Skip to content

Checking and training

How to check a recorded dataset before training, leave out episodes you do not want, and train with LeRobot.

Checking a dataset

roboloom check-dataset [robot] checks a dataset by reading it only (CLI). Name a robot (with --with / --without as recorded) or a dataset directory.

uv run roboloom check-dataset --with policy
Output Content
Summary Episodes, rows, fps, robot, the roboloom version that recorded it, file counts; with episode records, the verdict counts (labels: 7 success, 3 failure, ...)
problems What is broken: unreadable files, counts that disagree between info.json and the episode metadata, episode indices with gaps, missing videos and videos that do not match their rows, a half-written episode
notes What to know before training: the share of rows that are 0, per *.valid column; the share of rows whose action is all zeros (recorded before rows without a command kept the held target)
  • With problems, the command exits with code 1.
  • When a save was interrupted and a half-written episode is left, --recover removes it before checking. Recording into the dataset again removes it too (What happens to a saved episode).
  • What valid means is in Validity flags; the values of rows that are 0 are in Missing data.

Leaving out bad episodes

The per-episode metrics (Episode records) tell which episodes to train on.

from roboloom.recording import episodes

records = episodes.read(".roboloom/rakuda2/datasets/rakuda2")
keep = [
    index
    for index, record in sorted(records.items())
    if record.get("metrics", {}).get("action_valid_rate", 0) >= 0.9
]
print(keep)

Pass the indices to lerobot-train as --dataset.episodes, and it trains on those episodes only.

uv run lerobot-train \
  --dataset.repo_id=local/rakuda2 \
  --dataset.root=.roboloom/rakuda2/datasets/rakuda2 \
  --dataset.episodes='[0, 1, 3, 4]' \
  --policy.type=act --policy.device=cuda --policy.push_to_hub=false \
  --trackio.enable=false \
  --output_dir=outputs/train/act_rakuda2
  • --dataset.episodes selects rows only. The normalization statistics stay those of the whole dataset (meta/stats.json). When the left-out episodes hold outlying values, make a dataset without them instead, as below.
  • LeRobot's lerobot-edit-dataset --operation.type delete_episodes --operation.episode_indices "[2, 5]" writes a new dataset without those episodes and aggregates the statistics of the kept ones again (--root, --new_root and --new_repo_id give the places). The episodes are renumbered from 0. roboloom's files (meta/roboloom.json, meta/roboloom_episodes.jsonl) are not copied to the new dataset.
  • To recompute the statistics of a whole dataset, use lerobot-edit-dataset --operation.type recompute_stats.
  • Both need lerobot[dataset], which --extra train (lerobot[training]) installs. roboloom has not checked these tools; see LeRobot's documentation for their use.
  • The --dataset.* options of lerobot-train have no setting that selects rows by action.valid or the like. To select rows, read LeRobotDataset(...).hf_dataset and drop them yourself (Missing data).

Training with LeRobot

A dataset roboloom recorded (LeRobotDataset v3.0) trains with lerobot-train as it is. The extras, the command and how to run the trained checkpoint are in Training with LeRobot.

  • --dataset.repo_id is the configuration's repo_id; --dataset.root is its output, from the robot's directory.
  • lerobot-train does not look at the *.valid columns. In rows whose action.valid is 0, action holds the target each follower keeps.
  • Before training on a GPU, check with labstatus that a GPU is free.
  • An example of training your own model is in From a recording to your own model.