record_test_3 / README.md
zjt24's picture
chore: sync metadata, drop original files
c3b7a7e verified
|
Raw
History Blame Contribute Delete
2.59 kB
---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- timeseries
- tsfile
- format:tsfile
modality:
- tabular
- timeseries
pretty_name: record-test_3 TsFile
configs:
- config_name: default
data_files:
- split: train
path: data/record_test_3_train.tsfile
size_categories:
- 10K<n<100K
---
# record-test_3 TsFile
This dataset is a TsFile conversion of the LeRobot dataset
[`sunparticle/record-test_3`](https://huggingface.co/datasets/sunparticle/record-test_3).
The source dataset was created with [LeRobot](https://github.com/huggingface/lerobot)
and is licensed under `apache-2.0`.
## Dataset Details
- Source dataset: [`sunparticle/record-test_3`](https://huggingface.co/datasets/sunparticle/record-test_3)
- Source format: LeRobot `v2.1`
- Robot type: `so101_follower`
- Modalities: Time-series robot state and action data
- Source task: `place crown on capybara`
- Split: `train`
- Episodes: `40`
- Frames / rows: `17403`
- Tasks: `1`
- Sampling rate: `30 fps`
- Source videos: `40`
- Converted data file: `data/record_test_3_train.tsfile`
## TsFile Layout
The conversion writes one TsFile table for the train split:
`record_test_3_train`.
- `Time` is the TsFile time column in milliseconds.
- `episode_index` and `task_index` are TAG columns.
- `frame_index` and `sample_index` are scalar FIELD columns.
- `action_0` through `action_5` are flattened action FIELD columns.
- `observation_state_0` through `observation_state_5` are flattened observation FIELD columns.
Vector column flattening preserves the source column name, replaces `.` with
`_`, and appends the element index. For example, `observation.state` becomes
`observation_state_0` through `observation_state_5`.
## Conversion Notes
- `Time = round(timestamp * 1000)` and restarts per episode.
- The source `timestamp` column is dropped because it is represented by `Time / 1000` seconds.
- The source `index` column is renamed to `sample_index`.
- No numeric time-series rows are dropped; the converted TsFile has `17403` rows.
- Source `meta/` files are mirrored, with `meta/info.json` updated to describe the converted TsFile artifact.
- Videos are not uploaded here. The source video feature `observation.images.front`
and the `videos/` tree remain in the original dataset:
<https://huggingface.co/datasets/sunparticle/record-test_3/tree/main/videos>.
## Read Example
Use the Apache TsFile SDK or TsFile tools to read:
```text
data/record_test_3_train.tsfile
```
For episode-level filtering, query the table with the TAG column, for example
`WHERE episode_index = 0`.