THULab/lerobot_unitreeh1_warehouse

Unitree H1 Warehouse TsFile

Apache TsFile form of lerobot/unitreeh1_warehouse, a LeRobot robotics dataset containing Unitree H1 warehouse demonstrations.

Source and Attribution

• Original repository: lerobot/unitreeh1_warehouse
• Publishing organization: LeRobot
• Repository history: initially created by cadene, with later maintenance by aliberts and aractingi
• Formal data-collection authors: not specified by the original dataset card
• License: Apache-2.0
• Paper/citation: the original dataset card does not provide one
• Robot type: unknown in the source metadata; the repository name identifies Unitree H1
• Task: "Grab the spray paint on the shelf and place it in the bin on top of the robot dog."
• Split: train, episodes 0 through 23
• Scale: 24 episodes, 11,275 rows, one task, one source frame Parquet shard, 50 fps

Data and Videos

• TsFile path: data/lerobot_unitreeh1_warehouse.tsfile
• Table: lerobot_unitreeh1_warehouse
• Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
• Source video layout: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
• Original videos: videos/
• Left camera: videos/observation.images.cam_left/chunk-000/file-000.mp4
• Right camera: videos/observation.images.cam_right/chunk-000/file-000.mp4
• Video inventory: two MP4 container files (244,916,943 bytes), with 24 episode-aligned segments in each stream

Videos are not included in this TsFile repository. Use episode_index and frame_index together with the source meta/episodes offsets to align numeric rows with both camera streams in the original Hugging Face repository.

Schema

Column	TsFile role	Type	Description
Time	TIME	INT64	round(timestamp * 1000) milliseconds; restarts at 0 per episode
episode_index	TAG	STRING/device segment	Original episode index
task_index	TAG	STRING/device segment	Original task index; only task 0 is present
frame_index	FIELD	INT64	Frame number within the episode
sample_index	FIELD	INT64	Source index, renamed to avoid ambiguity
next_done	FIELD	BOOLEAN	Source next.done terminal marker
observation_state_0 ... observation_state_18	FIELD	FLOAT	Flattened observation.state[19]
action_0 ... action_39	FIELD	FLOAT	Flattened action[40]

Conversion

• The single source frame Parquet shard is written as one TsFile table.
• Rows are sorted by episode_index, task_index, then Time.
• episode_index and task_index use the TsFile table-model TAG/device mechanism.
• Vector names retain their source prefixes with dots replaced by underscores.
• Source timestamp is dropped after Time synthesis because it is exactly represented by Time / 1000 seconds.
• Source index is retained as sample_index, and next.done is retained as next_done.
• The two video features are omitted from TsFile because their payloads live in separate MP4 files, not in the frame Parquet shard.
• No numeric frame rows, state dimensions, action dimensions, episodes, or tasks are dropped.
• Only JSON metadata is included under meta/; source Parquet metadata files remain in the original repository and are not copied.

Read Example

from tsfile import TsFileReader

path = "data/lerobot_unitreeh1_warehouse.tsfile"
table_name = "lerobot_unitreeh1_warehouse"
reader = TsFileReader(path)
with reader.query_table(
    table_name,
    ["episode_index", "task_index", "Time", "frame_index", "action_0"],
    batch_size=4096,
) as result:
    batch = result.read_arrow_batch()
reader.close()