LabSafe-1700

LabSafe-1700 is the first VLA (Vision-Language-Action) dataset targeting safe robotic manipulation in biochemical laboratory scenarios. It contains 1,700 real-robot teleoperation trajectories collected on a SO-100 robotic arm, covering 17 task configurations including obstacle-free grasping and active obstacle avoidance tasks.

This dataset is introduced in the paper:

NeuroMPC-VLA: Embedding MPC-Structured Safety Mechanisms into Vision-Language-Action Models for Safe Manipulation in Biochemical Laboratories

---

📦 Dataset Overview

Property	Value
Total trajectories	1,700 (17 tasks × ~100 episodes each)
Task configurations	17 (10 simple + 7 obstacle avoidance)
Robot platform	SO-100 6-DOF robotic arm
Camera views	3× RGB cameras (OBS_IMAGE_1/2/3), 480×640, 30 fps
Annotation	Task labels per episode
Scene	Biochemical laboratory tabletop manipulation
Format	LeRobot v3.0 native (parquet + AV1 video)

Task Configurations

Type	Tasks	Description
Simple (obstacle-free)	T01–T10	Grasping and placing lab instruments (test tubes, beakers, funnels, etc.) without obstacles
Obstacle avoidance	T11–T17	Same manipulation goals with one or two obstacles placed randomly on the path

---

🗂️ Repository Structure

LeRobot records data in the following structure:

LabSafe-1700/
├── data/
│   └── chunk-000/
│       ├── file-000.parquet    # frame-level data (actions, states, timestamps)
│       └── ...
├── videos/
│   ├── observation.images.OBS_IMAGE_1/
│   │   └── chunk-000/
│   │       ├── file-000.mp4   # AV1, 480×640, 30fps
│   │       └── ...
│   ├── observation.images.OBS_IMAGE_2/
│   └── observation.images.OBS_IMAGE_3/
├── meta/
│   ├── info.json               # dataset-level metadata (LeRobot v3.0)
│   ├── episodes.jsonl          # per-episode metadata (task index, length)
│   ├── tasks.jsonl             # task descriptions
│   └── stats.json              # feature statistics
├── LICENSE
└── README.md

---

🚀 Quick Start

Install LeRobot

pip install lerobot

Load the dataset

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("Merlyn-L/LabSafe-1700")

# Iterate over episodes
for i in range(dataset.num_episodes):
    episode = dataset.get_episode(i)
    frames_cam1 = episode["observation.images.OBS_IMAGE_1"]  # (T, 480, 640, 3)
    frames_cam2 = episode["observation.images.OBS_IMAGE_2"]  # (T, 480, 640, 3)
    frames_cam3 = episode["observation.images.OBS_IMAGE_3"]  # (T, 480, 640, 3)
    actions = episode["action"]                    # (T, 6) joint positions
    print(f"Episode {i}: {len(actions)} steps")

---

📊 Dataset Statistics

Split	Episodes	Avg. steps	Description
Simple tasks (T01–T10)	~1,000	~120	Obstacle-free
Obstacle avoidance (T11–T17)	~700	~180	With obstacles
Total	1,700	—	—

---

🤖 Hardware Setup

• Robot: SO-100 follower arm (so100_follower)
• Teleoperation: Leader-follower SO-100 arm via LeRobot
• Cameras: 3× RGB cameras (OBS_IMAGE_1/2/3), 480×640 @ 30 fps, AV1 codec
• Lab instruments: Test tubes, beakers, funnels, micro-well plates, centrifuge tubes

---

📄 Citation

If you use LabSafe-1700 in your research, please cite:

@misc{neurompc_vla_2025,
  title  = {NeuroMPC-VLA: Embedding MPC-Structured Safety Mechanisms into
            Vision-Language-Action Models for Safe Manipulation in
            Biochemical Laboratories},
  author = {NEU-ECAL},
  year   = {2025}
}

---

📜 License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
You are free to share and adapt the data for any purpose, provided appropriate credit is given.

---

🙏 Acknowledgements

Data collection was conducted using the LeRobot teleoperation framework on a SO-100 robotic arm. We thank the Hugging Face Robotics team for the open-source tools that made this dataset possible.