franka-lift TsFile

This dataset is a TsFile conversion of the Hugging Face dataset purewater-TRI/franka-lift, which was created with LeRobot.

Modalities: Time-series. The source robot dataset also contains videos, but this repository stores only the numeric time-series data and metadata. Source videos remain available in the original dataset under videos/.

Source Dataset

• Original dataset: purewater-TRI/franka-lift
• License: apache-2.0
• Task: FrankaLiftEnv
• Robot type: isaac_sim
• Episodes: 500 total
• Frames: 25,785 total
• Splits: train episodes 0..449, test episodes 450..499
• Sampling rate: 50 fps
• Source video streams: observation.images.image_front, observation.images.image_side
• Source video count: 1,000

Converted Files

The source train/test split is preserved:

data/franka_lift_train.tsfile
data/franka_lift_test.tsfile

The source meta/ files are mirrored in this repository. meta/info.json is updated so data_path describes the converted TsFile layout, and the tsfile_conversion object records source paths, output paths, split ranges, feature mapping, row counts, and the video policy.

Schema

The TsFile table names are franka_lift_train and franka_lift_test.

• Time: integer millisecond timestamp, computed as round(timestamp * 1000).
• TAG columns: episode_index, task_index.
• FIELD columns: frame_index, sample_index, next_done, observation_state_0 through observation_state_8, and action_0 through action_7.

The source vector columns are flattened into scalar fields:

• observation.state with shape [9] becomes observation_state_0 ... observation_state_8.
• action with shape [8] becomes action_0 ... action_7.

All flattened vector values are stored as single-precision FLOAT fields. episode_index and task_index are declared as TAG columns, so a single episode can be selected with a predicate such as WHERE episode_index = 450.

Conversion Notes

• A dataset-specific converter is used so the train/test episode ranges declared in meta/info.json are preserved as separate TsFiles.
• Train split: 450 episodes, 23,243 rows.
• Test split: 50 episodes, 2,542 rows.
• Total converted rows: 25,785.
• Time restarts per episode, matching the source timestamp values.
• The source timestamp column is not retained as a separate FIELD because it is represented by Time / 1000 seconds.
• The source index column is renamed to sample_index.
• The source next.done column is renamed to next_done.
• Video features are not uploaded here. They remain in the original Hugging Face dataset and are referenced by the preserved metadata.

Validation after conversion:

• franka_lift_train.tsfile: 23,243 TsFile metadata rows, matching staged Parquet.
• franka_lift_test.tsfile: 2,542 TsFile metadata rows, matching staged Parquet.
• TsFile file count: 2.

Read Example

from tsfile import TsFileReader

path = "data/franka_lift_train.tsfile"
table = "franka_lift_train"

with TsFileReader(path) as reader:
    columns = [
        "episode_index",
        "task_index",
        "frame_index",
        "observation_state_0",
        "action_0",
    ]
    with reader.query_table(table, columns, batch_size=4096) as result:
        batch = result.read_arrow_batch()
        if batch is not None:
            print(batch.to_pandas().head())