Datasets:

kalanosai
/

synthetic-missing-task-labels

Episodes 0 and 1 have empty task instructions, producing 400 rows with empty labels. The other 1,600 rows have normal task labels. This is intentional missing instruction content, not an invalid task-index mapping. Ground truth records the affected episodes.

ground_truth.json is the fixture reference for generator parameters and planted defect locations. Expected findings describe the data, not a promised grader score. No Kalanos grader assessment is claimed here.

Source and intended use

Generated by the Kalanos Hub synthetic-data generator (kalanos-public-synthetic, generator version 5, specification version 2). Intended for dataset QA, loader checks, defect-detection tests, regression tests, and demonstrations.

These are synthetic signals, not recorded robot demonstrations or physics simulation. There are no videos or useful visual observations. The fixtures are not intended for policy training, measuring real-world task success, or robot deployment. The metadata robot type franka_synthetic does not mean a physical Franka robot recorded these data.

Dataset configuration and size

10 episodes, 200 steps per episode, 2,000 rows.
Nominal sampling frequency: 50 Hz; random seed: 42.
LeRobot v3 Parquet structure, with data/ and meta/ at repository root.
Five data/metadata/ground-truth files total 168,068 bytes, excluding this card and .gitattributes.
meta/info.json defines feature names, dtypes, shapes, paths and task counts. meta/tasks.parquet maps task indices to instructions.

Schema and units

observation.state: float32 vector of six synthetic joint-position signals.
action: float32 vector of six arm signals plus one binary gripper signal (-1 open, +1 closed).
observation.eef_pos: float32 vector of three synthetic end-effector coordinates.
observation.eef_quat: float32 quaternion in qx, qy, qz, qw order.
timestamp: float32 time in seconds within an episode; nominal interval 0.02 seconds, with configured jitter and any intentional temporal defect.
frame_index, episode_index, index, task_index: int64 indices.

Joint, action and end-effector coordinate values are synthetic numerical signals. No calibrated physical unit or hardware-command convention is asserted. Do not use them directly to control a robot.

Load locally with LeRobot 0.7.0

The local example below was tested with LeRobot 0.7.0 against these fixture files. Download this repository first and point root to the directory containing data/ and meta/. For a reproducible download, select and record an explicit commit from Files and versions. This example does not claim a tested remote-download workflow.

from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = Path("./synthetic-missing-task-labels").resolve()
dataset = LeRobotDataset(
    "kalanosai/synthetic-missing-task-labels",
    root=root,
    download_videos=False,
    video_backend="pyav",
)
assert len(dataset) == 2000
sample = dataset[0]
print(sample["action"].shape)  # torch.Size([7])

Validation and interpretation

Local validation with LeRobot 0.7.0 read all 2,000 frames and found no task-label mismatches against the intended fixture. This establishes loader and fixture consistency, not a Kalanos grade, a guarantee that all checks pass, or evidence of training usefulness. Intentional defects remain in the three defect fixtures.

Related synthetic QA fixtures

All four share the basic configuration, but should not be assumed to have byte-identical baseline signals: defect injection can affect the random-number sequence. Use each repository’s own ground truth.

License

Apache License 2.0. Kalanos has confirmed rights to license the generated content.

Downloads last month: 87