THULab/sponge_marker_merged_video

Sponge Marker Merged TsFile

This dataset is an Apache TsFile conversion of mkpongm/sponge_marker_merged_video, a LeRobot v2.1 SO follower robot-manipulation dataset containing sponge-placement and marker-placement trajectories.

Modalities: Time-series. The converted repository contains numeric robot state, action, frame timing, episode/task tags, and mirrored source metadata. The camera videos remain in the original Hugging Face dataset.

Source Dataset and Author

• Original dataset: mkpongm/sponge_marker_merged_video
• Pinned source revision: ab00f90ddda51d18ec83df836fb8d9cb808834e5
• Original author, repository owner, and uploader: Mboutidem Mkpong (mkpongm)
• License: Apache-2.0
• Robot type: so_follower
• LeRobot codebase version: v2.1
• Sampling rate: 30 fps
• Split: train, episodes 0 through 99
• Source scale: 100 episodes, 78,522 frames, 2 tasks
• Source frame layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source frame shards: 100
• Paper, homepage, and completed citation: not provided by the source card

The source repository commit describes this snapshot as the H.264/decord video version of the merged sponge-and-marker dataset. The source README embeds an older meta/info.json excerpt that says total_videos: 0; the actual source meta/info.json and repository tree contain 200 videos, with 100 files in each of two streams.

Tasks and Scale

task_index	Task	Episodes	Rows
0	grab the sponge and place it in the bowl	50	36,024
1	pick up the marker on the table and place it vertically into the cup	50	42,498

The converted train split contains 78,522 rows across 100 episode devices.

Converted Files

• TsFile: data/sponge_marker_merged_train.tsfile
• Table: sponge_marker_merged_train
• Rows: 78,522
• TsFile size: 1.33 MiB (1,391,553 bytes)
• FIELD compression: LZ4
• FIELD encodings: TS_2DIFF for INT64 and GORILLA for FLOAT
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten to describe the TsFile schema, source revision, conversion mapping, author, and video policy.
• Configuration: sponge_marker_merged.yaml

TsFile Schema

Time is an INT64 millisecond timestamp computed as round(timestamp * 1000). It restarts within each episode.

This is relative elapsed time, not an absolute wall-clock timestamp. The source timestamp column is present, so Time is not synthesized for missing data. A timestamp viewer may render Time=0 as 1970-01-01 00:00:00 in UTC or 1970-01-01 08:00:00 in UTC+8; that display is the Unix epoch plus the viewer's time-zone offset, not the dataset capture date.

TAG columns:

• episode_index — source episode number, 0 through 99
• task_index — source task number, 0 or 1

Scalar FIELD columns:

• frame_index — frame number within an episode
• sample_index — the source global index column, renamed

Flattened FLOAT FIELD groups:

• observation.state[6] -> observation_state_0 ... observation_state_5
• action[6] -> action_0 ... action_5

The table has 17 columns including Time: 2 TAG columns, 14 FIELD columns, and the TIME column.

Conversion Notes

• The shared config-driven lerobot converter is used; the dataset-specific Python file is a thin local orchestration and documentation entry point.
• All 100 source episode Parquet files in the train split are merged into one table-model TsFile. Filter by episode_index and task_index to select a trajectory or task.
• Vector columns are fully flattened into scalar fields. Full source prefixes are preserved and . is replaced with _.
• The source timestamp column is the only numeric source column not stored separately. It is redundant after the millisecond Time mapping (timestamp = Time / 1000 seconds at the source's 30 Hz cadence).
• The source index column is retained as sample_index; frame_index is retained unchanged.
• No source frames, episodes, tasks, state dimensions, or action dimensions are intentionally dropped.
• This local candidate fixes the Apache TsFile 2.2.1 import tool's schema-mode default of PLAIN + UNCOMPRESSED. TAG/device construction and Time mapping are unchanged; only FIELD encoding and compression are replaced.

Videos

Videos are not duplicated in the converted repository. The pinned source has 200 frame-aligned H.264 MP4 files in two streams:

• videos/chunk-000/observation.images.front/ — 100 MP4 files
• videos/chunk-000/observation.images.handeye/ — 100 MP4 files

The source template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Each stream contains one file for each of the 100 episodes. Numeric TsFile rows remain aligned with the original videos through episode_index, frame_index, and the source episode metadata under meta/.

Validation

The generated TsFile was queried through the Java SDK with every TIME, TAG, and FIELD value selected. It contains 78,522 rows and no duplicate (episode_index, task_index, Time) keys. Its full-content fingerprint is 452abb22e95739a0, identical to the original 5,322,542-byte TsFile and the ZSTD candidate.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/sponge_marker_merged_train.tsfile")
print(reader.get_all_table_schemas()["sponge_marker_merged_train"])
reader.close()

Source and License

The source dataset is published by Mboutidem Mkpong (mkpongm) under the Apache License 2.0 and was created with LeRobot. The source card does not provide a paper or completed citation, so none is inferred here.