release5_i_top SAM EVT2 (TsFile)

Apache TsFile version of 1g0rrr/release5_i_top, a LeRobot v2.1 robot trajectory dataset. Its recorded task is "Insert connector into the pcb."

Overview

• Robot type: sam_evt2
• Sampling rate: 30 fps
• Split: train, episodes 0:50
• Scale: 50 episodes and 65,700 frames / TsFile rows
• Task: Insert connector into the pcb
• Video coverage upstream: 200 source videos across front, side, left-wrist, and right-wrist camera streams
• Converted artifact: one 6,011,523-byte TsFile

Schema (TsFile structure)

• Time (INT64, milliseconds) - Time = round(timestamp * 1000) and restarts within each episode.
• episode_index, task_index (TAG) - original LeRobot identifiers; query one trajectory with WHERE episode_index = 0.
• frame_index, sample_index (FIELD, INT64) - source frame position and source index renamed to sample_index.
• action_0..13 (FIELD, FLOAT) - the 14-dimensional bimanual robot action.
• observation_state_0..13 (FIELD, FLOAT) - the 14-dimensional bimanual robot state.

All 50 source Parquet files are merged into data/release5_i_top_train.tsfile. The redundant source timestamp column is omitted after conversion to Time; no numeric trajectory rows or vector elements are dropped. The four camera streams are not uploaded here. The 200 MP4 files remain in the original dataset videos tree.

Usage

Read the .tsfile file with the Apache TsFile Python SDK:

from tsfile import TsFileReader

path = "data/release5_i_top_train.tsfile"
table = "release5_i_top_train"

reader = TsFileReader(path)
with reader.query_table(
    table,
    ["episode_index", "task_index", "frame_index", "action_0", "observation_state_0"],
    batch_size=1024,
) as result:
    batch = result.read_arrow_batch()
    if batch is not None:
        print(batch.to_pandas().head())

Source & license

• Original dataset: https://huggingface.co/datasets/1g0rrr/release5_i_top
• Author / publisher: 1g0rrr
• License: Apache-2.0