fix(datasets): render flow-only low_level recipes instead of dropping them

A recipe whose only supervision is the action-expert flow loss (e.g.
`low_level_execution`: `user(${subtask})` with `stream: low_level` and no
`target` turn) was rejected at render time by `_render_message_recipe` and
`_validate_rendered`, both of which required at least one target turn.

The result: every blend draw of the flow-only recipe rendered to `None`,
`predict_actions` was never set, `run_flow` never fired, and the action
expert received no flow loss — leaving it at random init. Both gates now
also accept a `low_level`-stream turn as valid supervision.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/lerobot/datasets/language_render.py

@@ -376,7 +376,15 @@ def _render_message_recipe(
         if turn.target:
             target_indices.append(message_idx)
 
-    if not target_indices:
+    # A render is meaningful if it supervises *something*: either a
+    # text-CE target turn, or a ``low_level`` stream turn (flow / action
+    # supervision — e.g. the flow-only ``low_level_execution`` recipe,
+    # ``user(${subtask})`` with ``stream: low_level`` and no target).
+    # Without this, a flow-only recipe renders to ``None`` every time
+    # the blend draws it → ``predict_actions`` is never True → the
+    # action expert never receives a flow loss.
+    has_low_level = any(stream == "low_level" for stream in streams)
+    if not target_indices and not has_low_level:
         return None
 
     rendered = {
@@ -433,8 +441,13 @@ def _validate_rendered(rendered: RenderedMessages) -> None:
 
     if len(streams) != len(messages):
         raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
+    # Valid iff it supervises something: a text-CE target turn OR a
-        raise ValueError("Rendered samples must contain at least one target message.")
+    # ``low_level`` stream turn (flow / action supervision).
+    if not target_indices and not any(s == "low_level" for s in streams):
+        raise ValueError(
+            "Rendered samples must contain a target message or a "
+            "low_level-stream message."
+        )
     for idx in target_indices:
         if idx < 0 or idx >= len(messages):
             raise ValueError(f"Target message index {idx} is out of bounds.")

tests/datasets/test_language_render.py

@@ -370,6 +370,38 @@ def test_resolve_task_explicit_override_beats_rephrasings():
     assert rendered["messages"][0]["content"] == "explicit override wins"
 
 
+def test_flow_only_low_level_recipe_renders_without_target():
+    """Regression: a flow-only ``low_level`` recipe has no ``target`` turn —
+    its supervision is the action-expert flow loss, not text-CE. It must
+    still render (not ``None``), otherwise every blend draw of it is dropped
+    and the action expert never receives a flow loss."""
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(
+                role="user",
+                content="${subtask}",
+                stream="low_level",
+                if_present="subtask",
+            ),
+        ],
+        bindings={"subtask": "active_at(t, style=subtask)"},
+    )
+
+    rendered = render_sample(
+        recipe=recipe,
+        persistent=PERSISTENT,
+        events=[],
+        t=0.5,
+        sample_idx=0,
+        task="clean kitchen",
+    )
+
+    assert rendered is not None
+    assert rendered["messages"] == [{"role": "user", "content": "subtask 0"}]
+    assert rendered["message_streams"] == ["low_level"]
+    assert rendered["target_message_indices"] == []
+
+
 def test_canonical_recipe_can_render_low_level_branch():
     recipe = TrainingRecipe.from_yaml(Path("src/lerobot/configs/recipes/pi05_hirobot.yaml"))
     low_level = TrainingRecipe(blend={"low": recipe.blend["low_level_execution"]})