Skip to content

Recording

The part of roboloom.recording that reads and checks a recorded dataset. The columns are listed in Dataset features and the procedure is in Checking the data and training. This page holds the descriptions generated from the docstrings.

Checking a dataset

roboloom.recording.check.check_dataset(root, recover=False)

Check the dataset under root. recover=True first undoes a save that was interrupted (what recording into the dataset again does too).

roboloom.recording.check.DatasetReport dataclass

What check_dataset found.

Attributes:

Name Type Description
root Path

The dataset directory.

episodes int

The episodes the dataset holds.

frames int

The rows it holds.

summary list[str]

Lines describing the dataset (counts, verdicts).

problems list[str]

What makes the dataset wrong as it is (unreadable files, counts that disagree, a video shorter or longer than its episode).

notes list[str]

What a training run should know (rows whose action or observation is not valid).

ok property

True when there are no problems (notes do not count).

describe()

The report as roboloom check-dataset prints it.

Episode records

roboloom.recording.episodes

What each saved episode was: meta/roboloom_episodes.jsonl in the dataset.

One line is appended when an episode is saved: its task, rows, quality metrics, what the controllers reported when it ended (a policy's inference time, the chunks it dropped, its error) and the operator's verdict when one was given. roboloom label appends a line with the verdict afterwards. Lines are only ever appended; read merges the lines of an episode, later ones winning.

LeRobot does not read this file: a dataset trains the same with or without it. It is what an evaluation run is counted from (summary).

read(root)

Episode index -> its record. A line cut off by a crash is skipped.

label(root, episode, verdict, note=None)

Give the saved episode its verdict (replacing an earlier one); returns the episode's record.

summary(root, episodes)

How many of the dataset's episodes carry each verdict; None when no episode has a record.

total(root)

How many episodes the dataset under root holds.

saved(root, episode, task, frames, metrics, label=None, note=None, controllers=None)

Record that episode was saved; returns the line written.

check_label(label)

label as it is stored: None (no verdict) or one of LABELS.

Features

roboloom.recording.features(cfg)

The LeRobot features a configuration records: name -> {"dtype", "shape", "names"}. action* only with followers, action.intervention only with a policy that has teleop:, action.filtered only for YAML hook write claims, then each recorded source's own features and observation.meta.<source>.*, then the joined ones (observation.state, ...).