README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - robotics
  - timeseries
  - tsfile
  - format:tsfile
modality:
  - tabular
  - timeseries
pretty_name: 2view_random50 TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/two_view_random50.tsfile
size_categories:
  - 10K<n<100K

2view_random50 TsFile

This dataset is an Apache TsFile conversion of the Hugging Face dataset iiyudana/2view_random50. The source dataset was created with LeRobot and is licensed under Apache 2.0.

Modalities: Time-series. The original visual MP4 streams remain available in the source dataset; this repository stores the numeric robot state, action, frame metadata, task index, and episode tags as TsFile.

Source Dataset

• Original dataset: iiyudana/2view_random50
• Source task: "Pick up the cube with the right arm and transfer it to the left arm."
• Codebase version: LeRobot v2.1
• Robot type: aloha
• Split: train (0:50)
• Episodes: 50
• Frames: 20,000
• Sampling rate: 50 fps
• Tasks: 1
• Videos: 100 MP4 files across observation.images.top and observation.images.right_wrist
• Source data path: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video path: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Converted Files

• TsFile: data/two_view_random50.tsfile
• Rows: 20,000
• Table name: two_view_random50
• Time precision: milliseconds
• Mirrored metadata: meta/, with meta/info.json updated for the TsFile artifact

Schema

Time is generated as round(timestamp * 1000) milliseconds. Time restarts within each episode, and devices are identified by the original LeRobot tag columns.

• TAG columns: episode_index, task_index
• FIELD metadata columns: frame_index, sample_index
• FIELD vectors: observation_state_0..observation_state_13 and action_0..action_13

Conversion Notes

• The source timestamp column is dropped after being mapped to Time; it is recoverable as Time / 1000 seconds.
• The source index column is renamed to sample_index.
• Vector columns are flattened by preserving the full source column name, replacing . with _, and appending the element index.
• Source video features are not uploaded here: observation.images.top and observation.images.right_wrist. Use the original dataset's videos/ tree for the visual streams.
• Aside from the redundant timestamp column and omitted video files, no numeric time-series rows are intentionally dropped.

Read Example

from tsfile import TsFileReader

path = "data/two_view_random50.tsfile"

with TsFileReader(path) as reader:
    tables = reader.get_all_table_schemas()
    print(tables.keys())