Episodes Preview UR5 Visualizer
280 episodes · 20 fps

LIBERO Plus Goal Camera All UR5e Failures TsFile

This repository contains an Apache TsFile conversion of LSY-lab/libero_plus_goal_camera_all_ur5e_failures, a LeRobot v3.0 UR5e manipulation dataset with successful and failed trajectories.

Source and Attribution

  • Publishing organization: LSY-lab / Learning Systems Lab
  • Original uploader and sole contributor shown in the source file history: Maximilian Christof (max-chr)
  • License: Apache-2.0
  • Source split: train (episodes 0:280)
  • Source layout: one Parquet shard at data/chunk-000/file-000.parquet
  • Source size: 10,870,000 bytes, 61,618 rows
  • Robot: UR5e
  • Sampling rate: 20 fps
  • Episodes: 280
  • Tasks: 6
  • Paper/citation: the source dataset card does not provide a dataset-specific paper or BibTeX citation.

Task indexes are: 0 put the wine bottle on the rack; 1 open the middle layer of the drawer; 2 put the bowl on the top of the drawer; 3 open the top layer of the drawer and put the bowl inside; 4 push the plate to the front of the stove; and 5 put the cream cheese on the bowl.

Converted Artifact

  • TsFile: data/lsy_lab_libero_plus_goal_camera_all_ur5e_failures.tsfile
  • Table: lsy_lab_libero_plus_goal_camera_all_ur5e_failures
  • Files/shards: 1
  • Rows: 61,618
  • TAG devices: 280, one for each distinct (episode_index, task_index)
  • Time precision: milliseconds
  • TsFile size: 7,260,077 bytes (66.79% of the 10,870,000-byte source frame Parquet)

Schema

TsFile columns Type Role Source
Time TIMESTAMP/INT64 TIME round(timestamp * 1000) ms
episode_index, task_index STRING TAG Source indexes represented by the TsFile device/tag mechanism
frame_index, sample_index INT64 FIELD frame_index; index renamed to sample_index
observation_state_0..7 FLOAT FIELD observation.state[8]
observation_states_ee_state_0..5 FLOAT FIELD observation.states.ee_state[6]
observation_states_joint_state_0..5 FLOAT FIELD observation.states.joint_state[6]
observation_states_gripper_state_0..1 FLOAT FIELD observation.states.gripper_state[2]
action_0..6 FLOAT FIELD action[7]

Vector columns are flattened in row-major order. Source periods are replaced by underscores and the element index is appended. Rows are sorted by TAG columns and then Time; Time restarts for each episode.

The source timestamp field is dropped only after producing Time because it is redundant with Time / 1000 seconds. The source index field is retained as sample_index. No rows or other numeric time-series fields are dropped.

Encoding and Compression

Physical type Encoding Compression
FLOAT / DOUBLE GORILLA LZ4
INT32 / INT64 TS_2DIFF LZ4
Time TS_2DIFF LZ4
BOOLEAN RLE LZ4
TAG TsFile table device/tag storage TsFile-managed

The source frame schema contains no BOOLEAN FIELD, so the BOOLEAN mapping is documented but not exercised by this artifact.

Videos

Videos are intentionally not included in this TsFile repository. The original 706 MB video tree remains at the source videos/ directory. It contains 280 episode files for each of four streams:

  • videos/observation.images.image/chunk-000/file-*.mp4
  • videos/observation.images.wrist_image/chunk-000/file-*.mp4
  • videos/observation.images.frontview_image/chunk-000/file-*.mp4
  • videos/observation.images.sideview_image/chunk-000/file-*.mp4

The preserved meta/episodes/chunk-000/file-000.parquet records per-episode video chunk/file indexes and timestamp ranges for frame alignment.

Validation

The TsFile was reopened with the Apache TsFile Python SDK. Its table schema contains 34 columns; authoritative TsFile metadata and full query readback both returned 61,618 rows across 280 TAG devices, matching the staged Parquet exactly. Time is monotonic within each (episode_index, task_index) group and Time == round(source_timestamp * 1000).

Usage

from tsfile import ColumnCategory, TsFileReader

path = "data/lsy_lab_libero_plus_goal_camera_all_ur5e_failures.tsfile"
reader = TsFileReader(path)
table_name = "lsy_lab_libero_plus_goal_camera_all_ur5e_failures"
schema = reader.get_all_table_schemas()[table_name]
columns = [
    c.get_column_name()
    for c in schema.get_columns()
    if c.get_category() in (ColumnCategory.FIELD, ColumnCategory.TAG)
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
Downloads last month
68