codebase_version string | robot_type string | total_episodes int64 | total_frames int64 | total_tasks int64 | chunks_size int64 | data_files_size_in_mb int64 | video_files_size_in_mb int64 | fps int64 | splits dict | data_path string | features dict | video_path_original string | tsfile_conversion dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
v3.0 | panda | 5,000 | 575,101 | 128 | 1,000 | 100 | 500 | 10 | {
"train": "0:5000"
} | data/vlabench_primitive_ft_lerobot_video.tsfile | {
"Time": {
"dtype": "int64",
"shape": [
1
],
"tsfile_role": "TIME",
"unit": "ms"
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"tsfile_role": "TAG"
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"tsfile_role": "TAG"
},
... | videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 | {
"source_dataset": "VLABench/vlabench_primitive_ft_lerobot_video",
"source_data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"converted_data_path": "data/vlabench_primitive_ft_lerobot_video.tsfile",
"table_name": "vlabench_primitive_ft_lerobot_video",
"granularity": "merged",
"time_... |
VLABench Primitive Tasks — LeRobot v3.0 (TsFile)
Apache TsFile version of VLABench/vlabench_primitive_ft_lerobot_video.
Overview
This dataset is organized in the LeRobot v3.0 format and is used for integrating VLABench into the LeRobot framework officially. Compared with the v2.0 and the RLDS versions, this release stores the visual observations in a video-compressed format rather than as individual image files, giving better storage efficiency and data-loading performance.
It contains demonstrations for 10 primitive manipulation tasks on a Franka Panda arm, collected across 128 concrete language-instructed task variants (e.g. "Pick up the mahjong of 1 pin", "Insert the rose into the vase"). Models trained on this data can be evaluated on multiple generalization tracks in VLABench.
The 10 primitive task categories are: add_condiment, insert_flower, select_book, select_chemistry_tube, select_drink, select_fruit, select_mahjong, select_painting, select_poker, select_toy.
- Robot: Franka Panda, 7-DoF.
- Episodes: 5,000 (episode_index 0–4999).
- Frames: 575,101 time-series rows.
- Tasks: 128 language-instructed variants (task_index 0–127); the full
task_index → instructionmapping is inmeta/tasks.parquet. - Sampling rate: 10 fps (from
meta/info.json).
Baseline models: pi05-primitive-10task, pi0-fast-primitive-10task, pi0-primitive-10task.
Schema (TsFile structure)
All 5,000 episodes are stored in a single TsFile table; episode_index and task_index are TAG columns (the TsFile device dimension), so one episode is selected with WHERE episode_index=0.
- Time (INT64, milliseconds) —
round(timestamp * 1000); restarts at 0 for each episode, stepping by 100 ms (10 fps). - episode_index (TAG) — source episode id, 0–4999.
- task_index (TAG) — source task id, 0–127; resolve to the instruction text via
meta/tasks.parquet. - frame_index (FIELD, INT64) — frame number within the episode.
- sample_index (FIELD, INT64) — the source global
indexcolumn, renamed. - state_0 … state_6 (FIELD, FLOAT) — robot proprioceptive state, flattened from the 7-element
statevector. - actions_0 … actions_6 (FIELD, FLOAT) — action command, flattened from the 7-element
actionsvector.
The source timestamp column is dropped because it equals Time / 1000 seconds. No other columns or rows are dropped.
Videos
The three camera video streams of the original dataset (image, second_image, wrist_image) are NOT included in this repository. They are large and non–time-series; obtain them from the original dataset:
https://huggingface.co/datasets/VLABench/vlabench_primitive_ft_lerobot_video (the videos/ directory).
Usage
Read the .tsfile file with the Apache TsFile Java or Python SDK.
Source & license
- Original dataset: https://huggingface.co/datasets/VLABench/vlabench_primitive_ft_lerobot_video
- Author / publisher: VLABench
- License: MIT (as declared by the original dataset).
- Downloads last month
- 75