address review

docs/source/hil_data_collection.mdx

@@ -68,12 +68,12 @@ The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with
 
 ## Script
 
-Use `lerobot-rollout` with `--strategy.type=dagger` for HIL data collection. Toggle RTC with `--rtc.enabled=true`:
+Use `lerobot-rollout` with `--strategy.type=dagger` for HIL data collection. Select the inference backend with `--inference.type=sync|rtc`:
 
-| Mode                     | Flag                 | Models                |
+| Mode                     | Flag                   | Models                |
-| ------------------------ | -------------------- | --------------------- |
+| ------------------------ | ---------------------- | --------------------- |
-| Standard (default)       | _(no flag needed)_   | ACT, Diffusion Policy |
+| Standard (default)       | _(no flag needed)_     | ACT, Diffusion Policy |
-| Real-Time Chunking (RTC) | `--rtc.enabled=true` | Pi0, Pi0.5, SmolVLA   |
+| Real-Time Chunking (RTC) | `--inference.type=rtc` | Pi0, Pi0.5, SmolVLA   |
 
 ---
 
@@ -122,10 +122,10 @@ For models with high inference latency, enable RTC for smooth execution:
 
 ```bash
 lerobot-rollout --strategy.type=dagger \
-    --rtc.enabled=true \
+    --inference.type=rtc \
-    --rtc.execution_horizon=20 \
+    --inference.rtc.execution_horizon=20 \
-    --rtc.max_guidance_weight=5.0 \
+    --inference.rtc.max_guidance_weight=5.0 \
-    --rtc.prefix_attention_schedule=LINEAR \
+    --inference.rtc.prefix_attention_schedule=LINEAR \
     --robot.type=bi_openarm_follower \
     --robot.left_arm_config.port=can1 \
     --robot.left_arm_config.side=left \

docs/source/il_robots.mdx

@@ -542,4 +542,4 @@ The `--strategy.type` flag selects the execution mode:
 - `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
 - `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
 
-All strategies support `--rtc.enabled=true` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).
+All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/rtc.mdx

@@ -34,7 +34,7 @@ pip install -e ".[smolvla]"
 
 ### Using RTC with Pi0
 
-You can use `lerobot-rollout --strategy.type=base --rtc.enabled=true` for RTC deployment on real robots.
+You can use `lerobot-rollout --strategy.type=base --inference.type=rtc` for RTC deployment on real robots.
 The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
 
 ```python
@@ -140,9 +140,9 @@ The script generates a visualization of the denoising process, comparing standar
 lerobot-rollout \
     --strategy.type=base \
     --policy.path=${HF_USERNAME}/policy_repo_id \
-    --rtc.enabled=true \
+    --inference.type=rtc \
-    --rtc.execution_horizon=10 \
+    --inference.rtc.execution_horizon=10 \
-    --rtc.max_guidance_weight=10.0 \
+    --inference.rtc.max_guidance_weight=10.0 \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58FA0834591 \
     --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \

docs/source/unitree_g1.mdx

@@ -284,7 +284,7 @@ python examples/rtc/eval_with_real_robot.py \
   --task="task_description" \
   --duration=1000 \
   --fps=30 \
-  --rtc.enabled=true
+  --inference.type=rtc
 ```
 
 ---

src/lerobot/rollout/context.py

@@ -242,18 +242,25 @@ def build_rollout_context(
     }
     action_features_hw = robot.action_features
 
-    dataset_features = combine_feature_dicts(
+    # The action side is always needed: sync inference reads action names from
-        aggregate_pipeline_dataset_features(
+    # ``dataset_features[ACTION]`` to map policy tensors back to robot actions.
-            pipeline=teleop_action_processor,
+    action_dataset_features = aggregate_pipeline_dataset_features(
-            initial_features=create_initial_features(action=action_features_hw),
+        pipeline=teleop_action_processor,
-            use_videos=cfg.dataset.video if cfg.dataset else True,
+        initial_features=create_initial_features(action=action_features_hw),
-        ),
+        use_videos=cfg.dataset.video if cfg.dataset else True,
-        aggregate_pipeline_dataset_features(
+    )
+    # Observation-side aggregation only feeds the dataset schema; skip it for
+    # the base strategy to avoid running observation pipelines that may have
+    # dataset-specific dependencies.
+    if cfg.dataset is not None:
+        observation_dataset_features = aggregate_pipeline_dataset_features(
             pipeline=robot_observation_processor,
             initial_features=create_initial_features(observation=observation_features_hw),
-            use_videos=cfg.dataset.video if cfg.dataset else True,
+            use_videos=cfg.dataset.video,
-        ),
+        )
-    )
+        dataset_features = combine_feature_dicts(action_dataset_features, observation_dataset_features)
+    else:
+        dataset_features = action_dataset_features
     hw_features = hw_to_dataset_features(observation_features_hw, "observation")
     raw_action_keys = list(robot.action_features.keys())
     policy_action_names = getattr(policy_config, "action_feature_names", None)

src/lerobot/rollout/inference/factory.py

@@ -68,6 +68,9 @@ class SyncInferenceConfig(InferenceStrategyConfig):
 class RTCInferenceConfig(InferenceStrategyConfig):
     """Real-Time Chunking: async policy inference in a background thread."""
 
+    # ``RTCConfig`` is a small dataclass with default-only fields, so eagerly
+    # constructing one here costs nothing and keeps draccus' CLI surface flat
+    # (``--inference.rtc.execution_horizon=...`` etc.).  No need to lazy-init.
     rtc: RTCConfig = field(default_factory=RTCConfig)
     queue_threshold: int = 30
 

src/lerobot/rollout/inference/rtc.py

@@ -51,6 +51,15 @@ from .base import InferenceStrategy
 
 logger = logging.getLogger(__name__)
 
+# How long the RTC loop sleeps when paused, idle, or backpressured by a full queue.
+_RTC_IDLE_SLEEP_S: float = 0.01
+# Backoff between transient inference errors (per consecutive failure).
+_RTC_ERROR_RETRY_DELAY_S: float = 0.5
+# Consecutive transient errors tolerated before giving up and propagating shutdown.
+_RTC_MAX_CONSECUTIVE_ERRORS: int = 10
+# Hard timeout for joining the RTC thread on stop().
+_RTC_JOIN_TIMEOUT_S: float = 3.0
+
 
 # ---------------------------------------------------------------------------
 # RTC helpers (extracted from examples/rtc and examples/hil)
@@ -208,7 +217,7 @@ class RTCInferenceStrategy(InferenceStrategy):
         self._shutdown_event.set()
         self._policy_active.clear()
         if self._rtc_thread is not None and self._rtc_thread.is_alive():
-            self._rtc_thread.join(timeout=3.0)
+            self._rtc_thread.join(timeout=_RTC_JOIN_TIMEOUT_S)
             self._rtc_thread = None
 
     def pause(self) -> None:
@@ -252,17 +261,18 @@ class RTCInferenceStrategy(InferenceStrategy):
 
             warmup_required = max(1, self._compile_warmup_inferences) if self._use_torch_compile else 0
             inference_count = 0
+            consecutive_errors = 0
 
             while not self._shutdown_event.is_set():
                 if not self._policy_active.is_set():
-                    time.sleep(0.01)
+                    time.sleep(_RTC_IDLE_SLEEP_S)
                     continue
 
                 queue = self._action_queue
                 with self._obs_lock:
                     obs = self._obs_holder.get("obs")
                 if queue is None or obs is None:
-                    time.sleep(0.01)
+                    time.sleep(_RTC_IDLE_SLEEP_S)
                     continue
 
                 if queue.qsize() <= self._rtc_queue_threshold:
@@ -310,6 +320,7 @@ class RTCInferenceStrategy(InferenceStrategy):
                         new_delay = math.ceil(new_latency / time_per_chunk)
 
                         inference_count += 1
+                        consecutive_errors = 0
                         is_warmup = self._use_torch_compile and inference_count <= warmup_required
                         if is_warmup:
                             latency_tracker.reset()
@@ -329,11 +340,20 @@ class RTCInferenceStrategy(InferenceStrategy):
                         logger.debug("RTC inference latency=%.2fs, queue=%d", new_latency, queue.qsize())
 
                     except Exception as e:
-                        logger.error("RTC inference error: %s", e)
+                        consecutive_errors += 1
+                        logger.error(
+                            "RTC inference error (%d/%d): %s",
+                            consecutive_errors,
+                            _RTC_MAX_CONSECUTIVE_ERRORS,
+                            e,
+                        )
                         logger.debug(traceback.format_exc())
-                        time.sleep(0.5)
+                        if consecutive_errors >= _RTC_MAX_CONSECUTIVE_ERRORS:
+                            # Persistent failure: stop retrying and propagate shutdown.
+                            raise
+                        time.sleep(_RTC_ERROR_RETRY_DELAY_S)
                 else:
-                    time.sleep(0.01)
+                    time.sleep(_RTC_IDLE_SLEEP_S)
 
         except Exception as e:
             logger.error("Fatal error in RTC thread: %s", e)

src/lerobot/rollout/inference/sync.py

@@ -17,6 +17,7 @@
 from __future__ import annotations
 
 import logging
+from contextlib import nullcontext
 from copy import copy
 
 import torch
@@ -55,7 +56,7 @@ class SyncInferenceStrategy(InferenceStrategy):
         self._dataset_features = dataset_features
         self._ordered_action_keys = ordered_action_keys
         self._task = task
-        self._device = device or "cpu"
+        self._device = torch.device(device or "cpu")
         self._robot_type = robot_type
 
     def start(self) -> None:
@@ -72,16 +73,18 @@ class SyncInferenceStrategy(InferenceStrategy):
     def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
         if obs_frame is None:
             return None
+        # Shallow copy is intentional: the caller (`send_next_action`) builds
+        # ``obs_frame`` fresh per tick via ``build_dataset_frame``, so the
+        # tensor/array values are not shared with any other reader.
         observation = copy(obs_frame)
-        policy_device = torch.device(self._device)
+        autocast_ctx = (
-        with (
+            torch.autocast(device_type=self._device.type)
-            torch.inference_mode(),
+            if self._device.type == "cuda" and self._policy.config.use_amp
-            torch.autocast(device_type=policy_device.type)
+            else nullcontext()
-            if policy_device.type == "cuda" and self._policy.config.use_amp
+        )
-            else torch.inference_mode(),
+        with torch.inference_mode(), autocast_ctx:
-        ):
             observation = prepare_observation_for_inference(
-                observation, policy_device, self._task, self._robot_type
+                observation, self._device, self._task, self._robot_type
             )
             observation = self._preprocessor(observation)
             action = self._policy.select_action(observation)

src/lerobot/rollout/ring_buffer.py

@@ -19,6 +19,7 @@ from __future__ import annotations
 from collections import deque
 
 import numpy as np
+import torch
 
 
 class RolloutRingBuffer:
@@ -28,6 +29,12 @@ class RolloutRingBuffer:
     time (``max_frames``) and memory (``max_memory_bytes``).  When either
     limit is reached the oldest frames are evicted.
 
+    .. note::
+       This class is **single-threaded**.  ``append``/``drain``/``clear``
+       must all be called from the same thread (the rollout main loop).
+       Concurrent access from a background thread will corrupt
+       ``_current_bytes`` accounting.
+
     Parameters
     ----------
     max_seconds:
@@ -91,7 +98,11 @@ def _estimate_frame_bytes(frame: dict) -> int:
     """Rough byte estimate for a single frame dictionary."""
     total = 0
     for v in frame.values():
-        if isinstance(v, np.ndarray) or hasattr(v, "nbytes"):
+        if isinstance(v, torch.Tensor):
+            # ``torch.Tensor`` has no ``nbytes``; compute it explicitly so the
+            # memory cap is honoured even when frames hold unconverted tensors.
+            total += v.nelement() * v.element_size()
+        elif isinstance(v, np.ndarray) or hasattr(v, "nbytes"):
             total += v.nbytes
         elif isinstance(v, (int, float)):
             total += 8

src/lerobot/rollout/strategies/dagger.py

@@ -33,7 +33,7 @@ import contextlib
 import enum
 import logging
 import time
-from threading import Lock
+from threading import Event, Lock
 from typing import Any
 
 import numpy as np
@@ -90,14 +90,16 @@ class DAggerEvents:
         self._phase = DAggerPhase.AUTONOMOUS
         self._pending_transition: str | None = None
 
-        # Episode-level flags (written by keyboard, consumed by main loop)
+        # Episode-level flags written by keyboard/pedal threads, consumed by
-        self.exit_early: bool = False
+        # the main loop.  ``threading.Event`` gives us atomic set/clear/check
-        self.rerecord_episode: bool = False
+        # semantics without taking ``self._lock``.
-        self.stop_recording: bool = False
+        self.exit_early = Event()
+        self.rerecord_episode = Event()
+        self.stop_recording = Event()
 
-        # Reset-phase flags (simpler lifecycle, shared between threads)
+        # Reset-phase flags (simpler lifecycle, shared between threads).
-        self.in_reset: bool = False
+        self.in_reset = Event()
-        self.start_next_episode: bool = False
+        self.start_next_episode = Event()
 
     # -- Thread-safe phase access ------------------------------------------
 
@@ -140,8 +142,8 @@ class DAggerEvents:
         with self._lock:
             self._phase = DAggerPhase.AUTONOMOUS
             self._pending_transition = None
-        self.exit_early = False
+        self.exit_early.clear()
-        self.rerecord_episode = False
+        self.rerecord_episode.clear()
 
 
 # ---------------------------------------------------------------------------
@@ -198,24 +200,24 @@ def _reset_loop(
     """Reset period where the human repositions the environment."""
     logger.info("RESET — press any key to enable teleoperation")
 
-    events.in_reset = True
+    events.in_reset.set()
-    events.start_next_episode = False
+    events.start_next_episode.clear()
 
     obs = robot.get_observation()
     robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
     _teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
 
-    while not events.start_next_episode and not events.stop_recording:
+    while not events.start_next_episode.is_set() and not events.stop_recording.is_set():
         precise_sleep(0.05)
 
-    if events.stop_recording:
+    if events.stop_recording.is_set():
         return
 
-    events.start_next_episode = False
+    events.start_next_episode.clear()
     _teleop_disable_torque(teleop)
     logger.info("Teleop enabled — press any key to start episode")
 
-    while not events.start_next_episode and not events.stop_recording:
+    while not events.start_next_episode.is_set() and not events.stop_recording.is_set():
         loop_start = time.perf_counter()
         obs = robot.get_observation()
         action = teleop.get_action()
@@ -224,8 +226,8 @@ def _reset_loop(
         robot.send_action(robot_action_to_send)
         precise_sleep(1 / fps - (time.perf_counter() - loop_start))
 
-    events.in_reset = False
+    events.in_reset.clear()
-    events.start_next_episode = False
+    events.start_next_episode.clear()
     events.reset_for_episode()
 
 
@@ -242,16 +244,16 @@ def _init_dagger_keyboard(events: DAggerEvents):
 
     def on_press(key):
         try:
-            if events.in_reset:
+            if events.in_reset.is_set():
                 if (
                     key in [keyboard.Key.space, keyboard.Key.right]
                     or hasattr(key, "char")
                     and key.char == "c"
                 ):
-                    events.start_next_episode = True
+                    events.start_next_episode.set()
                 elif key == keyboard.Key.esc:
-                    events.stop_recording = True
+                    events.stop_recording.set()
-                    events.start_next_episode = True
+                    events.start_next_episode.set()
                 return
 
             phase = events.phase
@@ -275,15 +277,15 @@ def _init_dagger_keyboard(events: DAggerEvents):
 
             elif key == keyboard.Key.right:
                 logger.info("End episode")
-                events.exit_early = True
+                events.exit_early.set()
             elif key == keyboard.Key.left:
                 logger.info("Re-record episode")
-                events.rerecord_episode = True
+                events.rerecord_episode.set()
-                events.exit_early = True
+                events.exit_early.set()
             elif key == keyboard.Key.esc:
                 logger.info("Stop recording...")
-                events.stop_recording = True
+                events.stop_recording.set()
-                events.exit_early = True
+                events.exit_early.set()
         except Exception as e:
             logger.debug("Key error: %s", e)
 
@@ -301,8 +303,8 @@ def _dagger_pedal_callback(events: DAggerEvents):
     def on_press(code: str) -> None:
         if code not in _DAGGER_PEDAL_KEYS:
             return
-        if events.in_reset:
+        if events.in_reset.is_set():
-            events.start_next_episode = True
+            events.start_next_episode.set()
             return
         phase = events.phase
         if phase == DAggerPhase.CORRECTING:
@@ -362,22 +364,22 @@ class DAggerStrategy(RolloutStrategy):
         with VideoEncodingManager(dataset):
             try:
                 recorded = 0
-                while recorded < self.config.num_episodes and not events.stop_recording:
+                while recorded < self.config.num_episodes and not events.stop_recording.is_set():
                     log_say(f"Episode {dataset.num_episodes}", self.config.play_sounds)
 
                     self._run_episode(ctx)
 
-                    if events.rerecord_episode:
+                    if events.rerecord_episode.is_set():
                         log_say("Re-recording", self.config.play_sounds)
-                        events.rerecord_episode = False
+                        events.rerecord_episode.clear()
-                        events.exit_early = False
+                        events.exit_early.clear()
                         dataset.clear_episode_buffer()
                         continue
 
                     dataset.save_episode()
                     recorded += 1
 
-                    if recorded < self.config.num_episodes and not events.stop_recording:
+                    if recorded < self.config.num_episodes and not events.stop_recording.is_set():
                         _reset_loop(
                             ctx.hardware.robot_wrapper,
                             teleop,
@@ -448,8 +450,8 @@ class DAggerStrategy(RolloutStrategy):
         while timestamp < self.config.episode_time_s:
             loop_start = time.perf_counter()
 
-            if events.exit_early:
+            if events.exit_early.is_set():
-                events.exit_early = False
+                events.exit_early.clear()
                 break
 
             transition = events.consume_transition()

src/lerobot/rollout/strategies/highlight.py

@@ -118,6 +118,14 @@ class HighlightStrategy(RolloutStrategy):
                         action_frame = build_dataset_frame(features, action_dict, prefix=ACTION)
                         frame = {**obs_frame, **action_frame, "task": task_str}
 
+                        # NOTE: ``is_set()`` then ``clear()`` is not atomic
+                        # against the keyboard thread setting the flag again
+                        # in between — but that is benign: we lose at most one
+                        # toggle, processed on the next iteration.  The
+                        # ``_recording_live`` branch below is reached in the
+                        # SAME iteration after ``clear()`` runs, so a frame
+                        # finalised by ``save_episode()`` is never re-added to
+                        # the next episode.
                         if self._save_requested.is_set():
                             self._save_requested.clear()
                             if not self._recording_live.is_set():

src/lerobot/rollout/strategies/sentry.py

@@ -108,10 +108,18 @@ class SentryStrategy(RolloutStrategy):
                         obs_frame = build_dataset_frame(features, obs_processed, prefix=OBS_STR)
                         action_frame = build_dataset_frame(features, action_dict, prefix=ACTION)
                         frame = {**obs_frame, **action_frame, "task": task_str}
+                        # ``add_frame`` writes to the in-progress episode buffer; the
+                        # background pusher only ever touches *finalised* episode
+                        # artifacts on disk.  The two operate on disjoint state, so
+                        # ``add_frame`` does not need ``_episode_lock``.
                         dataset.add_frame(frame)
 
                     elapsed = time.perf_counter() - episode_start
                     if elapsed >= self.config.episode_duration_s:
+                        # ``save_episode`` finalises the in-progress episode and
+                        # flushes it to disk; ``_episode_lock`` serialises this with
+                        # ``push_to_hub`` (run in the background executor) so the
+                        # pusher never reads a half-written episode.
                         with self._episode_lock:
                             dataset.save_episode()
                         episodes_since_push += 1