Datasets:

THULab
/

first_humanoid_dataset

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - humanoid
  - robotics
  - tsfile
  - modality:timeseries
  - timeseries
  - format:tsfile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/first_humanoid_dataset.tsfile
modality:
  - tabular
  - text
  - timeseries
size_categories:
  - 1K<n<10K

first_humanoid_dataset

Converted Apache TsFile version of skyzos/first-humanoid-dataset, a LeRobot v2.1 humanoid manipulation dataset.

Source Dataset

Original dataset: skyzos/first-humanoid-dataset
License: apache-2.0
Task: pick the block with the right arm and put it in the middle, then use the left arm and put it in the middle.
Source split: train
Source scale from meta/info.json: 19 episodes, 2,285 frames, 1 task, 57 videos, 30 fps
Robot type: not specified in the source meta/info.json
Modalities: Time-series. The source repository also contains videos, but this converted repository stores only numeric robot state/action time-series data and metadata.

TsFile Layout

Data file: data/first_humanoid_dataset.tsfile
Table name: first_humanoid_dataset
Rows: 2,285
Time precision: milliseconds
TAG columns: episode_index, task_index
FIELD columns: frame_index, sample_index, observation_state_0 through observation_state_31, and action_0 through action_13

The converted table uses the original LeRobot episode_index and task_index columns as TsFile TAG columns. Query one episode by filtering on episode_index.

Conversion Notes

Converted with the generic lerobot converter in the huggingface_to_tsfile pipeline.
Time = round(timestamp * 1000) in milliseconds and restarts per episode.
The source timestamp column is not retained as a separate FIELD because it is redundant with Time / 1000.
The source index column is renamed to sample_index.
Vector columns are flattened while preserving the full source column name with . replaced by _: observation.state[32] becomes observation_state_0..observation_state_31, and action[14] becomes action_0..action_13.
Numeric arrays are stored as single-precision FLOAT fields.
Source videos are not uploaded here. They remain in the original dataset under videos/.
The source repository also contains unrelated nvim-linux-x86_64/ files; those files are intentionally excluded from this time-series conversion.
meta/ is mirrored from the source except that meta/info.json describes the converted TsFile artifact and records the original video path separately.

Read Example

Use Apache TsFile tooling or SDKs to read:

data/first_humanoid_dataset.tsfile