fix(train): restrict legacy RA-BC migration to JSON checkpoints only

[masato-ka]

_migrate_legacy_rabc_fields was called for all config files, causing json.load to raise DecodeError when a YAML config was passed to lerobot-train for a new training run. Guard the block with an .endswith(".json") check so migration only runs when resuming from a JSON checkpoint.

Also adds types-PyYAML to pre-commit mypy additional_dependencies to fix a pre-existing [import-untyped] error in envs/robotwin.py.

Title

Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). See CONTRIBUTING.md for PR conventions.

Summary / Motivation

PR #3480 (cd6b43e) introduced _migrate_legacy_rabc_fields, which was called unconditionally
whenever config_file is not None. Since this function calls json.load() as its first step, passing
a YAML config file via --config_path to start a new training run raised a JSONDecodeError and
aborted the process. This fix adds an .endswith(".json") guard to the migration block so that the
legacy RA-BC migration only runs when resuming from a JSON checkpoint. A clarifying comment was
also added to make the intent clear to future readers. Additionally, a pre-existing mypy
[import-untyped] error in lerobot/envs/robotwin.py (missing types-PyYAML stubs) was resolved by
adding types-PyYAML to the pre-commit mypy additional_dependencies.

Reproduction steps (before fix)

Raises JSONDecodeError on main with #3480 applied

  lerobot-train --config_path=act-training.yaml             

  # act-training.yaml
  dataset:                                                                                          
    repo_id: lerobot/pusht                                  
  policy:
    type: act
    push_to_hub: false
    device: cuda
  output_dir: outputs/train/act-pusht                                                               
  job_name: pusht
  batch_size: 8                                                                                     
  save_freq: 20000                                          
  steps: 40000
  wandb:
    enable: false
    project: lerobot

Related issues

• Fixes / Closes:
• Related: Migrate legacy RA-BC fields in train config loading #3480, Fix policy.path not working in YAML config files #3145

What changed

• src/lerobot/configs/train.py: Changed the legacy RA-BC migration guard in
  TrainPipelineConfig.from_pretrained from if config_file is not None: to if config_file is not None
  and os.path.basename(config_file).endswith(".json"):, preventing json.load from being called on
  YAML configs.
  • src/lerobot/configs/train.py: Added a comment explaining the intent of the guard condition.
  • .pre-commit-config.yaml: Added types-PyYAML to the mypy hook's additional_dependencies to fix
    the pre-existing [import-untyped] error in lerobot/envs/robotwin.py.

How was this tested (or how to run locally)

1. Resume from legacy RA-BC JSON checkpoint (regression check for Migrate legacy RA-BC fields in train config loading #3480)
   Confirmed that resuming from a JSON checkpoint correctly migrates sample_weighting and
   sub-options, and that training resumes without error.

  lerobot-train \                                           
    --config_path=./outputs/train/policy_smolbla_rabc/checkpoints/last/pretrained_model/train_config
  .json \                                                                                           
    --steps=60000 \
    --resume=true                                                                                   

2. New training run with YAML config
   Confirmed that running the command below with the fix applied starts training successfully without
   a JSONDecodeError.

  lerobot-train --config_path=act-training.yaml   

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green
• Community Review: I have reviewed another contributor's open PR and linked it here: # (insert PR number/link)

Reviewer notes

• Anything the reviewer should focus on (performance, edge-cases, specific files) or general notes.
• Anyone in the community is free to review the PR.