THULab/adityx23_pnp_act_v4

PnP ACT v4 TsFile

Apache TsFile conversion of adityx23/pnp_act_v4, a LeRobot v3.0 30 Hz pick-and-place ACT dataset.

Modalities: Time-series. Numeric state/action trajectories, frame timing, and episode/task TAG metadata are included. Camera videos remain in the original Hugging Face dataset and are not included in this repository.

Source

• Author/uploader: adityx23
• License: Apache-2.0
• Revision: cf5d5843f37f09e137949b983c0dc4a8f626193b
• Split: train; 1 Parquet shard; 465 episodes; 114,278 rows; 30 fps
• task_index contains only 0; no textual task label is supplied by the source metadata.
• Source frames: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet

Converted Dataset

• TsFile: data/pnp_act_v4.tsfile
• Table: pnp_act_v4
• Rows: 114,278
• Devices/episodes: 465
• Time precision: milliseconds

Schema and Conversion

Time = round(timestamp * 1000) milliseconds and restarts at zero per episode. The source timestamp is dropped because it is redundant with Time / 1000 seconds. Source index is renamed to sample_index.

TAG columns use the TsFile table/device mechanism:

• episode_index
• task_index

FIELD columns:

• frame_index, sample_index
• observation.state[7] flattened to observation_state_0 ... observation_state_6
• action[8] flattened to action_0 ... action_7

Codec profile: FLOAT/DOUBLE GORILLA + LZ4; INT32/INT64 TS_2DIFF + LZ4; Time TS_2DIFF + LZ4; BOOLEAN RLE + LZ4 when present (none occur in this source). No numeric rows or vector dimensions are dropped; video features are omitted from TsFile.

Videos

Videos are not uploaded here. The original dataset contains 465 MP4 files for each stream:

• observation.images.zed
• observation.images.fish0

Use episode_index and frame_index to align rows with the original video files and meta/episodes records.

Validation

Source Parquet: 3,592,037 bytes and 114,278 rows. Converted TsFile: 4,314,182 bytes and 114,278 metadata rows. The TsFile contains 465 devices, 2 TAG columns, and 15 FIELD columns. The size ratio is approximately 1.2003; this is documented for review before further compression work.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/pnp_act_v4.tsfile")
columns = ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"]
with reader.query_table("pnp_act_v4", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())