README.md

metadata

pretty_name: eval2_compositional_augmented TsFile
tags:
  - TsFile
  - time-series
  - robotics
  - LeRobot
  - SO-101
  - imitation-learning
modality: timeseries
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/eval2_compositional_augmented.tsfile

eval2_compositional_augmented

This dataset was converted from robot-learning-group47/eval2_compositional_augmented to Apache TsFile format.

Modalities: Time-series.

Source Dataset

• Original dataset: robot-learning-group47/eval2_compositional_augmented
• Source format: LeRobot v3 parquet frames with metadata and one front-camera video stream
• Robot type: so_follower
• Split: train
• Episodes: 720
• Frames / TsFile rows: 161,780
• Tasks: 22
• Sampling rate: 15 fps

The source contains compositional prompt augmentations for banana-and-bowl manipulation tasks. The file combo_augmented_map.csv is included with this conversion and maps prompt-combination groups to source and augmented episode ranges.

Files

• data/eval2_compositional_augmented.tsfile: converted numeric time-series table
• meta/: source metadata mirrored from the original dataset, with meta/info.json rewritten to describe the converted TsFile layout
• combo_augmented_map.csv: source augmentation map

Videos are not included in this repository. The original camera video stream observation.images.front remains in the source dataset under videos/.

Schema

The TsFile table name is eval2_compositional_augmented.

• Time: integer timestamp in milliseconds, synthesized as round(timestamp * 1000)
• TAG columns: episode_index, task_index
• FIELD columns: frame_index, sample_index, action_0..action_5, observation_state_0..observation_state_5

The six action and observation-state dimensions use the original LeRobot joint order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos.

Conversion Notes

• Converted with the generic lerobot converter in the HuggingFace-to-TsFile pipeline.
• All 720 episodes are stored in one TsFile. Use episode_index and task_index as device TAG filters when querying.
• Source action[6] was flattened to action_0..action_5 as FLOAT fields.
• Source observation.state[6] was flattened to observation_state_0..5 as FLOAT fields.
• Source index was renamed to sample_index.
• Source timestamp is not retained as a separate field because the same information is represented by Time / 1000 seconds.
• Source video field observation.images.front is omitted from this TsFile conversion and remains available in the original dataset.
• No rows were dropped.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/eval2_compositional_augmented.tsfile")
schemas = reader.get_all_table_schemas()
print(schemas.keys())