MSaalaamaa/marine_vla_dataset

Marine VLA Dataset

Vision-Language-Action dataset for autonomous marine vessel navigation using SmolVLA.

Dataset Structure

LeRobot-style format with 37 episodes, 12175 frames:

data/
  episode_000000/
    episode_data.json       # frame-by-frame labels + metadata
    observation.images.camera_0/
      000000.jpg             # 640x480 RGB frames
      000001.jpg
      ...
  episode_000001/
  ...
dataset_info.json            # schema, label names, stats
marine_vla_dataset.tar.gz    # single-file archive of entire dataset

Labels (8 classes)

Label	Description
PROCEED	Go straight at normal speed
PORT_15	Turn 15° port (left)
STARBOARD_15	Turn 15° starboard (right)
REDUCE_SPEED	Slow down
STOP	Full stop
ASTERN	Reverse
AVOID_OBSTACLE	Emergency obstacle avoidance maneuver
DOCK_APPROACH	Slow approach to dock

Usage

Option 1: Extract archive (fastest)

# Download the archive
wget https://huggingface.co/datasets/MSaalaamaa/marine_vla_dataset/resolve/main/marine_vla_dataset.tar.gz

# Extract
tar -xzf marine_vla_dataset.tar.gz

Option 2: Load with Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("MSaalaamaa/marine_vla_dataset")

Option 3: Load with LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("MSaalaamaa/marine_vla_dataset")

Data Collection

Recorded with ROS Noetic, UUV Simulator, Gazebo 11 in a marine harbor environment. Expert controller uses an 8-label finite-state machine with obstacle detection via collision cone (90° field, 25m danger radius).

Stats

• Episodes: 37
• Total frames: 12,175
• Resolution: 640x480
• Rate: ~10.3 Hz
• AVOID_OBSTACLE frames: 171