rollout_pi05_reasoning_step_076596_counting

rollout pi0.5 Reasoning Step 076596 Counting (TsFile)

Apache TsFile version of justintiensmith/rollout_pi05_Reasoning_Step_076596_Counting.

Overview

A SO follower rollout dataset from a pi0.5 policy on a counting task family. The 32 language instructions ask the robot to move objects (a blue pen, a green pencil sharpener, a yellow block and an airplane toy) into or out of a bowl so that the bowl ends up holding a requested number of objects. Each frame records the 6-D joint state and the 6-D commanded action.

• Scale: 150 episodes, 254,544 frames, 32 tasks, sampled at 30 fps; one train split in data/rollout_pi05_reasoning_step_076596_counting.tsfile.

Tasks (task_index 0-31) include instructions such as "Leave exactly one object in the bowl", "Adjust the bowl so that it contains two objects", "Leave exactly three objects in the bowl", "Adjust the bowl so that it contains zero objects", "Move the green pencil sharpener into the bowl", and "Move the yellow block from the bowl to the table" (full list in meta/tasks.parquet).

Schema (TsFile structure)

• Time (INT64, milliseconds) - per-episode sample time, round(timestamp * 1000) in milliseconds (the source timestamp column equals Time / 1000 seconds and is therefore not stored separately).
• episode_index, task_index (TAG) - device dimensions identifying one demonstration episode and one task; query a single episode with WHERE episode_index=0.
• action_0..5, observation_state_0..5 (FIELD) - FLOAT 6-D SO-follower joint action and 6-D observation state.
• frame_index, sample_index (INT64, FIELD) - source frame counter and source row counter (index renamed to sample_index).

The dataset is stored as one wide TsFile table; within the train split all episodes share that single file with episode_index and task_index as TAG columns. State/action vectors are flattened to one FLOAT field per element (. -> _, element index appended). Camera video files are not included in this repository; the original dataset's videos are at videos/. meta/ is mirrored from the source. No data rows or non-video columns are dropped, apart from the redundant source timestamp column (replaced by Time). Two camera image/video features (observation.images.wrist, observation.images.middle) are omitted as non-time-series.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("data/rollout_pi05_reasoning_step_076596_counting.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], 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/justintiensmith/rollout_pi05_Reasoning_Step_076596_Counting
• Author / publisher: justintiensmith
• License: apache-2.0