Title

fix(train): initialize policy before accelerator creation

Type / Scope

• Type: Bug
• Scope: lerobot_train

Summary / Motivation

When training with --policy.path, the cfg.policy object remains None until cfg.validate() is explicitly called. This is due to the custom argument filtering in the parser that handles loading configs from local or hub paths. Currently, the code attempts to access cfg.policy.device to initialize the Accelerator before cfg.validate() is executed, leading to an AttributeError: 'NoneType' object has no attribute 'device'. This change moves the validation step to the beginning of the training pipeline to ensure the policy is correctly initialized from the provided path before being accessed.

Related issues

• Fixes: # (if any)

What changed

• Moved cfg.validate() call to the beginning of the train() function in src/lerobot/scripts/lerobot_train.py.
• This ensures cfg.policy is populated when using --policy.path before it's used for device configuration in Accelerator.

How was this tested

• Manual verification by running training with the --policy.path argument, which previously crashed with AttributeError.

How to run locally (reviewer)

• Run a training job using a pretrained policy path:

  python src/lerobot/scripts/lerobot_train.py \
    --policy.path=lerobot/smolvla_base \
    --dataset.repo_id=your/dataset \
    --policy.device=cuda

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green

Reviewer notes

• The change is minimal and only affects the order of initialization. It is necessary because draccus doesn't handle the --policy.path logic natively, and our validate() method is what actually loads the policy config into the cfg object.