Datasetpublic

THULab/g1_dex3_object_placement_dataset

G1 Dex3 Object Placement Dataset (TsFile) Converted from unitreerobotics/G1_Dex3_ObjectPlacement_Dataset at pinned revision 2e1d63e5fd9587e1240b889c51c6553be6910027. Modalities: Time-series. Dataset description This Unitree G1 dataset records dual-arm, three-fingered-hand demonstrations for picking up a toothpaste box and trash bag from a desk and placing them into a blue storage container. The source card documents approximately 20-second operations, 30 Hz… See the full description on the dataset page: https://huggingface.co/datasets/THULab/g1_dex3_object_placement_dataset.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes34downloads

Dataset Card

G1 Dex3 Object Placement Dataset (TsFile)

Converted from `unitreerobotics/G1_Dex3_ObjectPlacement_Dataset` at pinned revision `2e1d63e5fd9587e1240b889c51c6553be6910027`. Modalities: Time-series.

Dataset description

This Unitree G1 dataset records dual-arm, three-fingered-hand demonstrations for picking up a toothpaste box and trash bag from a desk and placing them into a blue storage container. The source card documents approximately 20-second operations, 30 Hz recording, and randomized object placement.

Source task label: object placement. The source card gives the objective above and links the AVP teleoperation documentation for camera/setup guidance.

—Source repository owner/publisher: unitreerobotics
—License: apache-2.0 (declared by the source card).
—Paper/homepage/citation: no completed paper, homepage, or citation is documented in the pinned source card unless linked above.

Dataset Scale

Split	Episodes	Tasks	Source trajectory Parquets	TsFile rows	Sampling rate	TsFile files/shards
`train`	210	1	1	98,266	30 Hz	1

The staged Parquet has 98,266 rows and 61 columns including Time; TsFile chunk metadata independently reports the same 98,266 rows.

TsFile schema

Column	Role	TsFile type	Observed/source range
`Time`	TIME	INT64	0–31,967 ms; restarts per episode
`episode_index`	TAG	STRING	source integer 0–209
`task_index`	TAG	STRING	source integer 0–0
`frame_index`	FIELD	INT64	0–959 within an episode
`sample_index`	FIELD	INT64	0–98,265 globally

Exact remaining FIELD names/ranges and imported types:

—observation_state_0–observation_state_27 (FLOAT): 14 arm joints and 14 Dex3 hand coordinates
—action_0–action_27 (FLOAT): the corresponding next-action coordinates

Conversion

—All source episodes in the train split are merged into data/g1_dex3_object_placement_dataset_train.tsfile; episode_index and task_index are TAG dimensions.
—Time = round(timestamp * 1000) in milliseconds. The source timestamp column is omitted because it is redundant with Time / 1000 seconds.
—frame_index is retained. Source index is retained as sample_index.
—Every vector is fully flattened: the complete source name is kept, . becomes _, and element indices are appended. Float vectors are imported as single-precision FLOAT fields.
—Other scalar source columns shown above are retained; no trajectory rows are intentionally dropped.
—Source metadata is mirrored for publication, with copied meta/info.json rewritten to describe the converted data path and conversion semantics.

Video policy

The source declares four 640×480 video features (left/right head cameras and left/right wrist cameras) but no current total_videos counter. MP4 files were not downloaded or uploaded; they remain in the pinned source `videos/` tree.

Minimal read example

python

from tsfile import TsFileReader

reader = TsFileReader("data/g1_dex3_object_placement_dataset_train.tsfile")
print(reader.get_all_table_schemas().keys())
reader.close()

Source and provenance

—Source dataset: `unitreerobotics/G1_Dex3_ObjectPlacement_Dataset`
—Pinned source revision: `2e1d63e5fd9587e1240b889c51c6553be6910027`
—Converted artifact: data/g1_dex3_object_placement_dataset_train.tsfile