Public training code should be policy-agnostic. This removes:
- resume_pretrain config field (pretrained_path without --resume already
provides the same "load weights, fresh optimizer" behavior)
- evo1-specific DDP static graph and stage2 logging in lerobot_train.py
- drop_last sourced from policy config (DataLoader concern, not policy)
- type-inference fallback in PreTrainedConfig that was added to handle
evo1 checkpoints missing a "type" field
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- add viz.video_keys for selecting and ordering camera views
- add viz.smooth_window with default 1 (disabled)
- wire new viz options through lerobot_value_infer export path
- migrate validated overlay style and multiview export logic