kjw6213/OmniVLA-serbotII-dataset

SERBOT Navigation Dataset

This dataset contains SERBOT navigation episodes collected for OmniVLA-style vision-language-action experiments.

Dataset Summary

• Episodes: 82
• Valid episodes: 82
• Frames: 4,343
• Actions: 4,342
• Action dimension: 3 (linear_x, linear_y, angular_z)
• Action horizon used by the local index: 8
• Tasks: task1 through task5

Structure

serbot_data/
  dataset_index.json
  manifest.jsonl
  task1/
    episode_.../
      actions.csv
      frames/
      goal_img.jpg
      instruction.txt
      meta.json

Each manifest.jsonl row points to a frame, the corresponding goal image, instruction text, task, episode, action index, and velocity command fields.

Task Counts

Task	Episodes	Samples
task1	22	1,430
task2	20	986
task3	20	826
task4	10	494
task5	10	606

Files

• dataset_index.json: dataset-level summary and per-episode metadata.
• manifest.jsonl: frame-level sample manifest.
• actions.csv: per-episode action commands.
• frames/*.jpg: observation frames.
• goal_img.jpg: goal image for each episode.
• instruction.txt: language instruction for each episode.
• meta.json: episode metadata.