physicalai-train-working-with-datasets

Working with Studio Datasets

Studio data lives in library/src/physicalai/data/. Datasets use the LeRobot format and are consumed through Lightning datamodules. The datamodules are first-class Python API objects; YAML/CLI configs are a serialization of the same construction path.

Key modules:

• data/lerobot/datamodule.py — LeRobotDataModule (the class configs reference as physicalai.data.lerobot.LeRobotDataModule).
• data/lerobot/dataset.py — LeRobot dataset wrapper.
• data/lerobot/converters.py — DataFormat (StrEnum: physicalai, lerobot) and bidirectional field mapping between the two layouts.
• data/observation.py — Observation, Feature, FeatureType, NormalizationParameters.
• data/datamodules.py — base DataModule (Lightning LightningDataModule, auto num-workers heuristic).
• data/dataset.py — base Dataset; data/gym.py — GymDataset for gym-generated data.

Python API usage

Use this path for notebooks, tests, direct batch inspection, or debugging dataloading without involving the training CLI.

from physicalai.data import LeRobotDataModule

datamodule = LeRobotDataModule(repo_id="lerobot/pusht", train_batch_size=2) datamodule.prepare_data() datamodule.setup("fit") batch = next(iter(datamodule.train_dataloader()))

Done when: the batch contains the observation/action fields the policy expects, with the expected batch/action dimensions.

## Wiring data into a training config

In a `physicalai fit` config, the `data` block selects the datamodule and its `repo_id`:

```yaml
data:
  class_path: physicalai.data.lerobot.LeRobotDataModule
  init_args:
    repo_id: lerobot/pusht
    train_batch_size: 64

repo_id points at a LeRobot/HuggingFace dataset; the datamodule pulls it on first use. See the physicalai-train-training-a-policy skill for the full config.

Workflow

1. Pick the dataset by repo_id and confirm its features (image keys, state dim, action dim) match the target policy's Config.
   • Done when: the policy's expected Feature names and action dimension line up with the dataset.
2. Verify a batch through the Python API before training:

   datamodule.prepare_data()
   datamodule.setup("fit")
   batch = next(iter(datamodule.train_dataloader()))
   

   • Done when: the batch has correct keys and shapes without invoking the CLI.
3. Verify CLI parity when the dataset is configured through YAML:

   physicalai fit --config <config.yaml> --trainer.fast_dev_run=true
   

   • Done when: one batch flows through with correct shapes and no missing-feature errors.
4. Convert layouts only when needed via converters.py (DataFormat.physicalai ↔ DataFormat.lerobot); keep field names stable, since they propagate to training and export.
5. Set normalization through NormalizationParameters/Feature consistently with what the policy expects at inference.

Debugging dataloading

• Missing/renamed feature → the config's dataset features disagree with the policy; align Feature names in data/observation.py conventions.
• Slow/stalled first batch → the LeRobot repo_id is downloading; expected on first run (see the requires_download test marker for tests that need this).
• Wrong batch dimensions → check train_batch_size and the datamodule's collate/observation handling before changing the policy.

Required checks

• Feature names, FeatureType, action dim, and normalization match between dataset, Config, and any export metadata.
• Conversions round-trip without dropping or renaming fields.
• Direct datamodule API construction and YAML config construction produce compatible batches.
• Tests that require downloads are marked requires_download; keep default uv run pytest runnable offline.

Verify

# from library/
uv run pytest tests/unit/data tests/unit/datamodules

Related skills

• physicalai-train-training-a-policy — the data block is one half of a training config.
• physicalai-train-adding-a-policy — align observation features with the policy Config.