THULab/svla_so100_stacking

SVLA SO100 Stacking (TsFile)

This dataset is a TsFile conversion of the Hugging Face dataset lerobot/svla_so100_stacking. It contains the numeric LeRobot frame data for an so100 robot dataset. The source dataset was created using LeRobot.

Modalities: Time-series. The original dataset also contains videos; those video files are not mirrored in this converted repository.

Source dataset

• Original dataset: lerobot/svla_so100_stacking
• Source commit used for conversion: 476c6810591b5e75b385d6e8d9c648525a153eb9
• License: apache-2.0
• LeRobot codebase version: v3.0
• Robot type: so100
• Published task metadata: one task index (0); the source task table does not include a text label

Dataset scale

• Episodes: 56
• Frames / converted TsFile rows: 22,956
• Tasks: 1
• Split: train (0:56)
• Sampling rate: 30 FPS
• Converted files: 1 TsFile file (429,148 bytes)
• Original video streams: observation.images.top and observation.images.wrist

Converted layout

data/svla_so100_stacking_train.tsfile
meta/
  episodes/chunk-000/file-000.parquet
  info.json
  stats.json
  tasks.parquet

The uploaded meta/info.json mirrors the source metadata and adds a tsfile_conversion object describing the source-to-TsFile mappings.

TsFile schema

Table name: svla_so100_stacking_train

• Time: millisecond timestamp, computed as round(timestamp * 1000).
• TAG columns: episode_index, task_index.
• FIELD columns: frame_index, sample_index, action_0 to action_5, and observation_state_0 to observation_state_5.

The six action_* and six observation_state_* fields preserve the source joint order:

main_shoulder_pan
main_shoulder_lift
main_elbow_flex
main_wrist_flex
main_wrist_roll
main_gripper

Conversion notes

• Source vector column action is flattened to action_0 through action_5.
• Source vector column observation.state is flattened to observation_state_0 through observation_state_5.
• Source column timestamp is mapped to the TsFile Time column in milliseconds and is not retained as a separate field because it is equivalent to Time / 1000 seconds.
• Source column index is renamed to sample_index; frame_index is retained.
• Source video features observation.images.top and observation.images.wrist are omitted from this TsFile conversion. The original videos remain available in the source dataset under videos/.

Minimal read example

import pyarrow as pa
from tsfile import ColumnCategory, TsFileReader

path = "data/svla_so100_stacking_train.tsfile"
table_name = "svla_so100_stacking_train"

reader = TsFileReader(path)
schema = reader.get_all_table_schemas()[table_name]
columns = [
    column.get_column_name()
    for column in schema.get_columns()
    if column.get_category() in (ColumnCategory.FIELD, ColumnCategory.TAG)
]

batches = []
with reader.query_table(table_name, columns, batch_size=65536) as result_set:
    while True:
        batch = result_set.read_arrow_batch()
        if batch is None:
            break
        if batch.num_rows:
            batches.append(batch)

table = pa.concat_tables(batches)
print(table.to_pandas().head())