refactor(pi052): trim PR — remove say tool, debug gates, dead code; move runtime

Cleanup pass over the language-support PR to cut LOC and scope creep.

Removals:
- SayTool + tools/ package (registry, Tool protocol, [tools] extra) and the
  runtime's tool-dispatch path. Kept <say> training supervision and inference
  stripping so speech-annotated datasets still train.
- WeightedEpisodeAwareSampler + VQA oversampling wiring
  (_build_vqa_oversample_weights, vqa_target_fraction) — training uses plain
  EpisodeAwareSampler again.
- Debug env-gates PI052_DEBUG_TENSORS, PI052_SUBTASK_USE_TASK, EVAL_TASK_OVERRIDE.
- Dead code: broken _tp._DUMP_BUDGET block, unused imports (copy/Tensor,
  RevisionNotFoundError, LeRobotDataset, os), messages_for_vqa, steps.py shim
  (modeling imports pi052_adapter directly), duplicated _emit, builtins.type[T].

Moves:
- Policy-agnostic runtime -> src/lerobot/runtime/ (LanguageConditionedRuntime +
  adapter Protocol + state); pi052 keeps only its adapter + CLI. Tests -> tests/runtime/.

Other:
- Compacted verbose AI-authored comments/docstrings across pi052 (kept the
  hard-won DDP / barrier-timeout / reduce-max / VQA-routing notes).
- Relocated LM-head prediction debug helper to pi052/debug_utils.py.
- Fixed test_render_messages: assert task-fallback render (current behavior)
  instead of the stale no-op expectation.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/datasets/test_sampler.py

@@ -25,7 +25,7 @@ from datasets import Dataset  # noqa: E402
 from lerobot.datasets.io_utils import (
     hf_transform_to_torch,
 )
-from lerobot.datasets.sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler, compute_sampler_state
+from lerobot.datasets.sampler import EpisodeAwareSampler, compute_sampler_state
 
 
 def calculate_episode_data_index(hf_dataset: Dataset) -> dict[str, torch.Tensor]:
@@ -152,52 +152,6 @@ def test_partial_episode_drop_warns(caplog):
     assert "Episode 0" in caplog.text
 
 
-# --- WeightedEpisodeAwareSampler --------------------------------------------
-
-
-def test_weighted_sampler_respects_episode_drop_and_length():
-    """The episode-boundary frame filtering is applied before weighting,
-    and one epoch still yields ``len(indices)`` samples."""
-    # One episode, 10 frames; drop the last 2.
-    sampler = WeightedEpisodeAwareSampler([0], [10], frame_weights=torch.ones(10), drop_n_last_frames=2)
-    assert sampler.indices == list(range(8))
-    assert len(sampler) == 8
-    draws = list(sampler)
-    assert len(draws) == 8
-    # Dropped frames 8 and 9 must never be sampled.
-    assert all(d in set(range(8)) for d in draws)
-
-
-def test_weighted_sampler_oversamples_high_weight_frames():
-    """A heavily-weighted frame dominates the draws."""
-    torch.manual_seed(0)
-    # 100 frames, frame 7 is weighted 1000x.
-    weights = torch.ones(100)
-    weights[7] = 1000.0
-    sampler = WeightedEpisodeAwareSampler([0], [100], frame_weights=weights)
-    counts = {}
-    for _ in range(20):  # 20 epochs
-        for d in sampler:
-            counts[d] = counts.get(d, 0) + 1
-    total = sum(counts.values())
-    # Frame 7 should be the overwhelming majority of the 2000 draws.
-    assert counts.get(7, 0) / total > 0.9
-
-
-def test_weighted_sampler_zero_weights_fall_back_to_uniform():
-    """If every surviving frame has zero weight, sampling is uniform
-    rather than crashing."""
-    sampler = WeightedEpisodeAwareSampler([0], [6], frame_weights=torch.zeros(6))
-    draws = set(sampler)
-    assert draws.issubset(set(range(6)))
-    assert len(list(sampler)) == 6
-
-
-def test_weighted_sampler_rejects_short_weight_vector():
-    with pytest.raises(ValueError, match="frame_weights"):
-        WeightedEpisodeAwareSampler([0], [10], frame_weights=torch.ones(5))
-
-
 # --- seeded (seed, epoch) shuffling, resume, and state ---
 
 EPISODE_BOUNDS = ([0, 2, 3], [2, 3, 6])  # episodes of 2, 1 and 3 frames

tests/policies/pi052/test_pi052_runtime_adapter.py

@@ -1,7 +1,7 @@
 from types import SimpleNamespace
 
-from lerobot.policies.language_conditioned import RuntimeState
 from lerobot.policies.pi052.inference.pi052_adapter import PI052PolicyAdapter, split_plan_and_say
+from lerobot.runtime import RuntimeState
 
 
 def test_pi052_adapter_builds_recipe_prompts_from_runtime_state():
@@ -28,13 +28,11 @@ def test_pi052_adapter_builds_recipe_prompts_from_runtime_state():
     ]
 
 
-def test_pi052_adapter_parses_say_tool_calls_and_plan_text():
+def test_pi052_adapter_strips_say_markers_from_plan_text():
     adapter = PI052PolicyAdapter(policy=object())
     text = "Move to the sink. <say>heading to the sink</say>"
 
     assert split_plan_and_say(text) == ("Move to the sink.", "heading to the sink")
-    assert adapter.parse_tool_calls(text)[0].name == "say"
-    assert adapter.parse_tool_calls(text)[0].arguments == {"text": "heading to the sink"}
     assert adapter.plan_from_text(text) == "Move to the sink."
 
 
@@ -48,7 +46,6 @@ def test_pi052_runtime_cli_smoke_does_not_load_model(monkeypatch):
         "_load_policy_and_preprocessor",
         lambda policy_path, dataset_repo_id: (fake_policy, None, None, None),
     )
-    monkeypatch.setattr(runtime_cli, "_build_tools", lambda no_tts, tts_voice: {})
     monkeypatch.setattr(runtime_cli, "_run_repl", lambda runtime, initial_task, max_ticks: 0)
 
     assert runtime_cli.main(["--policy.path=fake", "--no_robot", "--task=clean", "--max_ticks=0"]) == 0

tests/processor/test_render_messages_processor.py

@@ -12,7 +12,9 @@ from lerobot.processor.render_messages_processor import RenderMessagesStep  # no
 from lerobot.types import TransitionKey  # noqa: E402
 
 
-def test_render_messages_step_noops_without_language_columns():
+def test_render_messages_step_renders_task_fallback_without_language_columns():
+    """No language columns + a task string → low-level task fallback render,
+    matching what the policy sees at eval time on unannotated observations."""
     recipe = TrainingRecipe(
         messages=[
             MessageTurn(role="user", content="${task}", stream="high_level"),
@@ -21,6 +23,24 @@ def test_render_messages_step_noops_without_language_columns():
     )
     transition = create_transition(complementary_data={"task": "do it"})
 
+    out = RenderMessagesStep(recipe)(transition)
+    data = out[TransitionKey.COMPLEMENTARY_DATA]
+
+    assert data["messages"] == [{"role": "user", "content": "do it"}]
+    assert data["message_streams"] == ["low_level"]
+    assert data["target_message_indices"] == []
+    assert data["task"] == "do it"
+
+
+def test_render_messages_step_noops_without_language_columns_or_task():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="${subtask}", stream="low_level", target=True),
+        ]
+    )
+    transition = create_transition(complementary_data={})
+
     assert RenderMessagesStep(recipe)(transition) == transition
 
 

tests/runtime/test_language_runtime.py

@@ -1,7 +1,6 @@
-from lerobot.policies.language_conditioned import (
+from lerobot.runtime import (
     LanguageConditionedRuntime,
     RuntimeState,
-    ToolCall,
     VQAResult,
 )
 
@@ -18,11 +17,7 @@ class FakeAdapter:
 
     def select_text(self, kind, observation, state, user_text=None):
         self.text_calls.append((kind, user_text))
-        return "new plan <say>ok</say>"
-
-    def parse_tool_calls(self, text):
-        assert text == "new plan <say>ok</say>"
-        return [ToolCall("say", {"text": "ok"})]
+        return "new plan"
 
     def answer_vqa(self, question, camera, observation, state):
         return VQAResult(answer=f"answer: {question}")
@@ -32,14 +27,6 @@ class FakeAdapter:
         state.set_context("subtask", "pick cup", label="subtask")
 
 
-class FakeTool:
-    def __init__(self):
-        self.calls = []
-
-    def call(self, args):
-        self.calls.append(args)
-
-
 def test_runtime_tick_updates_language_enqueues_and_dispatches_action():
     adapter = FakeAdapter()
     executed = []
@@ -59,24 +46,20 @@ def test_runtime_tick_updates_language_enqueues_and_dispatches_action():
     assert "  subtask: pick cup" in logs
 
 
-def test_runtime_handles_user_interjection_and_dispatches_tools():
+def test_runtime_handles_user_interjection():
     adapter = FakeAdapter()
-    tool = FakeTool()
     runtime = LanguageConditionedRuntime(
         policy_adapter=adapter,
         observation_provider=lambda: {"observation.state": 1},
-        tools={"say": tool},
     )
     runtime.set_task("clean")
     runtime.state.extra["recent_interjection"] = "please say ok"
     runtime.state.emit("user_interjection")
 
-    logs = runtime.step_once()
+    runtime.step_once()
 
     assert ("interjection", "please say ok") in adapter.text_calls
-    assert runtime.state.language_context["plan"] == "new plan <say>ok</say>"
-    assert tool.calls == [{"text": "ok"}]
-    assert "  speech: ok" in logs
+    assert runtime.state.language_context["plan"] == "new plan"
 
 
 def test_runtime_state_aliases_legacy_keys_to_language_context():