📦 Dataset Card for dock_the_iron

Overview

dock_the_iron is a LeRobot-compatible robot learning dataset designed for fine-tuning vision-language-action (VLA) policies on a real-world manipulation task: docking a soldering iron into its station.

The dataset includes synchronized multi-view video observations, robot joint states, and continuous action trajectories of 50 episodes/demonstrations, making it suitable for imitation learning and policy fine-tuning.

---

📘 Dataset Details

---

---

🎯 Intended Uses

This dataset is released as an open-source resource for:

• Robot learning researchers
• Students and practitioners exploring imitation learning
• Fine-tuning vision-language-action (VLA) models

Direct Use

The dataset can be directly used with the LeRobot framework to fine-tune OpenPi VLA policies, including (but not limited to):

• Pi0
• Pi0.5

---

🗂 Dataset Structure

The dataset follows the standard LeRobot directory layout, consisting of structured parquet files and synchronized multi-view videos.

dock_the_iron/
├── data/
│   └── chunk-000/
│       └── file-000.parquet
│           # Key columns:
│           # - action: [j1, j2, j3, j4, j5, j6, j7, gripper_pos]
│           # - observation.state: [j1, j2, j3, j4, j5, j6, j7, gripper_pos]
├── meta/
│   ├── episodes/
│   ├── info.json
│   ├── stats.json
│   └── tasks.parquet
└── videos/
    ├── observation.images.front_image/
    │   └── chunk-000/
    │       ├── file-000.mp4
    │       ├── file-001.mp4
    │       ├── file-002.mp4
    │       └── file-003.mp4
    ├── observation.images.side_image/
    │   └── chunk-000/
    │       ├── file-000.mp4
    │       ├── file-001.mp4
    │       └── file-002.mp4
    └── observation.images.wrist_image/
        └── chunk-000/
            ├── file-000.mp4
            ├── file-001.mp4
            ├── file-002.mp4
            ├── file-003.mp4
            ├── file-004.mp4
            ├── file-005.mp4
            ├── file-006.mp4
            ├── file-007.mp4
            ├── file-008.mp4
            ├── file-009.mp4
            └── file-010.mp4

Dataset Creation

This dataset was collected in December 2025 at the
Department of Agricultural and Biological Engineering (ABE),
Mississippi State University.

---

Data Collection & Processing

Hardware Setup

• Robot Arm: Kinova Gen3 (7-DoF)
• Gripper: Robotiq 2F-85
• Teleoperation Interface: Xbox Controller
• Cameras:
  • Intel RealSense D435i × 2 (external views)
  • Kinova wrist-mounted RGB camera × 1

Task Specification

• Task Prompt:

  "Dock the soldering iron into the station."

Demonstrations were performed via teleoperation and recorded using the LeRobot data collection pipeline. All modalities were synchronized and stored following the standard LeRobot dataset format.

Actions

• Actions are represented as absolute joint positions of the robot arm.
• The gripper command is represented as an absolute gripper position.
• Action vector format: [j1, j2, j3, j4, j5, j6, j7, gripper_pos]

Citation [optional]

If you use this dataset, please cite it using the following.

BibTeX:

@dataset{adhikari2025docktheiron,
  author       = {Adhikari, Bishal},
  title        = {dock\_the\_iron: A LeRobot Dataset for Robotic Manipulation},
  year         = {2025},
  publisher    = {Hugging Face},
  organization = {Mississippi State University},
  url          = {https://huggingface.co/datasets/winnerBishal/dock_the_iron}
}

Dataset Card Contact

Bishal Adhikar ba974@msstate.edu