels"],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "action": {
            "dtype": "float32",
            "names": ["ee.x", "ee.y", "ee.z", "ee.wx", "ee.wy", "ee.wz", "ee.gripper_pos"],
            "shape": [7]
        },
        "timestamp": {"dtype": "float32", "shape": [1], "names": null},
        "frame_index": {"dtype": "int64", "shape": [1], "names": null},
        "episode_index": {"dtype": "int64", "shape": [1], "names": null},
        "index": {"dtype": "int64", "shape": [1], "names": null},
        "task_index": {"dtype": "int64", "shape": [1], "names": null}
    },
    "total_episodes": 100,
    "total_frames": 9274,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
    "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
    "robot_type": "so100",
    "splits": {"train": "0:100"}
}
```

> **Note:** `robot_type: "so100"` in `meta/info.json` is a hardcoded default of the conversion script (`sroi_to_lerobot.py`) and does **not** describe the collection rig — this data is human-collected with the SROI V2 handheld device.

## Citation

If you use this dataset, you are welcome to cite:

> Hou, L., Lu, W., Wang, Y., Peng, C., & Fei, Z. (2025). Strawberry Robotic Operation Interface: An Open-Source Device for Collecting Dexterous Manipulation Data in Robotic Strawberry Cultivation. *IFAC-PapersOnLine, 59*(23), 303–308.

## License

Apache-2.0