README.md

metadata

license: apache-2.0
authors:
  - Henry-Ellis
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
  - robotics
  - lerobot
modality:
  - timeseries
  - tabular
pretty_name: G1 Dex3 Object Placement Dataset TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/g1_dex3_objectplacement_dataset_train.tsfile
size_categories:
  - 10K<n<100K

G1 Dex3 Object Placement Dataset TsFile

Apache TsFile conversion of unitreerobotics/G1_Dex3_ObjectPlacement_Dataset, a LeRobot v3 robot-manipulation dataset.

Source and attribution

• Original dataset: https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset
• Original uploader/data author shown in the Hugging Face repository history: Henry-Ellis (https://huggingface.co/Henry-Ellis)
• Repository owner/organization: unitreerobotics (Unitree Robotics)
• License: Apache-2.0; no paper or formal citation is supplied by the source card.
• Task: pick up toothpaste and a trash bag and place them into the blue storage container.
• Robot: 7-DOF dual-arm Unitree_G1 with three-fingered dexterous hands.
• Recording frequency: 30 Hz; source frame resolution 640x480.

Source layout and videos

The source train split is one Parquet shard (data/chunk-000/file-000.parquet) with 210 episodes, 98,266 frames, one task, and 30 fps. Source metadata paths:

• Numeric data: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
• Videos: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
• Four camera streams: observation.images.cam_left_high, observation.images.cam_right_high, observation.images.cam_left_wrist, and observation.images.cam_right_wrist. They are listed in the original repository videos tree at https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset/tree/main/videos.
• Current source revision contains 9 physical MP4 shards and 840 logical episode videos.

Videos are not included in this TsFile repository. They remain at the original Hugging Face videos paths; episode_index and frame_index preserve frame alignment.

Converted artifact

• TsFile: data/g1_dex3_objectplacement_dataset_train.tsfile
• Table: g1_dex3_objectplacement_dataset_train
• Rows: 98,266; episodes: 210; source split: train
• Time precision: milliseconds; Time = round(timestamp * 1000) and restarts from zero for each episode.
• Original timestamp is dropped because it is redundant with Time / 1000; source index is renamed to sample_index.

Schema

Role	Columns	Source / notes
TIME	Time	INT64, milliseconds, TS_2DIFF + LZ4
TAG	episode_index, task_index	Original source columns; TsFile stores TAG values as strings while source_dtype: int64 is recorded in meta/info.json
FIELD	frame_index, sample_index	INT64, TS_2DIFF + LZ4; sample_index comes from source index
FIELD	observation_state_0 ... observation_state_27	Flattened source observation.state[28], FLOAT, GORILLA + LZ4
FIELD	action_0 ... action_27	Flattened source action[28], FLOAT, GORILLA + LZ4

All source numeric rows and 28-dimensional state/action elements are retained. The only dropped source column is timestamp; video/image columns are intentionally omitted because videos remain in the original dataset.

Conversion and validation

The dataset-specific converter in conversion_support/scripts/converters/unitreerobotics_g1_dex3_objectplacement.py uses LeRobot normalization rules, sorts by episode_index, task_index, then Time, and writes one merged TsFile for the train split. Encodings/compression are explicit: FLOAT/DOUBLE GORILLA, INT32/INT64/Time TS_2DIFF, BOOLEAN RLE (no boolean field is present), and LZ4 compression.

Local validation passed: source/staged/TsFile row counts are all 98,266; 210 TAG devices are present; Time exactly matches round(timestamp * 1000); vector dimensions round-trip exactly; and the TsFile is non-empty and readable. See VALIDATION.md and validation_report.json.

Minimal read example

from tsfile import TsFileReader
path = "data/g1_dex3_objectplacement_dataset_train.tsfile"
reader = TsFileReader(path)
columns = ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"]
with reader.query_table("g1_dex3_objectplacement_dataset_train", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())