G1 Dex3 Grasp Square Dataset TsFile
This repository is an Apache TsFile conversion of unitreerobotics/G1_Dex3_GraspSquare_Dataset, a LeRobot v3 robot-manipulation dataset. It contains numeric trajectories and metadata; camera videos remain in the original Hugging Face dataset.
Source dataset
- Original dataset: unitreerobotics/G1_Dex3_GraspSquare_Dataset
- Original uploader/contributors: Henry-Ellis and lv-1; hosted by Unitree Robotics.
- License: Apache-2.0. The source card provides no paper or formal citation.
- Task: stack three 5 cm cubic blocks from bottom to top in red, yellow, blue order on black tape.
- Robot: 7-DOF dual-arm
Unitree_G1with three-finger dexterous hands. - Sampling: 30 Hz; 301 episodes, 281,196 frames, one task, one source Parquet shard.
- Source numeric layout:
data/chunk-000/file-000.parquet.
Converted files
- TsFile:
data/g1_dex3_graspsquare_dataset_train.tsfile - Table:
g1_dex3_graspsquare_dataset_train - Rows: 281,196; devices: 301 (one per
episode_index/task_indexTAG combination). - Time precision: milliseconds;
Time = round(timestamp * 1000)and restarts at the first frame of each episode. - Size: source Parquet 63,247,041 bytes; flattened/staged Parquet 74,725,871 bytes; TsFile 51,443,386 bytes. The TsFile is smaller than both source and staged representations.
TsFile schema
| Role | Columns | Source/type |
|---|---|---|
| TIME | Time |
int64, milliseconds from timestamp |
| TAG/device | episode_index, task_index |
source int64, stored by TsFile device/tag mechanism as string segments |
| FIELD | frame_index, sample_index |
int64; sample_index is renamed from source index |
| FIELD | observation_state_0 … observation_state_27 |
28 FLOAT dimensions from observation.state |
| FIELD | action_0 … action_27 |
28 FLOAT dimensions from action |
Vector names retain their source prefix (. becomes _) and are flattened row-major to scalar fields. frame_index is preserved for video/frame alignment. The redundant source timestamp column is dropped after deriving Time; no numeric rows or vector dimensions are dropped. The four video columns are intentionally omitted from TsFile.
Encoding and compression
| TsFile type | Encoding | Compression |
|---|---|---|
| FLOAT/DOUBLE | GORILLA | LZ4 |
| INT32/INT64 | TS_2DIFF | LZ4 |
| Time | TS_2DIFF | LZ4 |
| BOOLEAN (if present) | RLE | LZ4 |
| TAG | TsFile device/tag mechanism | TsFile-managed |
Videos and frame alignment
Videos are not copied or uploaded with this TsFile dataset. The original repository's videos/ tree is about 9.12 GB and contains four streams:
The source path template is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Numeric rows remain frame-aligned through episode_index and frame_index, with episode metadata under meta/episodes/.
Validation
Local validation read the TsFile with the bundled TsFile Python SDK and compared it with the staged Parquet: 281,196 metadata rows and 281,196 query rows, 301 devices, unique (episode_index, task_index, Time) keys, monotonic Time within each episode, and all 28 action plus 28 observation-state dimensions present. Detailed local reports are validation_report.json and VALIDATION.md.
Usage
from tsfile import TsFileReader
path = "data/g1_dex3_graspsquare_dataset_train.tsfile"
reader = TsFileReader(path)
table = "g1_dex3_graspsquare_dataset_train"
columns = ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"]
with reader.query_table(table, columns, batch_size=65536) as result:
batch = result.read_arrow_batch()
print(batch.to_pandas().head())
- Downloads last month
- 30