Commit History
30 days ago
-
Drop policy_config from the tokenizer step
Verified signature`policy_config` duplicated fields that `config.json` already owns (`author_model_config`, `processor_metadata`, `camera_order`, `runtime_system`, ...). The copy could not be kept in sync, which is why the loader had to re-apply `runtime_system` onto the step after every load. The live `G05Config` is now authoritative: `G05TokenizerStep.get_config()` no longer serializes `policy_config`, and `reconcile_g05_processors` re-injects it on load. Dropping it here shrinks this file by about 80%. No behaviour change; weights are untouched.0ceb36c -
Update render step for lerobot #4183
Verified signature`policy_preprocessor.json` named `render_messages_processor`, a step removed from lerobot on 2026-09-09 by huggingface/lerobot#4183, which split it into `render_runtime_messages_processor` and `render_training_messages_processor`. The registry has no alias, so loading this checkpoint failed outright with an ImportError. Rewritten to the two-step layout that `make_g05_pre_post_processors` builds today; `render_training` was false here, so the training step gets a null recipe. Weights are untouched.3b2b5cc
-
Move G0.5 tokenization into policy processor
Verified signature87b9ffc -
Use unified semantic-message processor
Verified signatureReplace separate runtime/training renderers with one RenderMessagesStep, embed the policy recipe, and preserve all model and normalization tensors.2b87340
-
Add converted OpenGalaxea G0.5 base checkpoint
Verified signaturea74ab9a