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.

src/lerobot/configs/recipe.py

@@ -104,7 +104,9 @@ class TrainingRecipe:
 
     messages: list[MessageTurn] | None = None
     bindings: dict[str, str] | None = None
+    requires: list[str] | None = None
     blend: dict[str, TrainingRecipe] | None = None
+    select_from_applicable: bool = False
     weight: float | None = None
 
     def __post_init__(self) -> None:
@@ -151,6 +153,9 @@ class TrainingRecipe:
         assert self.messages is not None
         known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
 
+        missing_requirements = set(self.requires or ()) - known_bindings
+        if missing_requirements:
+            raise ValueError(f"TrainingRecipe requires unknown binding(s): {sorted(missing_requirements)}")
         for turn in self.messages:
             missing = self._referenced_bindings(turn) - known_bindings
             if missing:
@@ -168,6 +173,8 @@ class TrainingRecipe:
     def _validate_blend_recipe(self) -> None:
         """Ensure each blend component is a non-empty, weighted message recipe."""
         assert self.blend is not None
+        if self.requires:
+            raise ValueError("A blend recipe cannot declare requires; set it on its components.")
         if not self.blend:
             raise ValueError("Blend recipes must contain at least one component.")
 

src/lerobot/configs/recipes/g05_bbox_subtask.yaml

@@ -0,0 +1,60 @@
+# G0.5's released MixedSamplesBuilder semantics, limited to the CoT signals
+# LeRobot can currently provide: active subtasks and grounded bounding boxes.
+#
+# The renderer first removes branches whose required annotations are missing,
+# then deterministically samples among the remaining weights. The G0.5 policy
+# processor converts the selected target messages into the checkpoint's exact
+# BBox/Subtask/action template and supervises the complete assistant segment.
+
+select_from_applicable: true
+blend:
+  no_cot_action:
+    weight: 1
+    messages:
+      - {role: user, content: "${task}", stream: low_level}
+
+  subtask_action:
+    weight: 2
+    requires: [subtask]
+    messages:
+      - {role: user, content: "${task}", stream: low_level}
+      - {
+          role: assistant,
+          content: "Subtask: ${subtask}",
+          stream: low_level,
+          target: true,
+        }
+
+  bbox_action:
+    weight: 1
+    requires: [bbox]
+    bindings:
+      bbox: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.exterior)"
+    messages:
+      - {role: user, content: "${task}", stream: low_level}
+      - {
+          role: assistant,
+          content: "BBoxJSON: ${bbox}",
+          stream: low_level,
+          target: true,
+        }
+
+  bbox_subtask_action:
+    weight: 1
+    requires: [bbox, subtask]
+    bindings:
+      bbox: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.exterior)"
+    messages:
+      - {role: user, content: "${task}", stream: low_level}
+      - {
+          role: assistant,
+          content: "BBoxJSON: ${bbox}",
+          stream: low_level,
+          target: true,
+        }
+      - {
+          role: assistant,
+          content: "Subtask: ${subtask}",
+          stream: low_level,
+          target: true,
+        }

src/lerobot/datasets/language_render.py

@@ -169,6 +169,17 @@ def render_sample(
     persistent_rows = _normalize_rows(persistent or [])
     event_rows = _normalize_rows(events or [])
 
+    if recipe.blend is not None and recipe.select_from_applicable:
+        return _render_applicable_blend(
+            recipe,
+            persistent=persistent_rows,
+            events=event_rows,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+
     # Route sparse VQA frames to a matching view-specific component before weighted selection.
     # This avoids dropping annotated frames or selecting VQA without annotations.
     if recipe.blend is not None:
@@ -197,6 +208,54 @@ def render_sample(
     return _render_message_recipe(selected_recipe, bindings)
 
 
+def _render_applicable_blend(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> RenderedMessages | None:
+    """Select deterministically among components whose required bindings resolve.
+
+    Ordinary blends preserve their historical select-then-resolve behavior.
+    Annotation-dependent policies such as G0.5 instead need the author
+    ``MixedSamplesBuilder`` contract: discard unavailable CoT formats first,
+    then draw according to the remaining relative weights.
+    """
+
+    assert recipe.blend is not None
+    renderable: list[tuple[float, RenderedMessages]] = []
+    for component in recipe.blend.values():
+        bindings = _resolve_bindings(
+            component,
+            persistent=persistent,
+            events=events,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        rendered = _render_message_recipe(component, bindings)
+        if rendered is not None:
+            renderable.append((float(component.weight or 0.0), rendered))
+
+    if not renderable:
+        return None
+
+    total_weight = sum(weight for weight, _ in renderable)
+    digest = hashlib.blake2b(f"applicable:{sample_idx}".encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total_weight
+    cumulative = 0.0
+    for weight, rendered in renderable:
+        cumulative += weight
+        if draw < cumulative:
+            return rendered
+    return renderable[-1][1]
+
+
 def _render_vqa_if_present(
     recipe: TrainingRecipe,
     *,
@@ -385,6 +444,8 @@ def _render_message_recipe(
 ) -> RenderedMessages | None:
     """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
     assert recipe.messages is not None
+    if any(bindings.get(name) is None for name in recipe.requires or ()):
+        return None
     messages: list[dict[str, Any]] = []
     streams: list[str | None] = []
     target_indices: list[int] = []

src/lerobot/policies/g05/configuration_g05.py

@@ -91,6 +91,39 @@ def make_g05_prompt_template(num_images: int, *, predict_cot: bool, flow_only: b
     return f"{prefix}Action: <EOV><EOC><action_action>|<eos>"
 
 
+def make_g05_cot_prompt_template(
+    num_images: int,
+    *,
+    fields: tuple[str, ...],
+    flow_only: bool,
+) -> str:
+    """Build the exact author template for a selected Subtask/BBox CoT format."""
+
+    supported_fields = {"bbox", "subtask"}
+    if not fields or not set(fields) <= supported_fields:
+        raise ValueError(
+            f"G0.5 CoT fields must be a non-empty subset of {sorted(supported_fields)}, got {fields}."
+        )
+    # The released BBoxSubtaskCoTBuilder emits BBox before Subtask. Preserve
+    # checkpoint serialization even though the paper's schematic orders the
+    # independently composable labels differently.
+    ordered_fields = tuple(field for field in ("bbox", "subtask") if field in fields)
+    placeholders = {
+        "bbox": "<bbox_text>|",
+        "subtask": "<atomic_task_text>|",
+    }
+    images = "".join(f"<image{index}_image_!>" for index in range(num_images))
+    prefix = (
+        f"<chat_user_prefix>{images}<bos>"
+        "Embodiment: <embodiment_text_!>; Task: <command_text_!_200> "
+        "State: <proprio_proprio_!>;"
+        "<chat_user_suffix><chat_assistant_prefix>"
+        "<prompt_text_!>\n<EOC>"
+    )
+    action = "Action: <EOV><eos>" if flow_only else "Action: <EOV><action_action>|<eos>"
+    return prefix + "".join(placeholders[field] for field in ordered_fields) + action
+
+
 # Raw dimensions are inserted in these exact policy slots. The G0.5 shared layout is:
 # left_control[9] | left_gripper[1] | right_control[9] | right_gripper[1] | lower_body[7].
 # LIBERO uses only the right EEF delta and right gripper. atomic_4 is a single-arm mobile
@@ -198,6 +231,8 @@ class G05Config(PreTrainedConfig):
     processor_metadata: dict[str, Any] = field(default_factory=dict)
     action_codec_metadata: dict[str, Any] = field(default_factory=dict)
     prompt_template: str = ""
+    recipe_path: str | None = None
+    cot_bbox_camera: str | None = None
 
     normalization_mapping: dict[str, NormalizationMode] = field(
         default_factory=lambda: {
@@ -254,6 +289,8 @@ class G05Config(PreTrainedConfig):
             raise ValueError("runtime_system must be 'system1' or 'system2'.")
         if self.runtime_system == "system2" and not self.predict_cot:
             raise ValueError("G0.5 System 2 requires predict_cot=True in the packaged checkpoint.")
+        if self.recipe_path is not None and not self.predict_cot:
+            raise ValueError("G0.5 recipe-driven CoT training requires predict_cot=True.")
         if not 1 <= self.n_action_steps <= self.chunk_size:
             raise ValueError("n_action_steps must be between 1 and chunk_size.")
         if self.action_head == "actioncodec" and not self.discrete_action:
@@ -314,6 +351,8 @@ class G05Config(PreTrainedConfig):
             raise ValueError("camera_sizes must contain exactly the ordered checkpoint camera keys.")
         if not set(self.optional_camera_keys) <= set(self.camera_order):
             raise ValueError("optional_camera_keys must be a subset of camera_order.")
+        if self.cot_bbox_camera is not None and self.cot_bbox_camera not in self.camera_order:
+            raise ValueError("cot_bbox_camera must be one of camera_order.")
         if self.num_input_images != len(self.camera_order) * self.n_obs_steps:
             raise ValueError(
                 "num_input_images must equal len(camera_order) * n_obs_steps for the selected checkpoint."

src/lerobot/policies/g05/modeling_g05.py

@@ -11,6 +11,7 @@
 from __future__ import annotations
 
 import importlib
+import json
 import shutil
 from collections import deque
 from collections.abc import Mapping
@@ -26,7 +27,12 @@ from lerobot.optim.optimizers import OptimizerParams
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION, OBS_STATE
 
-from .configuration_g05 import G05_POLICY_PARTS, G05Config
+from .configuration_g05 import (
+    G05_POLICY_PARTS,
+    G05Config,
+    make_g05_cot_prompt_template,
+    make_g05_prompt_template,
+)
 
 
 def _author_backend(config: G05Config) -> nn.Module:
@@ -229,6 +235,158 @@ class G05Policy(PreTrainedPolicy):
             return value[index]
         return value
 
+    def _recipe_cot_targets(
+        self,
+        batch: Mapping[str, Any],
+        index: int,
+        batch_size: int,
+    ) -> tuple[str | None, str | None]:
+        """Read the selected recipe's supervised Subtask/BBox messages."""
+
+        messages = batch.get("messages")
+        target_indices = batch.get("target_message_indices")
+        if messages is None or target_indices is None:
+            return None, None
+
+        sample_messages = messages
+        if (
+            isinstance(messages, list | tuple)
+            and len(messages) == batch_size
+            and (not messages or isinstance(messages[0], list | tuple))
+        ):
+            sample_messages = messages[index]
+        sample_target_indices = target_indices
+        has_batched_target_indices = (isinstance(target_indices, Tensor) and target_indices.ndim > 1) or (
+            isinstance(target_indices, list | tuple)
+            and len(target_indices) == batch_size
+            and (not target_indices or isinstance(target_indices[0], list | tuple | Tensor))
+        )
+        if has_batched_target_indices:
+            sample_target_indices = target_indices[index]
+        if isinstance(sample_messages, Mapping):
+            sample_messages = [sample_messages]
+        if isinstance(sample_target_indices, Tensor):
+            sample_target_indices = sample_target_indices.detach().cpu().tolist()
+        if not isinstance(sample_messages, list | tuple) or not isinstance(
+            sample_target_indices, list | tuple
+        ):
+            return None, None
+
+        subtask: str | None = None
+        bbox_json: str | None = None
+        for target_index in sample_target_indices:
+            message = sample_messages[int(target_index)]
+            content = message.get("content") if isinstance(message, Mapping) else None
+            if not isinstance(content, str):
+                continue
+            if content.startswith("Subtask:"):
+                value = content.removeprefix("Subtask:").strip()
+                if value:
+                    subtask = value
+            elif content.startswith("BBoxJSON:"):
+                value = content.removeprefix("BBoxJSON:").strip()
+                if value:
+                    bbox_json = value
+        return subtask, bbox_json
+
+    @staticmethod
+    def _format_bbox_target(bbox_json: str | None, image_size: tuple[int, int]) -> str | None:
+        """Convert LeRobot grounded-VQA JSON into G0.5's location-token format."""
+
+        if not bbox_json:
+            return None
+        try:
+            payload = json.loads(bbox_json)
+            if isinstance(payload, str):
+                payload = json.loads(payload)
+        except (json.JSONDecodeError, TypeError):
+            return None
+        if not isinstance(payload, Mapping):
+            return None
+        if isinstance(payload.get("answer"), Mapping):
+            payload = payload["answer"]
+
+        height, width = image_size
+        boxes: list[tuple[str, list[float]]] = []
+        detections = payload.get("detections")
+        if isinstance(detections, list):
+            for detection in detections:
+                if not isinstance(detection, Mapping) or detection.get("bbox_format", "xyxy") != "xyxy":
+                    continue
+                coords = detection.get("bbox")
+                if not isinstance(coords, list | tuple) or len(coords) != 4:
+                    continue
+                label = str(detection.get("label") or "object")
+                boxes.append((label, [float(value) for value in coords]))
+        else:
+            for label, coords in payload.items():
+                if isinstance(coords, list | tuple) and len(coords) == 4:
+                    boxes.append((str(label), [float(value) for value in coords]))
+        if not boxes:
+            return None
+
+        def normalize(coords: list[float]) -> list[float]:
+            if max(abs(value) for value in coords) <= 1.0:
+                return coords
+            x1, y1, x2, y2 = coords
+            return [x1 / width, y1 / height, x2 / width, y2 / height]
+
+        def location_token(value: float) -> str:
+            location = max(0, min(1023, round(value * 1024)))
+            return f"<loc{location:04d}>"
+
+        formatted = []
+        for label, raw_coords in boxes:
+            x1, y1, x2, y2 = normalize(raw_coords)
+            locations = "".join(location_token(value) for value in (y1, x1, y2, x2))
+            formatted.append(f"{label} {locations}")
+        return "BBox: " + "; ".join(formatted)
+
+    def _apply_recipe_cot(
+        self,
+        sample: dict[str, Any],
+        batch: Mapping[str, Any],
+        index: int,
+        batch_size: int,
+    ) -> bool:
+        """Populate one author sample from recipe-rendered CoT targets."""
+
+        subtask, bbox_json = self._recipe_cot_targets(batch, index, batch_size)
+        image_size = batch.get("g05_bbox_image_size")
+        if (
+            isinstance(image_size, list | tuple)
+            and len(image_size) == batch_size
+            and image_size
+            and isinstance(image_size[0], list | tuple | Tensor)
+        ):
+            image_size = image_size[index]
+        if isinstance(image_size, Tensor):
+            image_size = image_size.detach().cpu().tolist()
+        if not isinstance(image_size, list | tuple) or len(image_size) != 2:
+            camera = self.config.cot_bbox_camera or self.config.camera_order[0]
+            image_size = self.config.camera_sizes[camera]
+        bbox = self._format_bbox_target(bbox_json, (int(image_size[0]), int(image_size[1])))
+
+        fields = tuple(field for field, value in (("bbox", bbox), ("subtask", subtask)) if value)
+        if not fields:
+            return False
+        flow_only = "<action_action" not in self.config.prompt_template
+        sample["template"] = make_g05_cot_prompt_template(
+            self.config.num_prompt_images,
+            fields=fields,
+            flow_only=flow_only,
+        )
+        if bbox is not None:
+            sample["bbox"] = bbox
+        if subtask is not None:
+            sample["atomic_task"] = f"Subtask: {subtask}"
+        sample["prompt"] = {
+            ("bbox",): "predict bbox",
+            ("subtask",): "predict subtask",
+            ("bbox", "subtask"): "predict bbox, subtask and action",
+        }[fields]
+        return True
+
     def _prepare_author_batch(self, batch: Mapping[str, Any], task: str | None = None) -> dict[str, Any]:
         prepare = getattr(self.backend, "prepare_lerobot_batch", None)
         if callable(prepare):
@@ -284,13 +442,30 @@ class G05Policy(PreTrainedPolicy):
             if frequency is not None:
                 sample["frequency"] = frequency
             if self.config.predict_cot:
-                sample["prompt"] = "predict subtask"
-                atomic_task = batch.get("atomic_task")
-                if atomic_task is not None:
-                    atomic_task = str(self._batch_item(atomic_task, index, batch_size))
-                    sample["atomic_task"] = (
-                        atomic_task if atomic_task.startswith("Subtask:") else f"Subtask: {atomic_task}"
-                    )
+                rendered_recipe = "messages" in batch
+                applied_recipe_cot = rendered_recipe and self._apply_recipe_cot(
+                    sample, batch, index, batch_size
+                )
+                if not applied_recipe_cot:
+                    # During mixed-recipe training an applicable no-CoT branch is a
+                    # genuine target format. At inference, where actions are absent,
+                    # retain the checkpoint's configured System 2 prompt.
+                    if rendered_recipe and isinstance(batch.get(ACTION), Tensor):
+                        sample["template"] = make_g05_prompt_template(
+                            self.config.num_prompt_images,
+                            predict_cot=False,
+                            flow_only="<action_action" not in self.config.prompt_template,
+                        )
+                    else:
+                        sample["prompt"] = "predict subtask"
+                        atomic_task = batch.get("atomic_task")
+                        if atomic_task is not None:
+                            atomic_task = str(self._batch_item(atomic_task, index, batch_size))
+                            sample["atomic_task"] = (
+                                atomic_task
+                                if atomic_task.startswith("Subtask:")
+                                else f"Subtask: {atomic_task}"
+                            )
             for image_index in range(self.config.num_prompt_images):
                 camera = self.config.camera_order[image_index % len(self.config.camera_order)]
                 sample[f"image{image_index}"] = self.config.camera_sizes[camera]

src/lerobot/policies/g05/processor_g05.py

@@ -11,6 +11,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import Any
 
 import torch
@@ -48,12 +49,54 @@ from lerobot.utils.constants import (
 from .configuration_g05 import G05_EMBODIMENT_MAPPINGS, G05_POLICY_PARTS, G05Config
 
 
+def _load_recipe(path_str: str) -> Any:
+    """Load an absolute recipe path or one relative to ``lerobot/configs``."""
+
+    from lerobot.configs.recipe import TrainingRecipe
+
+    path = Path(path_str)
+    if not path.is_absolute() and not path.exists():
+        from lerobot.configs import recipe as recipe_module
+
+        candidate = Path(recipe_module.__file__).resolve().parent / path
+        if candidate.exists():
+            path = candidate
+    return TrainingRecipe.from_yaml(path)
+
+
 def _copy_feature_tree(
     features: dict[PipelineFeatureType, dict[str, PolicyFeature]],
 ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
     return {kind: values.copy() for kind, values in features.items()}
 
 
+@dataclass
+@ProcessorStepRegistry.register(name="g05_bbox_image_size")
+class G05BBoxImageSizeStep(ProcessorStep):
+    """Preserve the annotated camera's source size before checkpoint resizing."""
+
+    camera_key: str
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION) or {}
+        image = observation.get(self.camera_key)
+        if image is None:
+            return transition
+        image = torch.as_tensor(image)
+        if image.ndim < 3:
+            raise ValueError(f"G0.5 bbox camera {self.camera_key!r} has invalid shape {image.shape}.")
+        transition = transition.copy()
+        complementary = dict(transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
+        complementary["g05_bbox_image_size"] = (int(image.shape[-2]), int(image.shape[-1]))
+        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
 @dataclass
 @ProcessorStepRegistry.register(name="g05_image_transform")
 class G05ImageTransformStep(ProcessorStep):
@@ -690,17 +733,26 @@ def make_g05_pre_post_processors(
         action_names=list(config.action_feature_names) or None,
         num_obs_steps=config.n_obs_steps,
     )
-    steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+    steps: list[ProcessorStep] = [RenameObservationsProcessorStep(rename_map={})]
+    if config.recipe_path:
+        from lerobot.processor.render_messages_processor import RenderMessagesStep
+
+        steps.extend(
+            [
+                G05BBoxImageSizeStep(camera_key=config.cot_bbox_camera or config.camera_order[0]),
+                RenderMessagesStep(recipe=_load_recipe(config.recipe_path)),
+            ]
+        )
+    steps.append(AddBatchDimensionProcessorStep())
+    steps.append(
         G05ImageTransformStep(
             camera_order=config.camera_order,
             camera_sizes=config.camera_sizes,
             mean=config.image_mean,
             std=config.image_std,
             optional_camera_keys=config.optional_camera_keys,
-        ),
-    ]
+        )
+    )
     action_filter = config.processor_metadata.get("action_filter") or {}
     if str(action_filter.get("_target_", "")).endswith("R1LiteJointActionFilter"):
         action_parts = tuple(

tests/configs/test_recipe.py

@@ -57,6 +57,14 @@ def test_message_recipe_requires_at_least_one_target():
         )
 
 
+def test_message_recipe_requires_known_bindings():
+    with pytest.raises(ValueError, match="requires unknown binding"):
+        TrainingRecipe(
+            messages=[_minimal_target_turn()],
+            requires=["not_a_binding"],
+        )
+
+
 def test_recipe_rejects_both_messages_and_blend():
     with pytest.raises(ValueError, match="only one"):
         TrainingRecipe(
@@ -144,6 +152,35 @@ def test_from_dict_with_nested_blend():
     assert isinstance(recipe.blend["a"].messages[0], MessageTurn)
 
 
+def test_applicable_blend_round_trips_from_dict():
+    recipe = TrainingRecipe.from_dict(
+        {
+            "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"}],
+                },
+            },
+        }
+    )
+
+    assert recipe.select_from_applicable
+    assert recipe.blend["subtask"].requires == ["subtask"]
+
+
 def test_from_yaml_round_trips_through_load_recipe(tmp_path: Path):
     yaml_text = dedent(
         """

tests/datasets/test_language_render.py

@@ -176,6 +176,74 @@ def test_deterministic_blend_sampling():
     assert first == second
 
 
+def test_applicable_blend_filters_missing_bindings_before_weighted_selection():
+    recipe = TrainingRecipe(
+        select_from_applicable=True,
+        blend={
+            "missing": TrainingRecipe(
+                weight=1_000,
+                requires=["subtask"],
+                messages=[
+                    MessageTurn(
+                        role="assistant",
+                        content="${subtask}",
+                        stream="low_level",
+                        target=True,
+                    )
+                ],
+            ),
+            "action": TrainingRecipe(
+                weight=1,
+                messages=[MessageTurn(role="user", content="${task}", stream="low_level")],
+            ),
+        },
+    )
+
+    rendered = render_sample(
+        recipe=recipe,
+        persistent=[],
+        events=[],
+        t=0.0,
+        sample_idx=0,
+        task="pick the cup",
+    )
+
+    assert rendered["messages"] == [{"role": "user", "content": "pick the cup"}]
+    assert rendered["target_message_indices"] == []
+
+
+def test_applicable_blend_can_select_joint_bbox_subtask_target():
+    recipe = TrainingRecipe.from_yaml("src/lerobot/configs/recipes/g05_bbox_subtask.yaml")
+    persistent = [persistent_row("assistant", "grasp the cup", "subtask", 0.0)]
+    events = [
+        {
+            "role": "assistant",
+            "content": '{"detections": [{"label": "cup", "bbox": [1, 2, 3, 4]}]}',
+            "style": "vqa",
+            "camera": "observation.images.exterior",
+        }
+    ]
+
+    rendered = next(
+        candidate
+        for sample_idx in range(100)
+        if (
+            candidate := render_sample(
+                recipe=recipe,
+                persistent=persistent,
+                events=events,
+                t=0.0,
+                sample_idx=sample_idx,
+                task="pick the cup",
+            )
+        )
+        and candidate["target_message_indices"] == [1, 2]
+    )
+
+    assert rendered["messages"][1]["content"].startswith("BBoxJSON:")
+    assert rendered["messages"][2]["content"] == "Subtask: grasp the cup"
+
+
 def test_emitted_at_filters_vqa_by_camera():
     top = emitted_at(
         3.0,

tests/policies/g05/test_g05.py

@@ -492,6 +492,119 @@ def test_system2_training_target_is_forwarded_without_replacing_operator_task():
     assert prepared["samples"][0]["atomic_task"] == "Subtask: grasp the cup"
 
 
+def test_system2_recipe_subtask_target_selects_author_template():
+    policy = G05Policy(_config(predict_cot=True, runtime_system="system2"), backend=TinyG05Backend())
+    batch = _policy_batch("operator task")
+    batch["messages"] = [
+        [
+            {"role": "user", "content": "operator task"},
+            {"role": "assistant", "content": "Subtask: grasp the cup"},
+        ]
+    ]
+    batch["target_message_indices"] = [[1]]
+
+    sample = policy._prepare_author_batch(batch)["samples"][0]
+
+    assert sample["command"] == "operator task"
+    assert sample["prompt"] == "predict subtask"
+    assert sample["atomic_task"] == "Subtask: grasp the cup"
+    assert "<EOC><atomic_task_text>|Action: <EOV><action_action>|<eos>" in sample["template"]
+
+
+def test_system2_recipe_bbox_and_subtask_use_checkpoint_field_order():
+    policy = G05Policy(_config(predict_cot=True, runtime_system="system2"), backend=TinyG05Backend())
+    batch = _policy_batch("operator task")
+    batch["messages"] = [
+        [
+            {"role": "user", "content": "operator task"},
+            {
+                "role": "assistant",
+                "content": (
+                    'BBoxJSON: {"detections": [{"label": "cup", "bbox_format": "xyxy", '
+                    '"bbox": [20, 10, 100, 50]}]}'
+                ),
+            },
+            {"role": "assistant", "content": "Subtask: grasp the cup"},
+        ]
+    ]
+    batch["target_message_indices"] = [[1, 2]]
+    batch["g05_bbox_image_size"] = (100, 200)
+
+    sample = policy._prepare_author_batch(batch)["samples"][0]
+
+    assert sample["prompt"] == "predict bbox, subtask and action"
+    assert sample["bbox"] == "BBox: cup <loc0102><loc0102><loc0512><loc0512>"
+    assert sample["atomic_task"] == "Subtask: grasp the cup"
+    assert "<EOC><bbox_text>|<atomic_task_text>|Action:" in sample["template"]
+
+
+def test_system2_recipe_no_cot_branch_uses_action_only_training_template():
+    policy = G05Policy(_config(predict_cot=True, runtime_system="system2"), backend=TinyG05Backend())
+    batch = _policy_batch("operator task")
+    batch["messages"] = [[{"role": "user", "content": "operator task"}]]
+    batch["target_message_indices"] = [[]]
+
+    sample = policy._prepare_author_batch(batch)["samples"][0]
+
+    assert "prompt" not in sample
+    assert "atomic_task" not in sample
+    assert "<chat_assistant_prefix>Action: <EOV><EOC><action_action>|<eos>" in sample["template"]
+
+
+def test_recipe_preprocessor_resolves_lerobot_subtask_and_bbox_annotations():
+    pytest.importorskip("datasets", reason="recipe rendering requires lerobot[dataset]")
+    config = _config(
+        predict_cot=True,
+        runtime_system="system2",
+        recipe_path="recipes/g05_bbox_subtask.yaml",
+    )
+    preprocessor, _ = make_pre_post_processors(config)
+    policy = G05Policy(config, backend=TinyG05Backend())
+    raw = {
+        OBS_STATE: torch.zeros(7),
+        ACTION: torch.zeros(4, 7),
+        "observation.images.image": torch.zeros(3, 100, 200, dtype=torch.uint8),
+        "observation.images.wrist_image": torch.zeros(3, 100, 200, dtype=torch.uint8),
+        "task": "operator task",
+        "timestamp": torch.tensor(0.0),
+        "language_persistent": [
+            {
+                "role": "assistant",
+                "content": "grasp the cup",
+                "style": "subtask",
+                "timestamp": 0.0,
+                "camera": None,
+                "tool_calls": None,
+            }
+        ],
+        "language_events": [
+            {
+                "role": "assistant",
+                "content": (
+                    '{"detections": [{"label": "cup", "bbox_format": "xyxy", "bbox": [20, 10, 100, 50]}]}'
+                ),
+                "style": "vqa",
+                "camera": "observation.images.exterior",
+                "tool_calls": None,
+            }
+        ],
+    }
+
+    processed = next(
+        candidate
+        for sample_index in range(100)
+        if (candidate := preprocessor({**raw, "index": torch.tensor(sample_index)}))["target_message_indices"]
+        == [[1, 2]]
+    )
+    sample = policy._prepare_author_batch(processed)["samples"][0]
+
+    assert "language_persistent" not in processed
+    assert "language_events" not in processed
+    assert sample["bbox"] == "BBox: cup <loc0102><loc0102><loc0512><loc0512>"
+    assert sample["atomic_task"] == "Subtask: grasp the cup"
+    assert "<EOC><bbox_text>|<atomic_task_text>|Action:" in sample["template"]
+
+
 def test_author_inference_payload_synthesizes_required_dummy_action():
     policy = G05Policy(_config(), backend=TinyG05Backend())
     batch = _policy_batch()