THULab/kivod_so100_cam2

SO100 Cam2 TsFile

This is an Apache TsFile conversion of kivod/so100_cam2, a LeRobot v2.1 SO100 dataset for “Grasp a cucumber and put it in the hole.”

Modalities: Time-series. Numeric robot state, actions, timing, and episode/task metadata are stored in TsFile. The original dataset has two camera streams; videos are intentionally not copied here.

Source and provenance

• Source revision: e52e39a1adfea61bade108a17a366455bbe542d0
• Original repository owner/uploader: kivod
• License: Apache-2.0; robot: so100; split: train; sampling: 30 fps
• Scale: 52 episodes, 24,655 frame rows, 1 task, 52 source Parquet files, 104 source videos
• Source frame path: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video path: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

The source card provides no separate formal author, paper, or BibTeX citation; repository ownership is the available authorship attribution.

Converted artifact

• TsFile: data/kivod_so100_cam2_train.tsfile
• Table: kivod_so100_cam2_train
• Rows: 24,655; episodes/devices: 52
• Time precision: milliseconds
• meta/ is mirrored from the source. meta/info.json records source/converted paths, row count, mappings, and video policy.

Schema and encoding

Time is round(timestamp * 1000) (INT64 milliseconds), restarting at zero for each episode. Source timestamp is redundant and dropped; frame_index and source index (renamed sample_index) are retained.

TAG columns (TsFile table/device tags):

• episode_index
• task_index

FIELD columns:

• frame_index
• sample_index
• action_0 ... action_5
• observation_state_0 ... observation_state_5

Flattened vectors:

• action -> action_0 ... action_5 (FLOAT)
• observation.state -> observation_state_0 ... observation_state_5 (FLOAT)

Encoding profile: FLOAT/DOUBLE = GORILLA + LZ4; INT32/INT64 and Time = TS_2DIFF + LZ4; TAGs use TsFile device/tag storage. No source rows or numeric fields are dropped other than redundant timestamp.

Videos and frame alignment

Videos remain upstream:

• observation.images.laptop
• observation.images.phone

Each stream has 52 per-episode MP4 files. Join numeric rows to source videos using episode_index and frame_index.

Minimal read example

from tsfile import TsFileReader
reader = TsFileReader("data/kivod_so100_cam2_train.tsfile")
with reader.query_table("kivod_so100_cam2_train", [
    "episode_index", "task_index", "frame_index", "sample_index",
    "action_0", "observation_state_0",
], batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()