---

name: data-integration description: "Integrate LeRobot-format StarVLA datasets, modality schemas, robot DataConfig classes, mixtures, and registry discovery." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

StarVLA data integration

Use this sub-skill when the task is to add, debug, or review StarVLA robot-action data sources: LeRobot-format datasets, meta/modality.json, robot DataConfig classes, dataset mixtures, embodiment tags, registry discovery, statistics caches, and safe custom-dataset templates.

Route elsewhere

• Training command construction, YAML launch overrides, Accelerate/DeepSpeed, or checkpoint resume: training-config.
• Policy request/response schemas, server-side unnormalization, deployment clients, or real-robot bridge calls: policy-deployment.
• Simulator installation, benchmark evaluator loops, rendering, or benchmark environment setup: benchmark-evaluation.

Operating map

1. Confirm that each dataset is a LeRobot dataset with meta/modality.json, LeRobot metadata files, parquet data, and task metadata. Use LeRobot data format for the expected shape.
2. Draft or edit the robot DataConfig: define video_keys, state_keys, action_keys, language_keys, observation_indices, optional state_indices, and action_indices; implement modality_config() and transform(). Use dataset registry.
3. Register the robot and mixture by exporting ROBOT_TYPE_CONFIG_MAP and DATASET_NAMED_MIXTURES from an auto-discovered examples/**/train_files/data_registry/data_config.py module. Prefer an embodiment_tag class variable on the DataConfig; legacy ROBOT_TYPE_TO_EMBODIMENT_TAG overrides still work.
4. Keep len(action_indices) aligned with the model/training action_horizon for the selected framework. If you are editing launch YAML, route to training-config.
5. Validate candidate modality.json files before data loading with scripts/validate_modality_json.py from this sub-skill.
6. If the user starts from the bundled custom-dataset assets, adapt only the dataset and registry pieces here; route the training launcher and benchmark/policy bridge pieces as described in dataset integration templates.

Must-know StarVLA contracts

• StarVLA data loading uses data_root_dir plus DATASET_NAMED_MIXTURES[data_mix]; each mixture entry is (dataset_subdir, sampling_weight, robot_type).
• ROBOT_TYPE_CONFIG_MAP[robot_type] must produce a DataConfig instance whose modality keys match the meta/modality.json subkeys.
• Language keys are StarVLA keys such as annotation.human.task_description or annotation.human.action.task_description; the corresponding modality.json annotation entry must read task ids with original_key: "task_index".
• State/action start and end slices are end-exclusive slices into the LeRobot column named by original_key or the schema default.
• StateActionTransform owns training-time normalization (min_max, q99, mean_std, or binary) and reads dataset statistics generated from LeRobot parquet files.
• StarVLA may rebuild meta/stats_gr00t.json when cache format or action-mode configuration changes; keep the training-output dataset_statistics.json paired with checkpoints for deployment.

References

• LeRobot data format
• Dataset registry
• Dataset integration templates
• Troubleshooting

Safe helper

• scripts/validate_modality_json.py validates the JSON shape without importing StarVLA or opening data files. It catches missing top-level modalities, invalid state/action slices, and the common language-conditioning error where the annotation field does not use task_index.

Evidence notes

Distilled from StarVLA dataloader and registry sources (starVLA/dataloader/__init__.py, lerobot_datasets.py, gr00t_lerobot/registry.py, data_config.py, datasets.py, schema.py, transform/*), representative benchmark/robot data registries, the custom integration asset templates, and the README/guideline data sections.