feat(g05): train with LeRobot language recipes

docs/source/g05.mdx

@@ -123,9 +123,11 @@ For example, fine-tune the private SO-101 checkpoint on a LeRobot dataset:
 export HF_USER=your_hf_username
 
 lerobot-train \
-  --dataset.repo_id=${HF_USER}/my_so101_dataset \
+  --dataset.repo_id=${HF_USER}/my_so101_dataset_annotated \
   --policy.path=lerobot/g05_so101 \
   --policy.device=cuda \
+  --policy.recipe_path=recipes/g05_bbox_subtask.yaml \
+  --policy.cot_bbox_camera=observation.images.exterior \
   --policy.repo_id=${HF_USER}/g05_so101_finetuned \
   --policy.private=true \
   --output_dir=outputs/train/g05_so101 \
@@ -135,6 +137,31 @@ lerobot-train \
   --save_freq=1000
 ```
 
+The bundled `g05_bbox_subtask.yaml` recipe resolves the active
+`language_persistent` `subtask` and camera-scoped grounded `vqa` event at each
+sample timestamp. It filters out unavailable formats before selecting one of
+four author-compatible objectives:
+
+| Assistant sequence         | Weight |
+| -------------------------- | -----: |
+| Action only                |      1 |
+| Subtask, then action       |      2 |
+| BBox, then action          |      1 |
+| BBox, subtask, then action |      1 |
+
+Grounded VQA boxes are converted from pixel-space `xyxy` JSON using the source
+camera dimensions captured before image resizing, then serialized as G0.5
+`<locXXXX>` tokens. Joint samples preserve the released checkpoint's
+`BBox → Subtask → Action` order. The user/task conditioning tokens remain
+masked; the author backend applies its language/action objective to the
+assistant sequence.
+
+Generate the required `subtask` and grounded `vqa` language columns with
+`lerobot-annotate` as described in the
+[annotation pipeline](./annotation_pipeline). The bundled recipe targets
+`observation.images.exterior`; copy the YAML and change its camera-filtered
+bindings when training an embodiment with a different grounded camera.
+
 The SO-101 recipe uses AdamW at `8e-5` with 1,000 warmup steps. The packaged
 LIBERO and RoboTwin configurations use their released `1e-5` recipe, with
 1,000 and 500 warmup steps respectively. All profiles preserve G0.5's six

docs/source/language_and_recipes.mdx

@@ -146,6 +146,31 @@ The renderer does not apply a tokenizer chat template. Policy processors decide
 Blend recipes select one weighted sub-recipe deterministically from the sample index.
 `recipes/subtask_mem.yaml` trains the compact core blend — high-level subtask prediction, low-level execution, and memory. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds VQA and spoken interjection responses.
 
+Annotation-dependent blends can set `select_from_applicable: true` and declare
+`requires` on each component. The renderer first removes components whose
+required bindings resolve to `None`, then performs the deterministic weighted
+selection. This matches mixed-CoT policies where unavailable annotation formats
+must not consume probability:
+
+```yaml
+select_from_applicable: true
+blend:
+  subtask:
+    weight: 2
+    requires: [subtask]
+    messages:
+      - {
+          role: assistant,
+          content: "${subtask}",
+          stream: low_level,
+          target: true,
+        }
+  action:
+    weight: 1
+    messages:
+      - { role: user, content: "${task}", stream: low_level }
+```
+
 A message recipe with a supervised assistant turn on the `low_level` stream trains
 the π0.5 paper's joint sequence instead of a blend: the target span gets text CE
 while also conditioning the action losses in the same forward.