zfff/strawberry_picking_dataset_scara_clustered

Strawberry Picking Dataset — SCARA Clustered

This LeRobot v3.0 dataset contains human-teleoperated demonstrations of a 4-DoF SCARA robot performing clustered strawberry picking in a controlled physical mock-up. It accompanies the work Learning to Pick: A Visuomotor Policy for Clustered Strawberry Picking, which studies imitation-learned visuomotor control for reaching and picking a target strawberry amid leaves, stems, and neighboring fruit.

The scene uses artificial strawberry plants, red and white rubber strawberries, flexible-wire stems, and magnetic fruit attachment. Demonstrations include maneuvering around or gently displacing occluding objects before grasping the target stem and detaching the fruit. This is a physical robot dataset collected in a controlled artificial environment; it is not a physics-simulator dataset.

Dataset summary

Property	Value
LeRobot format	v3.0
Episodes	511
Frames	119,830
Stored sampling rate	30 Hz
Total stored duration	3,994.3 s (1 h 6 min 34.3 s)
Episode length	129–598 frames; median 219 frames
Split	train (all 511 episodes)
Image streams	top, wrist, wrist_down
Image shape	480 × 640 × 3
Video encoding	AV1, yuv420p
Robot state	4D SCARA joint position
Action	4D SCARA joint command + 1D gripper command

This corrected release is stored at 30 Hz, matching the policy operating rate reported in the publication. An earlier Hub revision incorrectly labeled and timestamped the data at a nominal 50 Hz; no frames were removed by the 30 Hz correction.

Data structure

Each frame contains:

Feature	Shape	Description
observation.state	[4]	SCARA joint positions: j1, j2, j3, j4
action	[5]	Joint commands j1–j4 and gripper
observation.images.top	[480, 640, 3]	RGB video stream retained from the source camera key top
observation.images.wrist	[480, 640, 3]	RGB video stream retained from the source camera key wrist
observation.images.wrist_down	[480, 640, 3]	RGB video stream retained from the source camera key wrist_down
timestamp	scalar	Episode-relative timestamp in seconds
frame_index	scalar	Frame index within the episode
episode_index	scalar	Global episode index
task_index	scalar	Source-session provenance index

The joint and gripper units and sign conventions are retained from the source ACT/ALOHA HDF5 demonstrations. Confirm them against your robot interface before deploying a trained policy.

Example camera views

The following synchronized frames illustrate the three RGB observations available in each demonstration:

observation.images.top	observation.images.wrist	observation.images.wrist_down

The top camera provides a broad view of the SCARA workspace, the wrist camera looks forward through the plant structure, and the wrist-down camera provides a close view around the gripper and target fruit.

Provenance groups

The six task strings (data1–data6) identify source recording sessions. They are provenance labels, not semantic picking-task or occlusion-state labels.

Label	Episodes	Frames
data1	53	15,255
data2	76	18,133
data3	71	16,821
data4	131	31,015
data5	69	13,785
data6	111	24,821

Per-episode source filenames, frame counts, source-session paths, excluded validation files, and color-conversion provenance are recorded in manifest.json.

Collection setup

The associated study uses a 4-DoF SCARA arm and a human teleoperation interface. The operator directs the end effector around or through clutter, grasps the stem above the strawberry calyx, and detaches the fruit. Initial robot positions and cluster arrangements vary between demonstrations. The paper describes five clustered configurations and an unobstructed baseline; the training demonstrations comprise the clustered scenarios and total 511 episodes.

The paper investigates ACT and End Pose Assisted ACT (EPACT) policies and the effect of multiple camera viewpoints. Consult the publication for robot design, teleoperation, policy architecture, training settings, and evaluation results.

Robot model and action visualization

hitbot_model.urdf is the SolidWorks-exported HitBot SCARA model used by the deployment visualization code. The code parses the URDF with kdl_parser_py and PyKDL, constructs the kinematic chain from world to the wrist-mounted camera_link, and applies forward kinematics to predicted four-joint action chunks. The resulting poses are projected into the wrist RGB image with the camera intrinsics to draw the predicted action trajectory.

The URDF is included primarily for its joint origins, axes, and kinematic structure. Its visual and collision elements refer to package://hitbot_model/meshes/*.STL; those mesh files are not included and are not needed for the trajectory-overlay calculation.

Loading with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="zfff/strawberry_picking_dataset_scara_clustered"
)

print(dataset.num_episodes)  # 511
sample = dataset[0]
print(sample["observation.state"].shape)       # (4,)
print(sample["action"].shape)                  # (5,)
print(sample["observation.images.wrist"].shape)  # (3, 480, 640)

An AV1-capable video backend is required. LeRobot may use TorchCodec or fall back to PyAV depending on the local installation.

Intended use

This dataset is intended for research on:

• imitation learning and action-chunking policies;
• visuomotor control for agricultural manipulation;
• clustered and occluded strawberry picking;
• multi-view robot learning;
• SCARA manipulation and teleoperation learning.

Limitations

• The environment uses artificial plants and fruit in a controlled laboratory setup. Appearance, deformation, lighting, and contact dynamics differ from commercial farms and real strawberry plants.
• All episodes are provided in one training split; no official test split is included in this release.
• Source groups are not semantic labels for the paper's cluster scenarios.
• The three camera keys retain their source names and should be visually checked before selecting or interpreting viewpoints.
• Policies trained on these demonstrations should not be deployed without independent safety validation, calibration, and real-robot testing.

Conversion notes

The release was converted from ACT/ALOHA HDF5 demonstrations to LeRobot v3.0. Vector features were preserved as float32. Source frames stored in BGR order were converted to RGB; conversion details are recorded per source session in manifest.json. Videos use AV1 encoding with CRF 30, GOP size 2, and yuv420p.

License

This dataset is released under the MIT License.

Citation

If you use this dataset, please cite:

@article{fei2026learning,
  title={Learning to Pick: A Visuomotor Policy for Clustered Strawberry Picking},
  author={Fei, Zhenghao and Lu, Wenwu and Hou, Linsheng and Ye, Zunzhong and Peng, Chen},
  journal={IEEE Transactions on AgriFood Electronics},
  year={2026},
  publisher={IEEE}
}

Paper: https://arxiv.org/abs/2509.14530