MSaalaamaa/marine_vla_dataset_v2

Marine VLA Dataset — v2 (Fixed Expert)

This dataset contains 12,070 frames collected in the Mangalia harbor Gazebo simulation, used to fine-tune a vision-language-action (VLA) model for marine navigation.

Built with opencode + DeepSeek V4 Flash — every component of this project (simulation setup, expert controller, data pipeline, ML training, inference nodes) was developed using opencode CLI with DeepSeek V4 Flash as the underlying model.

v2 Changes (Fixed Expert)

The v1 dataset was collected with a yaw-spin bug in the expert controller: the vessel SDF had zero hydrodynamic damping, causing uncontrolled yaw rotation. Labels oscillated between PORT_15 / STARBOARD_15 / PROCEED as the vessel wobbled.

Impact: A SmolVLA 450M classifier trained on v1 achieved only 25.7% accuracy (see ml_training_lead/scripts/launch_colab_training_fixes.ipynb, Cell 9). PROCEED, PORT_15, and STARBOARD_15 were all at 0% precision/recall.

v2 fixes:

• Added angular velocity decay (<velocity_decay angular=0.01>) to the vessel SDF
• Added proportional heading-hold during PROCEED to prevent oscillation
• Removed lateral force (Fy) from turns — uses only differential torque
• Corrected port/starboard sign convention

Label Vocabulary

ID	Label	Description
0	PROCEED	Continue on heading at cruise speed
1	PORT_15	Turn 15° to port (left)
2	STARBOARD_15	Turn 15° to starboard (right)
3	REDUCE_SPEED	Slow to minimum steerage
4	STOP	Stop engines
5	ASTERN	Reverse thrust
6	AVOID_OBSTACLE	Emergency evasive maneuver
7	DOCK_APPROACH	Orient and approach dock

Data Format

LeRobot-compatible format with image + label pairs:

• observation.images.camera_0: 640×480 RGB, ~50KB JPG
• action.goal_label: integer label ID 0-7

Collection Pipeline

1. ROS Noetic + Gazebo 11 with mangalia harbor world
2. Custom vessel SDF with camera + force plugin
3. Scenario loader randomizes start pose, waypoints, obstacles
4. Expert controller (fixed) drives vessel + publishes labels
5. Dataset recorder saves synchronized image + label pairs
6. Converter produces LeRobot-format JSON

Repository Structure

• marine_vla_dataset_v2.tar.gz — full dataset archive (~95MB)
• Episodes: ~250 frames each, 25-30s duration
• Train/test split: 80/20 by episode