Datasets:
SO100 Test 5 TsFile
This Apache TsFile dataset converts the numeric trajectories from
arulloomba/so100_test_5, a LeRobot v2.1 SO100 dataset
for the task "Grasp a block and put it in the bin."
Source Dataset
- Repository owner and uploader:
arulloomba - License: Apache-2.0
- Robot:
so100; LeRobot codebase version:v2.1 - Split:
train; sampling rate: 30 fps - Numeric data: 50 episode Parquet files, 37,316 rows total (746 or 747 rows per episode)
- Task count: 1; task index
0: "Grasp a block and put it in the bin." - Source frame layout:
data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet - Source video layout:
videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4 - Paper/citation: the source card supplies neither a paper nor a completed citation
TsFile Data
- File:
data/arulloomba_so100_test_5_train.tsfile - Table:
arulloomba_so100_test_5_train - Rows/devices: 37,316 rows across 50 episode/task TAG combinations
- Size: 0.29 MiB
- Time:
round(timestamp * 1000)milliseconds, ranging from 0 to 24,867 ms
Schema roles:
| Source | TsFile columns | Role | Type |
|---|---|---|---|
timestamp |
Time |
TIME | INT64 milliseconds |
episode_index |
episode_index |
TAG | STRING device segment |
task_index |
task_index |
TAG | STRING device segment |
frame_index |
frame_index |
FIELD | INT64 |
index |
sample_index |
FIELD | INT64 |
action[6] |
action_0 ... action_5 |
FIELD | FLOAT |
observation.state[6] |
observation_state_0 ... observation_state_5 |
FIELD | FLOAT |
Vector names retain the source prefix with . replaced by _. The six action
and state dimensions are, in order, main_shoulder_pan, main_shoulder_lift,
main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.
The source timestamp field is omitted after Time synthesis because it equals
Time / 1000; every numeric row and every other numeric dimension is retained.
Encoding uses GORILLA + LZ4 for FLOAT/DOUBLE, TS_2DIFF + LZ4 for INT32/INT64
and Time, and RLE + LZ4 for BOOLEAN fields. The dataset has no BOOLEAN field.
episode_index and task_index use the TsFile table-model TAG/device mechanism.
Videos
Videos are not included here. They remain in the source repository as 100 MP4 files totaling 450,731,429 bytes, split into two 50-file camera streams:
The upstream spelling detatched_arm is preserved. The numeric rows align to
each episode video through episode_index and
frame_index.
Read Example
from tsfile import TsFileReader
reader = TsFileReader("data/arulloomba_so100_test_5_train.tsfile")
columns = ["episode_index", "task_index", "frame_index", "action_0", "observation_state_0"]
with reader.query_table("arulloomba_so100_test_5_train", columns, batch_size=65536) as result:
print(result.read_arrow_batch().to_pandas().head())
reader.close()
Attribution
The source card identifies the Hugging Face repository owner/uploader as
arulloomba and provides no separate formal author, paper, or citation. Cite
the original dataset and its owner when using this conversion.
- Downloads last month
- 39