THULab/lerobot_cmu_stretch

CMU Stretch TsFile

TsFile conversion of lerobot/cmu_stretch, a LeRobot v3.0 robotics dataset licensed under MIT.

Source and attribution

• Contributors visible in the source repository: LeRobot, aractingi, and aliberts
• Task: CMU Stretch robot manipulation demonstrations
• Paper: Affordances from Human Videos as a Versatile Representation for Robotics
• Split: train (0:135), 135 episodes, 25,016 frames, 5 tasks, 5 fps
• Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet

Task labels from the source task metadata:

• 0: open drawer
• 1: lift up a lid from the pot
• 2: lift open green garbage can lid
• 3: pull open a dishwasher
• 4: open door

TsFile schema

Time = round(timestamp * 1000) milliseconds, restarting at zero per episode. The redundant source timestamp column is dropped. index becomes sample_index; vector fields preserve their names with dots replaced by _.

Column	Type	Role
episode_index	STRING	TAG
task_index	STRING	TAG
frame_index	INT64	FIELD
sample_index	INT64	FIELD
next_reward	FLOAT	FIELD
next_done	BOOLEAN	FIELD
observation_state_0	FLOAT	FIELD
observation_state_1	FLOAT	FIELD
observation_state_2	FLOAT	FIELD
observation_state_3	FLOAT	FIELD
action_0	FLOAT	FIELD
action_1	FLOAT	FIELD
action_2	FLOAT	FIELD
action_3	FLOAT	FIELD
action_4	FLOAT	FIELD
action_5	FLOAT	FIELD
action_6	FLOAT	FIELD
action_7	FLOAT	FIELD

TAG columns are episode_index and task_index. observation.state[4] is flattened to observation_state_0..3; action[8] to action_0..7. next.done is retained as BOOLEAN next_done; next.reward becomes next_reward.

Storage profile

• FLOAT/DOUBLE: GORILLA + ZSTD
• INT32/INT64: TS_2DIFF + ZSTD
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4
• TAG: native TsFile table/device mechanism

The source Parquet is 655,555 bytes; the TsFile is 629,392 bytes.

Videos

Source videos are under videos/, specifically videos/observation.images.image/ with path template videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Videos are not included in this TsFile repository. Frame alignment uses episode_index and frame_index.

Read example

from tsfile import TsFileReader
reader = TsFileReader("data/lerobot_cmu_stretch.tsfile")
with reader.query_table("lerobot_cmu_stretch", ["episode_index", "task_index", "Time", "frame_index", "next_done"], batch_size=4096) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()