MT State Recognition 200 TsFile
This repository is a compact Apache TsFile conversion of justintiensmith/MT_State_Recognition_200, a LeRobot v3 state-recognition/manipulation dataset.
Source dataset and attribution
- Original dataset: justintiensmith/MT_State_Recognition_200
- Original author/uploader: Justin Tien-Smith
- License: Apache-2.0
- Task: state-recognition demonstrations that place cups or a glass spice container into a bowl.
- Source card lineage: derived from MT_State_Recognition_200_OLD; 40 checklist-selected episodes were replaced by episodes 0-39 from MT_State_Recognition_40.
- Paper/citation: the source card does not specify a paper or BibTeX citation.
Task index mapping:
0: Put a yellow-white striped cup in the bowl.1: Put a brown cup in the bowl.2: Put a white cup in the bowl.3: Pick up the brown cup and place it into the bowl.4: Pick up a glass spice container and place it into the bowl.5: Place a white cup in the bowl.6: Place a yellow-white striped cup in the bowl.7: Place a brown cup in the bowl.8: Place a glass spice container in the bowl.
Dataset scale and files
- Split:
train(0:200) - Episodes: 200
- Rows/frames: 102,000
- Tasks: 9
- Sampling rate: 30 fps
- Source Parquet shards: 1 (data/chunk-000/file-000.parquet)
- Source row bytes: 4,897,269
- Converted TsFile:
data/justintiensmith_mt_state_recognition_200.tsfile(2,240,355 bytes; 0.457x source Parquet bytes) - Table:
justintiensmith_mt_state_recognition_200
The source video layout is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 with five streams: observation.images.middle, observation.images.above, observation.images.right, observation.images.left, and observation.images.wrist. Videos are not copied into this TsFile repository; use the original Hugging Face videos/ tree. Frame/video alignment is preserved through episode_index, frame_index, and meta/episodes/ offsets.
TsFile schema
Time is round(timestamp * 1000) in milliseconds and restarts at zero for each episode. The source timestamp is dropped because it equals Time / 1000 seconds.
| Column group | TsFile columns | Role/type |
|---|---|---|
| Time | Time |
TIME, INT64 milliseconds |
| Device metadata | episode_index, task_index |
TAG columns (TsFile table/device mechanism) |
| Scalar fields | frame_index, sample_index |
FIELD, INT64; sample_index is source index renamed |
| Action | action_0 ... action_5 |
FIELD, FLOAT32; flattened from action[6] |
| Observation state | observation_state_0 ... observation_state_5 |
FIELD, FLOAT32; flattened from observation.state[6] (dot replaced by underscore) |
No numeric rows or action/state dimensions are dropped. Video tensors are omitted because TsFile output is numeric time-series data.
Encoding and compression
The writer applies the requested compact profile: FLOAT/DOUBLE = GORILLA + LZ4; INT32/INT64 and Time = TS_2DIFF + LZ4; BOOLEAN (if present) = RLE + LZ4; TAG values use TsFile table/device storage. The resulting schema and row count were read back locally and compared with the staged Parquet.
Videos
Videos are not duplicated in this converted repository. The original Hugging Face source retains all five streams under videos/ using the path template above. The numeric rows remain aligned to source video frames using episode_index, frame_index and meta/episodes/ offsets.
Validation
The local validation report is kept outside the upload set under the source dataset's .conversion_workdir/reports/ directory. It records staged/converted row counts, schema, per-episode Time checks, and size comparison. Conversion scripts and reports are intentionally local and are not part of the TsFile upload set.
Minimal read example
from tsfile import TsFileReader
reader = TsFileReader("data/justintiensmith_mt_state_recognition_200.tsfile")
table = reader.get_all_table_schemas()["justintiensmith_mt_state_recognition_200"]
fields = [c.get_column_name() for c in table.get_columns() if c.get_column_name() != "Time"]
with reader.query_table("justintiensmith_mt_state_recognition_200", fields, batch_size=65536) as result:
batch = result.read_arrow_batch()
- Downloads last month
- 23