pusht_james172610 (TsFile)

Apache TsFile version of james172610/pusht.

Overview

A LeRobot v2.1 manipulation dataset recorded with a unknown robot arm, converted to a single Apache TsFile table.

• Scale: 206 episodes / 25,650 frames at 10 fps (~43 min of robot time) / 1 task(s).

  • (meta/info.json declares 26,030 frames; the frame Parquet files hold 25,650 rows, which is what was converted.)

• Per-frame vectors: observation.state and action carry 2 joint values per frame (flattened to FLOAT columns here).

• Task language: Push the T-shaped block onto the T-shaped target.

• Tables: 1 TsFile table(s), 25,650 rows in total.

  • pusht_james172610: 25,650 rows, Time restarts inside every episode.

Schema (TsFile structure)

pusht_james172610.tsfile

• Time (INT64, milliseconds) — round(timestamp × 1000); the source timestamp column is folded into Time.
• TAG: episode_index, task_index — device dimension; query one device with WHERE episode_index=....
• FIELD (9): frame_index (int64), sample_index (int64), next_reward (float), next_done (int64), next_success (int64), observation_state_0 (float), observation_state_1 (float), action_0 (float), action_1 (float)

Conversion notes

• LeRobot v2.1 robot dataset converted with the project's generic lerobot converter: all episodes of the source train split share one TsFile, with the source columns episode_index and task_index as TAG columns (the TsFile device dimension). Query a single episode with WHERE episode_index=0.
• Vector columns preserve the full source column name when flattened: . becomes _ and the element index is appended, so observation.state becomes observation_state_0..N-1 and action becomes action_0..N-1, all single-precision FLOAT. frame_index is kept and the source index column is kept as sample_index.
• Time = round(timestamp * 1000) ms and restarts inside each episode. The source timestamp column is not stored separately because it equals Time / 1000 seconds.
• Camera video streams are NOT included in this repository; see the original dataset for them: https://huggingface.co/datasets/james172610/pusht/tree/main/videos
• meta/ is mirrored from the source with meta/info.json updated to describe the converted TsFile artifact. No numeric time-series row is dropped.
• The source additionally ships data_t/, meta_s/ and *_t files; data_t/chunk-000/episode_*.parquet is byte-identical to data/ (checked file by file), so only the documented train split is converted and nothing is lost.

Usage

Read this dataset with the Java SDK. It is written with the zstd compressor. The Apache TsFile Java SDK reads every compressor, but the Python SDK (pip install tsfile) implements only UNCOMPRESSED, SNAPPY, GZIP and LZ4; opening this file with it yields no rows.

Read a converted file with the Apache TsFile Java SDK (org.apache.tsfile:tsfile):

import org.apache.tsfile.file.metadata.IDeviceID;
import org.apache.tsfile.read.TsFileReader;
import org.apache.tsfile.read.TsFileSequenceReader;
import org.apache.tsfile.read.common.Path;
import org.apache.tsfile.read.expression.QueryExpression;
import org.apache.tsfile.read.query.dataset.QueryDataSet;
import org.apache.tsfile.write.schema.MeasurementSchema;

import java.util.ArrayList;
import java.util.List;

String path = "data/pusht_james172610.tsfile";
try (TsFileSequenceReader sequenceReader = new TsFileSequenceReader(path);
     TsFileReader reader = new TsFileReader(sequenceReader)) {
    // Devices are enumerated as IDeviceID objects. Tables with TAG columns
    // have multi-segment device ids that cannot be rebuilt from their string
    // form, so always use getAllDevices() rather than getAllDeviceIds().
    for (IDeviceID device : sequenceReader.getAllDevices()) {
        List<Path> paths = new ArrayList<>();
        for (MeasurementSchema schema : reader.getMeasurement(device)) {
            String name = schema.getMeasurementName();
            if (name != null && !name.isEmpty() && !name.equalsIgnoreCase("Time")) {
                paths.add(new Path(device, name, true));
            }
        }
        QueryDataSet dataSet = reader.query(QueryExpression.create(paths, null));
        while (dataSet.hasNext()) {
            System.out.println(dataSet.next());
        }
    }
}

Source & license

• Original dataset: james172610/pusht
• License: mit (as declared by the original dataset).