chore(datasets): trim sampler comment and drop duplicate tests

Remove the verbose dataloader-guard comment and the two EpisodeAwareSampler tests
that duplicated existing validation/warning coverage (no coverage loss).

Co-authored-by: Cursor <cursoragent@cursor.com>

docs/source/il_robots.mdx

@@ -647,5 +647,6 @@ The `--strategy.type` flag selects the execution mode:
 - `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
 - `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))
+- `episodic`: Episode-oriented policy recording with reset phases between episodes
 
 All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -157,6 +157,44 @@ Foot pedal input is also supported via `--strategy.input_device=pedal`. Configur
 | `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
 | `--teleop.type`                      | **Required.** Teleoperator type                         |
 
+### Episodic (`--strategy.type=episodic`)
+
+Episode-oriented recording that mirrors the behavior of `lerobot-record`. The policy drives the robot for each episode; an optional teleoperator can drive the robot during the reset phase between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=episodic \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/my_eval_data \
+    --dataset.num_episodes=20 \
+    --dataset.episode_time_s=30 \
+    --dataset.reset_time_s=10 \
+    --dataset.single_task="Pick up the red cube"
+```
+
+Teleop is optional — if omitted the robot holds its position during the reset phase.
+
+**Keyboard controls:**
+
+| Key         | Action                           |
+| ----------- | -------------------------------- |
+| `→` (right) | End the current episode early    |
+| `←` (left)  | Discard episode and re-record it |
+| `ESC`       | Stop the recording session       |
+
+| Flag                                            | Description                                                                |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
+| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
+| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
+| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
+
 ---
 
 ## Inference Backends

pyproject.toml

@@ -216,7 +216,7 @@ robometer = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]", "lerobot
 topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
-hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.14,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 vla_jepa = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[qwen-vl-utils-dep]"]
 
 # Features
@@ -231,9 +231,9 @@ video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
 # NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["lerobot[dataset]", "gym-aloha>=0.1.4,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.4,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
 # NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
 # is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI

src/lerobot/common/control_utils.py

@@ -18,6 +18,7 @@ from __future__ import annotations
 # Utilities
 ########################################################################################
 import logging
+import time
 import traceback
 from contextlib import nullcontext
 from copy import copy
@@ -243,3 +244,72 @@ def sanity_check_dataset_robot_compatibility(
         raise ValueError(
             "Dataset metadata compatibility check failed with mismatches:\n" + "\n".join(mismatches)
         )
+
+
+########################################################################################
+# Teleoperator smooth handover helpers
+# NOTE(Maxime): These functions use minimal type hints to maintain compatibility with utils
+# being a root module.
+########################################################################################
+
+
+def teleop_supports_feedback(teleop) -> bool:
+    """Return True when the teleop can receive position feedback (is actuated).
+
+    Actuated teleops (e.g. SO-101, OpenArmMini) have non-empty ``feedback_features``
+    and expose ``enable_torque`` / ``disable_torque`` motor-control methods.
+
+    TODO(Maxime): See if it is possible to unify this interface across teleops instead of duck-typing.
+    """
+    return (
+        bool(teleop.feedback_features)
+        and hasattr(teleop, "disable_torque")
+        and hasattr(teleop, "enable_torque")
+    )
+
+
+def teleop_smooth_move_to(teleop, target_pos: dict, duration_s: float = 2.0, fps: int = 30) -> None:
+    """Smoothly move an actuated teleop to ``target_pos`` via linear interpolation.
+
+    Requires the teleoperator to support feedback (i.e. have non-empty
+    ``feedback_features`` and implement ``disable_torque`` / ``enable_torque``).
+
+    ``target_pos`` is expected to be in the teleop's action/feedback key space.
+    For homogeneous setups (e.g. SO-101 leader + SO-101 follower) this matches
+    the robot action key space directly.
+
+    TODO(Maxime): This blocks up to ``duration_s`` seconds; during this time the
+    follower robot does not receive new actions, which could be an issue on LeKiwi.
+    """
+    teleop.enable_torque()
+    current = teleop.get_action()
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {
+            k: current[k] * (1 - t) + target_pos[k] * t if k in target_pos else current[k] for k in current
+        }
+        teleop.send_feedback(interp)
+        time.sleep(1 / fps)
+
+
+def follower_smooth_move_to(
+    robot, current: dict, target: dict, duration_s: float = 1.0, fps: int = 30
+) -> None:
+    """Smoothly move the follower robot from ``current`` to ``target`` action.
+
+    Used when the teleop is non-actuated: instead of driving the leader arm to
+    the follower, the follower is brought to the teleop's current pose so the
+    robot meets the operator's hand rather than jumping to it on the first frame.
+
+    Both ``current`` and ``target`` must be in the robot action key space
+    (i.e. the output of ``robot_action_processor``).
+    """
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {k: current[k] * (1 - t) + target[k] * t if k in target else current[k] for k in current}
+        robot.send_action(interp)
+        time.sleep(1 / fps)

src/lerobot/common/train_utils.py

@@ -49,8 +49,19 @@ def get_step_checkpoint_dir(output_dir: Path, total_steps: int, step: int) -> Pa
     return output_dir / CHECKPOINTS_DIR / step_identifier
 
 
-def save_training_step(step: int, save_dir: Path) -> None:
-    write_json({"step": step}, save_dir / TRAINING_STEP)
+def save_training_step(
+    step: int, save_dir: Path, num_processes: int | None = None, batch_size: int | None = None
+) -> None:
+    state: dict = {"step": step}
+    # num_processes and batch_size are recorded so a resumed run can detect a changed world size or
+    # batch size: the sampler's resume offset is computed from the (num_processes, batch_size) that
+    # produced `step`, since both scale how many sampler positions a step consumes (see
+    # compute_sampler_state).
+    if num_processes is not None:
+        state["num_processes"] = num_processes
+    if batch_size is not None:
+        state["batch_size"] = batch_size
+    write_json(state, save_dir / TRAINING_STEP)
 
 
 def load_training_step(save_dir: Path) -> int:
@@ -58,6 +69,16 @@ def load_training_step(save_dir: Path) -> int:
     return training_step["step"]
 
 
+def load_training_num_processes(checkpoint_dir: Path) -> int | None:
+    """World size recorded at checkpoint time, or None for checkpoints written before it was stored."""
+    return load_json(checkpoint_dir / TRAINING_STATE_DIR / TRAINING_STEP).get("num_processes")
+
+
+def load_training_batch_size(checkpoint_dir: Path) -> int | None:
+    """Per-process batch size recorded at checkpoint time, or None for older checkpoints."""
+    return load_json(checkpoint_dir / TRAINING_STATE_DIR / TRAINING_STEP).get("batch_size")
+
+
 def update_last_checkpoint(checkpoint_dir: Path) -> Path:
     last_checkpoint_dir = checkpoint_dir.parent / LAST_CHECKPOINT_LINK
     if last_checkpoint_dir.is_symlink():
@@ -75,6 +96,8 @@ def save_checkpoint(
     scheduler: LRScheduler | None = None,
     preprocessor: PolicyProcessorPipeline | None = None,
     postprocessor: PolicyProcessorPipeline | None = None,
+    num_processes: int | None = None,
+    batch_size: int | None = None,
 ) -> None:
     """This function creates the following directory structure:
 
@@ -100,6 +123,10 @@ def save_checkpoint(
         scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
         preprocessor: The preprocessor/pipeline to save. Defaults to None.
         postprocessor: The postprocessor/pipeline to save. Defaults to None.
+        num_processes (int | None, optional): Distributed world size to record for sample-exact
+            resume. Defaults to None (not recorded).
+        batch_size (int | None, optional): Per-process batch size to record for sample-exact
+            resume. Defaults to None (not recorded).
     """
     pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
     policy.save_pretrained(pretrained_dir)
@@ -112,7 +139,9 @@ def save_checkpoint(
         preprocessor.save_pretrained(pretrained_dir)
     if postprocessor is not None:
         postprocessor.save_pretrained(pretrained_dir)
-    save_training_state(checkpoint_dir, step, optimizer, scheduler)
+    save_training_state(
+        checkpoint_dir, step, optimizer, scheduler, num_processes=num_processes, batch_size=batch_size
+    )
 
 
 def save_training_state(
@@ -120,6 +149,8 @@ def save_training_state(
     train_step: int,
     optimizer: Optimizer | None = None,
     scheduler: LRScheduler | None = None,
+    num_processes: int | None = None,
+    batch_size: int | None = None,
 ) -> None:
     """
     Saves the training step, optimizer state, scheduler state, and rng state.
@@ -131,10 +162,12 @@ def save_training_state(
             Defaults to None.
         scheduler (LRScheduler | None, optional): The scheduler from which to save the state_dict.
             Defaults to None.
+        num_processes (int | None, optional): Distributed world size to record. Defaults to None.
+        batch_size (int | None, optional): Per-process batch size to record. Defaults to None.
     """
     save_dir = checkpoint_dir / TRAINING_STATE_DIR
     save_dir.mkdir(parents=True, exist_ok=True)
-    save_training_step(train_step, save_dir)
+    save_training_step(train_step, save_dir, num_processes=num_processes, batch_size=batch_size)
     save_rng_state(save_dir)
     if optimizer is not None:
         save_optimizer_state(optimizer, save_dir)

src/lerobot/datasets/__init__.py

@@ -50,7 +50,7 @@ from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
 from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
-from .sampler import EpisodeAwareSampler
+from .sampler import EpisodeAwareSampler, compute_sampler_state
 from .streaming_dataset import StreamingLeRobotDataset
 from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
 from .video_utils import VideoEncodingManager
@@ -82,6 +82,7 @@ __all__ = [
     "aggregate_stats",
     "convert_image_to_video_dataset",
     "create_initial_features",
+    "compute_sampler_state",
     "create_lerobot_dataset_card",
     "column_for_style",
     "delete_episodes",

src/lerobot/datasets/sampler.py

@@ -14,14 +14,36 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import math
 from collections.abc import Iterator
 
+import numpy as np
 import torch
 
 logger = logging.getLogger(__name__)
 
 
 class EpisodeAwareSampler:
+    """Sampler over episode frames that stores only per-episode boundaries.
+
+    Logical positions map to frame indices on the fly (O(num_episodes) construction memory)
+    instead of materializing a Python list of every frame index.
+
+    Each epoch is shuffled with a `torch.randperm` seeded from `(seed, epoch)`, so the data order
+    is a pure function of `(seed, epoch)`: it reproduces on every rank without synchronizing the
+    global RNG (no `generator` to sync across distributed ranks), and `state_dict` /
+    `load_state_dict` resume a run sample-exactly by regenerating the epoch's permutation and
+    continuing from the saved offset. Each call to `__iter__` advances the epoch. During a
+    resumed epoch, `__len__` still reports the full length.
+
+    Epoch advancement: `__iter__` eagerly advances the epoch, and `set_epoch` / `load_state_dict`
+    set it explicitly. Within a single run callers should rely on exactly one of these mechanisms,
+    not both: advancing the epoch by hand *and* letting `__iter__` auto-advance over the same
+    iterations would skip or repeat epochs. The training loop drives it purely through `__iter__`
+    (via `cycle`); `set_epoch` / `load_state_dict` are used only to (re)position before iteration
+    starts (e.g. on resume or in tests).
+    """
+
     def __init__(
         self,
         dataset_from_indices: list[int],
@@ -30,57 +52,125 @@ class EpisodeAwareSampler:
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
         shuffle: bool = False,
+        seed: int = 0,
     ):
-        """Sampler that optionally incorporates episode boundary information.
-
+        """
         Args:
-            dataset_from_indices: List of indices containing the start of each episode in the dataset.
-            dataset_to_indices: List of indices containing the end of each episode in the dataset.
-            episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
-                                    Assumes that episodes are indexed from 0 to N-1.
-            drop_n_first_frames: Number of frames to drop from the start of each episode.
-            drop_n_last_frames: Number of frames to drop from the end of each episode.
+            dataset_from_indices: Start index of each episode in the dataset.
+            dataset_to_indices: End index of each episode in the dataset.
+            episode_indices_to_use: Episode indices to use; None means all.
+            drop_n_first_frames: Frames to drop from the start of each episode.
+            drop_n_last_frames: Frames to drop from the end of each episode.
             shuffle: Whether to shuffle the indices.
+            seed: Seed the permutation is derived from (together with the epoch).
         """
         if drop_n_first_frames < 0:
             raise ValueError(f"drop_n_first_frames must be >= 0, got {drop_n_first_frames}")
         if drop_n_last_frames < 0:
             raise ValueError(f"drop_n_last_frames must be >= 0, got {drop_n_last_frames}")
 
-        indices = []
-        for episode_idx, (start_index, end_index) in enumerate(
-            zip(dataset_from_indices, dataset_to_indices, strict=True)
-        ):
-            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
-                ep_length = end_index - start_index
-                if drop_n_first_frames + drop_n_last_frames >= ep_length:
-                    logger.warning(
-                        "Episode %d has %d frames but drop_n_first_frames=%d and "
-                        "drop_n_last_frames=%d removes all frames. Skipping.",
-                        episode_idx,
-                        ep_length,
-                        drop_n_first_frames,
-                        drop_n_last_frames,
-                    )
-                    continue
-                indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))
+        from_indices = np.asarray(dataset_from_indices, dtype=np.int64)
+        to_indices = np.asarray(dataset_to_indices, dtype=np.int64)
+        if from_indices.shape != to_indices.shape:
+            raise ValueError(
+                f"dataset_from_indices and dataset_to_indices must have the same length, "
+                f"got {len(from_indices)} and {len(to_indices)}"
+            )
 
-        if not indices:
+        used = np.ones(len(from_indices), dtype=bool)
+        if episode_indices_to_use is not None:
+            used = np.zeros(len(from_indices), dtype=bool)
+            used[np.asarray(episode_indices_to_use, dtype=np.int64)] = True
+
+        starts = from_indices + drop_n_first_frames
+        lengths = to_indices - drop_n_last_frames - starts
+        for episode_idx in np.flatnonzero(used & (lengths <= 0)):
+            logger.warning(
+                "Episode %d has %d frames but drop_n_first_frames=%d and "
+                "drop_n_last_frames=%d removes all frames. Skipping.",
+                episode_idx,
+                to_indices[episode_idx] - from_indices[episode_idx],
+                drop_n_first_frames,
+                drop_n_last_frames,
+            )
+        used &= lengths > 0
+        if not used.any():
             raise ValueError(
                 "No valid frames remain after applying drop_n_first_frames and drop_n_last_frames. "
                 "All episodes were either filtered out or had too few frames."
             )
 
-        self.indices = indices
+        self._starts = starts[used]
+        self._cum_lengths = np.cumsum(lengths[used])
+        self._num_frames = int(self._cum_lengths[-1])
         self.shuffle = shuffle
+        self.seed = seed
+        self._epoch = 0
+        self._start_index = 0
+
+    @property
+    def indices(self) -> list[int]:
+        """Materialized frame indices in unshuffled order; O(num_frames), introspection only."""
+        return [self._frame_index(k) for k in range(self._num_frames)]
+
+    def set_epoch(self, epoch: int) -> None:
+        self._epoch = epoch
+
+    def state_dict(self) -> dict:
+        return {"epoch": self._epoch, "start_index": self._start_index}
+
+    def load_state_dict(self, state: dict) -> None:
+        self._epoch = state["epoch"]
+        self._start_index = state["start_index"]
+
+    def _epoch_generator(self, epoch: int) -> torch.Generator:
+        # Derive a per-epoch seed from (seed, epoch) so the permutation is a pure function of both
+        # and reproduces identically on every rank without touching the global RNG.
+        epoch_seed = int(np.random.SeedSequence([self.seed, epoch]).generate_state(1, dtype=np.uint64)[0])
+        return torch.Generator().manual_seed(epoch_seed)
+
+    def _frame_index(self, position: int) -> int:
+        episode = int(np.searchsorted(self._cum_lengths, position, side="right"))
+        position_in_episode = position - (int(self._cum_lengths[episode - 1]) if episode > 0 else 0)
+        return int(self._starts[episode]) + position_in_episode
 
     def __iter__(self) -> Iterator[int]:
+        # Advance epoch state eagerly, not on first consumption of the generator.
+        epoch, start = self._epoch, self._start_index
+        self._epoch += 1
+        self._start_index = 0
+        return self._iter_epoch(epoch, start)
+
+    def _iter_epoch(self, epoch: int, start: int) -> Iterator[int]:
         if self.shuffle:
-            for i in torch.randperm(len(self.indices)):
-                yield self.indices[i]
+            order = torch.randperm(self._num_frames, generator=self._epoch_generator(epoch))
+            for k in range(start, self._num_frames):
+                yield self._frame_index(int(order[k]))
         else:
-            for i in self.indices:
-                yield i
+            for k in range(start, self._num_frames):
+                yield self._frame_index(k)
 
     def __len__(self) -> int:
-        return len(self.indices)
+        return self._num_frames
+
+
+def compute_sampler_state(step: int, num_frames: int, batch_size: int, num_processes: int) -> dict:
+    """Map an optimization step to an `EpisodeAwareSampler` state for sample-exact resume.
+
+    Under accelerate's batch sharding, one step consumes `batch_size * num_processes` sampler
+    positions and each rank sees `ceil(ceil(num_frames / batch_size) / num_processes)` batches
+    per epoch (`even_batches` padding included). The start index provably stays below
+    `num_frames`; the `min` is defensive.
+
+    Assumptions (resume is only sample-exact when they hold):
+        - `num_processes` and `batch_size` match the run that wrote the checkpoint. Both scale how
+          many positions a step consumes, so the epoch/offset are wrong if either changed. The
+          caller passes the checkpoint's `num_processes` and `batch_size` and warns on a mismatch.
+        - accelerate uses `even_batches=True` (its default). The `ceil(... / num_processes)` term
+          mirrors that padding; with `even_batches=False` the per-epoch batch count differs and
+          the boundary is off.
+    """
+    batches_per_epoch = math.ceil(math.ceil(num_frames / batch_size) / num_processes)
+    epoch, batches_into_epoch = divmod(step, batches_per_epoch)
+    start_index = min(batches_into_epoch * batch_size * num_processes, num_frames)
+    return {"epoch": epoch, "start_index": start_index}

src/lerobot/processor/pipeline.py

@@ -32,7 +32,6 @@ from __future__ import annotations
 
 import importlib
 import json
-import os
 import re
 from abc import ABC, abstractmethod
 from collections.abc import Callable, Iterable, Sequence
@@ -281,6 +280,11 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
     before_step_hooks: list[Callable[[int, EnvTransition], None]] = field(default_factory=list, repr=False)
     after_step_hooks: list[Callable[[int, EnvTransition], None]] = field(default_factory=list, repr=False)
+    _serialized_state_filenames: tuple[str | None, ...] | None = field(
+        default=None,
+        init=False,
+        repr=False,
+    )
 
     def __call__(self, data: TInput) -> TOutput:
         """Processes input data through the full pipeline.
@@ -338,30 +342,108 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             transition = processor_step(transition)
             yield transition
 
-    def _save_pretrained(self, save_directory: Path, **kwargs):
-        """Internal method to comply with `HubMixin`'s saving mechanism.
+    def _get_sanitized_name(self) -> str:
+        """Return a filename-safe version of the pipeline name.
 
-        This method does the actual saving work and is called by HubMixin.save_pretrained.
+        Returns:
+            The lower-cased pipeline name with non-alphanumeric characters replaced by underscores.
         """
-        config_filename = kwargs.pop("config_filename", None)
+        return re.sub(r"[^a-zA-Z0-9_]", "_", self.name.lower())
 
-        # Sanitize the pipeline name to create a valid filename prefix.
-        sanitized_name = re.sub(r"[^a-zA-Z0-9_]", "_", self.name.lower())
+    @staticmethod
+    def _get_state_filename(
+        *,
+        step_index: int,
+        registry_name: str | None,
+        sanitized_name: str,
+    ) -> str:
+        """Return the safetensors filename for one stateful processor step.
 
-        if config_filename is None:
-            config_filename = f"{sanitized_name}.json"
+        Args:
+            step_index: The index of the processor step in this pipeline.
+            registry_name: The registered processor step name, if available.
+            sanitized_name: The filename-safe pipeline name.
 
-        config: dict[str, Any] = {
+        Returns:
+            The state filename used by the existing disk serialization format.
+        """
+        if registry_name:
+            return f"{sanitized_name}_step_{step_index}_{registry_name}.safetensors"
+
+        return f"{sanitized_name}_step_{step_index}.safetensors"
+
+    @staticmethod
+    def _get_state_key(state_filename: str) -> str:
+        """Return the in-memory state key for a serialized state filename.
+
+        Args:
+            state_filename: The `.safetensors` filename from the serialized config.
+
+        Returns:
+            The state key used by the in-memory pipeline state dictionary.
+        """
+        return state_filename.removesuffix(".safetensors")
+
+    @staticmethod
+    def _get_state_filenames_from_config(loaded_config: dict[str, Any]) -> tuple[str | None, ...]:
+        """Return serialized state filenames in step order.
+
+        Args:
+            loaded_config: A validated processor pipeline config.
+
+        Returns:
+            A tuple containing each step's serialized state filename, or None for stateless steps.
+        """
+        return tuple(step_entry.get("state_file") for step_entry in loaded_config["steps"])
+
+    def _get_state_filenames_for_loading(self) -> tuple[str | None, ...]:
+        """Return expected state filenames in step order for `load_state_dict()`.
+
+        Returns:
+            The preserved serialized state filenames when available, otherwise filenames derived from
+            current non-empty step state.
+        """
+        if self._serialized_state_filenames is not None and len(self._serialized_state_filenames) == len(
+            self.steps
+        ):
+            return self._serialized_state_filenames
+
+        sanitized_name = self._get_sanitized_name()
+        state_filenames: list[str | None] = []
+
+        for step_index, processor_step in enumerate(self.steps):
+            step_state_dict = processor_step.state_dict()
+            if not step_state_dict:
+                state_filenames.append(None)
+                continue
+
+            registry_name = getattr(processor_step.__class__, "_registry_name", None)
+            state_filenames.append(
+                self._get_state_filename(
+                    step_index=step_index,
+                    registry_name=registry_name,
+                    sanitized_name=sanitized_name,
+                )
+            )
+
+        return tuple(state_filenames)
+
+    def get_config(self) -> dict[str, Any]:
+        """Return the JSON-serializable pipeline configuration.
+
+        Returns:
+            A dictionary with the same content that `save_pretrained()` writes as JSON.
+        """
+        sanitized_name = self._get_sanitized_name()
+        pipeline_config: dict[str, Any] = {
             "name": self.name,
             "steps": [],
         }
 
-        # Iterate through each step to build its configuration entry.
         for step_index, processor_step in enumerate(self.steps):
             registry_name = getattr(processor_step.__class__, "_registry_name", None)
-
             step_entry: dict[str, Any] = {}
-            # Prefer registry name for portability, otherwise fall back to full class path.
+
             if registry_name:
                 step_entry["registry_name"] = registry_name
             else:
@@ -369,31 +451,110 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                     f"{processor_step.__class__.__module__}.{processor_step.__class__.__name__}"
                 )
 
-            # Save step configuration if `get_config` is implemented.
-            if hasattr(processor_step, "get_config"):
-                step_entry["config"] = processor_step.get_config()
+            step_entry["config"] = processor_step.get_config()
 
-            # Save step state if `state_dict` is implemented and returns a non-empty dict.
-            if hasattr(processor_step, "state_dict"):
-                state = processor_step.state_dict()
-                if state:
-                    # Clone tensors to avoid modifying the original state.
-                    cloned_state = {key: tensor.clone() for key, tensor in state.items()}
+            step_state_dict = processor_step.state_dict()
+            if step_state_dict:
+                step_entry["state_file"] = self._get_state_filename(
+                    step_index=step_index,
+                    registry_name=registry_name,
+                    sanitized_name=sanitized_name,
+                )
 
-                    # Create a unique filename for the state file.
-                    if registry_name:
-                        state_filename = f"{sanitized_name}_step_{step_index}_{registry_name}.safetensors"
-                    else:
-                        state_filename = f"{sanitized_name}_step_{step_index}.safetensors"
+            pipeline_config["steps"].append(step_entry)
 
-                    save_file(cloned_state, os.path.join(str(save_directory), state_filename))
-                    step_entry["state_file"] = state_filename
+        return pipeline_config
 
-            config["steps"].append(step_entry)
+    def state_dict(self) -> dict[str, dict[str, torch.Tensor]]:
+        """Return pipeline state tensors grouped by state key.
 
-        # Write the main configuration JSON file.
-        with open(os.path.join(str(save_directory), config_filename), "w") as file_pointer:
-            json.dump(config, file_pointer, indent=2)
+        Returns:
+            A dictionary mapping suffixless state keys to cloned step state dictionaries.
+        """
+        sanitized_name = self._get_sanitized_name()
+        pipeline_state_dict: dict[str, dict[str, torch.Tensor]] = {}
+
+        for step_index, processor_step in enumerate(self.steps):
+            step_state_dict = processor_step.state_dict()
+            if not step_state_dict:
+                continue
+
+            registry_name = getattr(processor_step.__class__, "_registry_name", None)
+            state_filename = self._get_state_filename(
+                step_index=step_index,
+                registry_name=registry_name,
+                sanitized_name=sanitized_name,
+            )
+            state_key = self._get_state_key(state_filename)
+            pipeline_state_dict[state_key] = {
+                tensor_name: tensor.clone() for tensor_name, tensor in step_state_dict.items()
+            }
+
+        return pipeline_state_dict
+
+    def load_state_dict(
+        self,
+        state_dict: dict[str, dict[str, torch.Tensor]],
+    ) -> None:
+        """Load pipeline state tensors into the existing steps.
+
+        Args:
+            state_dict: A dictionary mapping suffixless state keys to step state dictionaries.
+
+        Raises:
+            KeyError: If loading finds missing expected state or unexpected extra state.
+        """
+        expected_state_filenames = self._get_state_filenames_for_loading()
+        used_state_keys: set[str] = set()
+
+        for step_index, (processor_step, state_filename) in enumerate(
+            zip(self.steps, expected_state_filenames, strict=True)
+        ):
+            if state_filename is None:
+                continue
+
+            state_key = self._get_state_key(state_filename)
+            if state_key not in state_dict:
+                raise KeyError(
+                    f"Missing state key '{state_key}' for processor step {step_index}. "
+                    f"Available state keys: {sorted(state_dict.keys())}"
+                )
+
+            processor_step.load_state_dict(state_dict[state_key])
+            used_state_keys.add(state_key)
+
+        unexpected_state_keys = set(state_dict) - used_state_keys
+        if unexpected_state_keys:
+            expected_state_key_set = {
+                self._get_state_key(state_filename)
+                for state_filename in expected_state_filenames
+                if state_filename is not None
+            }
+            raise KeyError(
+                f"Unexpected processor state keys: {sorted(unexpected_state_keys)}. "
+                f"Expected state keys: {sorted(expected_state_key_set)}"
+            )
+
+    def _save_pretrained(self, save_directory: Path, **kwargs) -> None:
+        """Internal method to comply with `HubMixin`'s saving mechanism.
+
+        This method does the actual saving work and is called by HubMixin.save_pretrained.
+        """
+        config_filename = kwargs.pop("config_filename", None)
+        sanitized_name = self._get_sanitized_name()
+
+        if config_filename is None:
+            config_filename = f"{sanitized_name}.json"
+
+        pipeline_config = self.get_config()
+        pipeline_state_dict = self.state_dict()
+
+        for state_key, step_state_dict in pipeline_state_dict.items():
+            state_filename = f"{state_key}.safetensors"
+            save_file(step_state_dict, save_directory / state_filename)
+
+        with open(save_directory / config_filename, "w") as file_pointer:
+            json.dump(pipeline_config, file_pointer, indent=2)
 
     def save_pretrained(
         self,
@@ -577,12 +738,54 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         cls._validate_overrides_used(validated_overrides, loaded_config)
 
         # 5. Construct and return the final pipeline instance
-        return cls(
+        pipeline = cls(
             steps=steps,
             name=loaded_config.get("name", "DataProcessorPipeline"),
             to_transition=to_transition or cast(Callable[[TInput], EnvTransition], batch_to_transition),
             to_output=to_output or cast(Callable[[EnvTransition], TOutput], transition_to_batch),
         )
+        pipeline._serialized_state_filenames = cls._get_state_filenames_from_config(loaded_config)
+        return pipeline
+
+    @classmethod
+    def from_config(
+        cls,
+        config: dict[str, Any],
+        *,
+        state_dict: dict[str, dict[str, torch.Tensor]] | None = None,
+        overrides: dict[str, Any] | None = None,
+        to_transition: Callable[[TInput], EnvTransition] | None = None,
+        to_output: Callable[[EnvTransition], TOutput] | None = None,
+    ) -> DataProcessorPipeline[TInput, TOutput]:
+        """Build a pipeline from an in-memory config and optional state tensors.
+
+        Args:
+            config: A config dictionary with the same structure as the saved processor JSON.
+            state_dict: Optional in-memory pipeline state grouped by suffixless state key.
+            overrides: Optional constructor overrides keyed by registry name or class name.
+            to_transition: Optional converter from input data to `EnvTransition`.
+            to_output: Optional converter from `EnvTransition` to output data.
+
+        Returns:
+            A processor pipeline built from the config and optional state.
+        """
+        cls._validate_loaded_config("<in-memory config>", config, "<in-memory config>")
+
+        steps, remaining_override_keys = cls._build_steps_from_config(config, overrides or {})
+        cls._validate_overrides_used(remaining_override_keys, config)
+
+        pipeline = cls(
+            steps=steps,
+            name=config.get("name", "DataProcessorPipeline"),
+            to_transition=to_transition or cast(Callable[[TInput], EnvTransition], batch_to_transition),
+            to_output=to_output or cast(Callable[[EnvTransition], TOutput], transition_to_batch),
+        )
+        pipeline._serialized_state_filenames = cls._get_state_filenames_from_config(config)
+
+        if state_dict is not None:
+            pipeline.load_state_dict(state_dict)
+
+        return pipeline
 
     @classmethod
     def _load_config(
@@ -666,9 +869,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                 ) from e
 
     @classmethod
-    def _validate_loaded_config(
-        cls, model_id: str, loaded_config: dict[str, Any], config_filename: str
-    ) -> None:
+    def _validate_loaded_config(cls, model_id: str, loaded_config: Any, config_filename: str) -> None:
         """Validate that a config was loaded and is a valid processor config.
 
         This method validates processor config format with intelligent migration detection:
@@ -688,7 +889,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
         Args:
             model_id: The model identifier (used for migration detection)
-            loaded_config: The loaded config dictionary (guaranteed non-None)
+            loaded_config: The loaded config value to validate (may be non-dict)
             config_filename: The config filename that was loaded (for error messages)
 
         Raises:
@@ -702,9 +903,14 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                     model_id,
                     f"Config file '{config_filename}' is not a valid processor configuration",
                 )
+            loaded_config_description = (
+                list(loaded_config.keys())
+                if isinstance(loaded_config, dict)
+                else type(loaded_config).__name__
+            )
             raise ValueError(
                 f"Config file '{config_filename}' is not a valid processor configuration. "
-                f"Expected a config with 'steps' field, but got: {list(loaded_config.keys())}"
+                f"Expected a config with 'steps' field, but got: {loaded_config_description}"
             )
 
     @classmethod
@@ -766,26 +972,41 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             ImportError: If a step class cannot be imported or found in registry
             ValueError: If a step cannot be instantiated with its configuration
         """
-        steps: list[ProcessorStep] = []
-        override_keys = set(overrides.keys())
+        steps, remaining_override_keys = cls._build_steps_from_config(loaded_config, overrides)
 
-        for step_entry in loaded_config["steps"]:
-            # 1. Get step class and key
-            step_class, step_key = cls._resolve_step_class(step_entry)
-
-            # 2. Instantiate step with overrides
-            step_instance = cls._instantiate_step(step_entry, step_class, step_key, overrides)
-
-            # 3. Load step state if available
+        for step_instance, step_entry in zip(steps, loaded_config["steps"], strict=True):
             cls._load_step_state(step_instance, step_entry, model_id, base_path, hub_download_kwargs)
 
-            # 4. Track used overrides
-            if step_key in override_keys:
-                override_keys.discard(step_key)
+        return steps, remaining_override_keys
 
-            steps.append(step_instance)
+    @classmethod
+    def _build_steps_from_config(
+        cls,
+        loaded_config: dict[str, Any],
+        overrides: dict[str, Any],
+    ) -> tuple[list[ProcessorStep], set[str]]:
+        """Build processor steps from config without loading tensor state.
 
-        return steps, override_keys
+        Args:
+            loaded_config: The loaded processor configuration.
+            overrides: User-provided constructor overrides keyed by step key.
+
+        Returns:
+            A tuple containing instantiated steps and override keys that did not match a step.
+        """
+        processor_steps: list[ProcessorStep] = []
+        remaining_override_keys = set(overrides.keys())
+
+        for step_entry in loaded_config["steps"]:
+            step_class, step_key = cls._resolve_step_class(step_entry)
+            processor_step = cls._instantiate_step(step_entry, step_class, step_key, overrides)
+
+            if step_key in remaining_override_keys:
+                remaining_override_keys.discard(step_key)
+
+            processor_steps.append(processor_step)
+
+        return processor_steps, remaining_override_keys
 
     @classmethod
     def _resolve_step_class(cls, step_entry: dict[str, Any]) -> tuple[type[ProcessorStep], str]:
@@ -1096,7 +1317,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         return True
 
     @classmethod
-    def _is_processor_config(cls, config: dict) -> bool:
+    def _is_processor_config(cls, config: Any) -> bool:
         """Check if config follows DataProcessorPipeline format.
 
         This method validates the processor configuration structure:
@@ -1147,6 +1368,9 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         Returns:
             True if config follows valid DataProcessorPipeline format, False otherwise
         """
+        if not isinstance(config, dict):
+            return False
+
         # Must have a "steps" field with a list of step configurations
         if not isinstance(config.get("steps"), list):
             return False

src/lerobot/rollout/__init__.py

@@ -23,6 +23,7 @@ from .configs import (
     DAggerKeyboardConfig,
     DAggerPedalConfig,
     DAggerStrategyConfig,
+    EpisodicStrategyConfig,
     HighlightStrategyConfig,
     RolloutConfig,
     RolloutStrategyConfig,
@@ -49,6 +50,7 @@ from .inference import (
 from .strategies import (
     BaseStrategy,
     DAggerStrategy,
+    EpisodicStrategy,
     HighlightStrategy,
     RolloutStrategy,
     SentryStrategy,
@@ -66,6 +68,8 @@ __all__ = [
     "HardwareContext",
     "HighlightStrategy",
     "HighlightStrategyConfig",
+    "EpisodicStrategy",
+    "EpisodicStrategyConfig",
     "InferenceEngine",
     "InferenceEngineConfig",
     "PolicyContext",

src/lerobot/rollout/configs.py

@@ -121,6 +121,35 @@ class DAggerPedalConfig:
     upload: str = "KEY_C"
 
 
+@RolloutStrategyConfig.register_subclass("episodic")
+@dataclass
+class EpisodicStrategyConfig(RolloutStrategyConfig):
+    """Episode-oriented recording that mirrors the behavior of ``lerobot-record``.
+
+    Records ``dataset.num_episodes`` episodes of maximum ``dataset.episode_time_s`` each.
+    After each episode, runs ``dataset.reset_time_s`` seconds of reset time.
+
+    Keyboard controls:
+        Right arrow  — end current episode or reset phase early
+        Left arrow   — discard current episode and re-record
+        Escape       — stop recording session
+
+    In between episodes:
+    - if there is no teleop leader, the robot is held at its initial joint positions captured at startup.
+    - else, the robot is moved smoothly to the position of the teleop leader.
+    """
+
+    # This only applies if there are no teleop leaders specified.
+    # When True (default), moves the robot back to the joint positions captured at startup.
+    # Otherwise, leave the robot in its current position.
+    reset_to_initial_position: bool = True
+
+    # Whether to turn on or off the leader -> follower smooth handover behavior.
+    # When False, fallback to follower -> leader handover.
+    # Note that leader -> follower handover is only supported when the leader has `send_feedback` capability.
+    smooth_leader_to_follower_handover: bool = True
+
+
 @RolloutStrategyConfig.register_subclass("dagger")
 @dataclass
 class DAggerStrategyConfig(RolloutStrategyConfig):
@@ -229,7 +258,13 @@ class RolloutConfig:
 
         # TODO(Steven): DAgger shouldn't require a dataset (user may want to just rollout+intervene without recording), but for now we require it to simplify the implementation.
         needs_dataset = isinstance(
-            self.strategy, (SentryStrategyConfig, HighlightStrategyConfig, DAggerStrategyConfig)
+            self.strategy,
+            (
+                SentryStrategyConfig,
+                HighlightStrategyConfig,
+                DAggerStrategyConfig,
+                EpisodicStrategyConfig,
+            ),
         )
         if needs_dataset and (self.dataset is None or not self.dataset.repo_id):
             raise ValueError(f"{self.strategy.type} strategy requires --dataset.repo_id to be set")

src/lerobot/rollout/strategies/__init__.py

@@ -17,6 +17,7 @@
 from .base import BaseStrategy
 from .core import RolloutStrategy, estimate_max_episode_seconds, safe_push_to_hub, send_next_action
 from .dagger import DAggerEvents, DAggerPhase, DAggerStrategy
+from .episodic import EpisodicStrategy
 from .factory import create_strategy
 from .highlight import HighlightStrategy
 from .sentry import SentryStrategy
@@ -27,6 +28,7 @@ __all__ = [
     "DAggerPhase",
     "DAggerStrategy",
     "HighlightStrategy",
+    "EpisodicStrategy",
     "RolloutStrategy",
     "SentryStrategy",
     "create_strategy",

src/lerobot/rollout/strategies/dagger.py

@@ -56,10 +56,14 @@ from typing import Any
 
 import numpy as np
 
-from lerobot.common.control_utils import is_headless
+from lerobot.common.control_utils import (
+    follower_smooth_move_to,
+    is_headless,
+    teleop_smooth_move_to,
+    teleop_supports_feedback,
+)
 from lerobot.datasets import VideoEncodingManager
 from lerobot.datasets.utils import DEFAULT_VIDEO_FILE_SIZE_IN_MB
-from lerobot.teleoperators import Teleoperator
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame
 from lerobot.utils.import_utils import _pynput_available
@@ -69,7 +73,6 @@ from lerobot.utils.utils import log_say
 
 from ..configs import DAggerKeyboardConfig, DAggerPedalConfig, DAggerStrategyConfig
 from ..context import RolloutContext
-from ..robot_wrapper import ThreadSafeRobot
 from .core import RolloutStrategy, estimate_max_episode_seconds, safe_push_to_hub, send_next_action
 
 PYNPUT_AVAILABLE = _pynput_available
@@ -171,64 +174,6 @@ class DAggerEvents:
         self.upload_requested.clear()
 
 
-# ---------------------------------------------------------------------------
-# Teleoperator helpers
-# ---------------------------------------------------------------------------
-
-
-def _teleop_supports_feedback(teleop: Teleoperator) -> bool:
-    """Return True when the teleop can receive position feedback (is actuated).
-    TODO(Maxime): See if it is possible to unify this interface across teleops instead of duck-typing.
-    """
-    return (
-        bool(teleop.feedback_features)
-        and hasattr(teleop, "disable_torque")
-        and hasattr(teleop, "enable_torque")
-    )
-
-
-def _teleop_smooth_move_to(
-    teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 30
-) -> None:
-    """Smoothly move an actuated teleop to ``target_pos`` via linear interpolation.
-
-    Requires the teleoperator to support feedback
-    (i.e. have non-empty ``feedback_features`` and implement ``disable_torque`` / ``enable_torque``).
-
-    TODO(Maxime): This blocks up to ``duration_s`` seconds, during this time
-    the follower robot doesn't receive new actions, this could be an issue on LeKiwi.
-    """
-    teleop.enable_torque()
-    current = teleop.get_action()
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {
-            k: current[k] * (1 - t) + target_pos[k] * t if k in target_pos else current[k] for k in current
-        }
-        teleop.send_feedback(interp)
-        time.sleep(1 / fps)
-
-
-def _follower_smooth_move_to(
-    robot: ThreadSafeRobot, current: dict, target: dict, duration_s: float = 1.0, fps: int = 30
-) -> None:
-    """Smoothly move the follower robot from ``current`` to ``target`` action.
-
-    Used when the teleop is non-actuated: instead of driving the leader arm
-    to the follower, we bring the follower to the teleop's current pose.
-    Both ``current`` and ``target`` must be in robot-action key space.
-    """
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {k: current[k] * (1 - t) + target[k] * t if k in target else current[k] for k in current}
-        robot.send_action(interp)
-        time.sleep(1 / fps)
-
-
 # ---------------------------------------------------------------------------
 # Input device handlers
 # ---------------------------------------------------------------------------
@@ -756,31 +701,31 @@ class DAggerStrategy(RolloutStrategy):
             logger.info("Pausing engine - robot holds position")
             engine.pause()
 
-            if _teleop_supports_feedback(teleop) and prev_action is not None:
+            if teleop_supports_feedback(teleop) and prev_action is not None:
                 # TODO(Maxime): prev_action is in robot action key space (output of robot_action_processor).
                 # send_feedback expects teleop feedback key space. For homogeneous setups (e.g. SO-101
                 # leader + SO-101 follower) the keys are identical so this works. If the processor pipeline
                 # does non-trivial key renaming (e.g. a rename_map on action keys), the interpolation in
-                # _teleop_smooth_move_to silently no-ops and the arm doesn't move.
+                # teleop_smooth_move_to silently no-ops and the arm doesn't move.
                 logger.info("Smooth handover: moving leader arm to follower position")
-                _teleop_smooth_move_to(teleop, prev_action)
+                teleop_smooth_move_to(teleop, prev_action)
 
         elif old_phase == DAggerPhase.PAUSED and new_phase == DAggerPhase.CORRECTING:
             logger.info("Entering correction mode - human teleop control")
-            if not _teleop_supports_feedback(teleop) and prev_action is not None:
+            if not teleop_supports_feedback(teleop) and prev_action is not None:
                 logger.info("Smooth handover: sliding follower to teleop position")
                 obs = robot.get_observation()
                 teleop_action = teleop.get_action()
                 processed = ctx.processors.teleop_action_processor((teleop_action, obs))
                 target = ctx.processors.robot_action_processor((processed, obs))
-                _follower_smooth_move_to(robot, prev_action, target)
+                follower_smooth_move_to(robot, prev_action, target)
 
             # unlock the teleop for human control
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.disable_torque()
 
         elif old_phase == DAggerPhase.CORRECTING and new_phase == DAggerPhase.PAUSED:
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.enable_torque()
 
         elif new_phase == DAggerPhase.AUTONOMOUS:
@@ -790,7 +735,7 @@ class DAggerStrategy(RolloutStrategy):
             engine.resume()
 
             # release teleop before resuming the policy
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.disable_torque()
 
     # ------------------------------------------------------------------

src/lerobot/rollout/strategies/episodic.py

@@ -0,0 +1,335 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Episodic rollout strategy: mirrors the behavior of ``lerobot-record``.
+
+- Policy drives the robot during each recording episode.
+- An optional teleoperator can drive the robot during reset phases so the
+  operator can bring the environment back to its starting configuration.
+  If no teleop is connected the robot stays in its current position.
+- Keyboard controls:
+
+      Right arrow  — end the current episode or reset phase early
+      Left arrow   — discard the current episode and re-record it
+      Escape       — stop the recording session
+
+Dataset naming follows the rollout convention: repo names must start with ``rollout_``.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+import time
+
+from lerobot.common.control_utils import (
+    follower_smooth_move_to,
+    init_keyboard_listener,
+    is_headless,
+    teleop_smooth_move_to,
+    teleop_supports_feedback,
+)
+from lerobot.datasets import VideoEncodingManager
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import log_rerun_data
+
+from ..configs import EpisodicStrategyConfig
+from ..context import RolloutContext
+from .core import RolloutStrategy, safe_push_to_hub, send_next_action
+
+logger = logging.getLogger(__name__)
+
+
+class EpisodicStrategy(RolloutStrategy):
+    """Policy-driven multi-episode recording, mirrors the behavior of ``lerobot-record``.
+
+    Each recording episode runs the policy for maximum ``dataset.episode_time_s``
+    seconds, recording every frame.  A reset phase of ``dataset.reset_time_s``
+    follows every episode (except the last) so the operator can manually
+    reset the environment.  During the reset phase, an optional teleoperator
+    drives the robot; if none is present the robot returns to its initial joint positions captured at startup.
+
+    The policy state (hidden state, RTC queue, interpolator) is reset at
+    the start of each recording episode.
+
+    Keyboard events:
+        right arrow  → end current episode or reset phase early
+        left arrow   → discard & re-record current episode
+        ESC          → stop the session
+    """
+
+    config: EpisodicStrategyConfig
+
+    def __init__(self, config: EpisodicStrategyConfig) -> None:
+        super().__init__(config)
+        self._listener = None
+        self._events: dict | None = None
+
+    def setup(self, ctx: RolloutContext) -> None:
+        """Start the inference engine and attach the keyboard listener."""
+        self._init_engine(ctx)
+        self._listener, self._events = init_keyboard_listener()
+        logger.info("Episodic strategy ready")
+
+    def run(self, ctx: RolloutContext) -> None:
+        """Main multi-episode recording loop."""
+        cfg = ctx.runtime.cfg
+        dataset_cfg = cfg.dataset
+        robot = ctx.hardware.robot_wrapper
+        teleop = ctx.hardware.teleop
+        dataset = ctx.data.dataset
+        events = self._events
+        features = ctx.data.dataset_features
+
+        fps = cfg.fps
+        episode_time_s = dataset_cfg.episode_time_s
+        reset_time_s = dataset_cfg.reset_time_s
+        num_episodes = dataset_cfg.num_episodes
+        single_task = dataset_cfg.single_task or cfg.task
+        play_sounds = cfg.play_sounds
+
+        display_compressed = (
+            True
+            if (cfg.display_data and cfg.display_ip is not None and cfg.display_port is not None)
+            else cfg.display_compressed_images
+        )
+
+        with VideoEncodingManager(dataset):
+            try:
+                recorded_episodes = 0
+                while recorded_episodes < num_episodes and not events["stop_recording"]:
+                    if ctx.runtime.shutdown_event.is_set():
+                        break
+
+                    # Reset policy state at episode start (discard leftover hidden state / queue)
+                    self._engine.reset()
+                    self._interpolator.reset()
+                    self._engine.resume()
+
+                    log_say(f"Recording episode {dataset.num_episodes}", play_sounds)
+                    self._policy_loop(
+                        ctx=ctx,
+                        robot=robot,
+                        events=events,
+                        features=features,
+                        fps=fps,
+                        control_time_s=episode_time_s,
+                        dataset=dataset,
+                        single_task=single_task,
+                    )
+
+                    # Reset phase, skip after the last episode (but run when re-recording)
+                    if not events["stop_recording"] and (
+                        recorded_episodes < num_episodes - 1 or events["rerecord_episode"]
+                    ):
+                        log_say("Reset the environment", play_sounds)
+
+                        if teleop:
+                            # Smooth handover so the transition to teleop control is jerk-free.
+                            # For actuated teleops: drive the leader arm to the follower's current
+                            # position so the operator takes over without fighting the arm.
+                            # For non-actuated teleops: slide the follower to the teleop's current
+                            # pose instead, since the leader cannot be driven.
+                            obs = robot.get_observation()
+                            current_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+                            if (
+                                teleop_supports_feedback(teleop)
+                                and self.config.smooth_leader_to_follower_handover
+                            ):
+                                logger.info("Smooth handover: moving leader arm to follower position")
+                                teleop_smooth_move_to(teleop, current_pos, duration_s=2)
+                                teleop.disable_torque()
+                            else:
+                                logger.info("Smooth handover: sliding follower to teleop position")
+                                teleop_action = teleop.get_action()
+                                processed = ctx.processors.teleop_action_processor((teleop_action, obs))
+                                target = ctx.processors.robot_action_processor((processed, obs))
+                                follower_smooth_move_to(robot, current_pos, target, duration_s=1)
+
+                        elif self.config.reset_to_initial_position:
+                            # No teleop: return the robot to its startup position.
+                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+
+                        self._reset_loop(
+                            ctx=ctx,
+                            robot=robot,
+                            teleop=teleop,
+                            events=events,
+                            fps=fps,
+                            control_time_s=reset_time_s,
+                            display_data=cfg.display_data,
+                            display_compressed=display_compressed,
+                        )
+
+                    if events["rerecord_episode"]:
+                        log_say("Re-record episode", play_sounds)
+                        events["rerecord_episode"] = False
+                        events["exit_early"] = False
+                        dataset.clear_episode_buffer()
+
+                        # returns to its initial joint positions captured at startup
+                        if not teleop and self.config.reset_to_initial_position:
+                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+
+                        continue
+
+                    dataset.save_episode()
+                    recorded_episodes += 1
+            finally:
+                # Save any frames buffered in the current episode so an unexpected
+                # exception or KeyboardInterrupt does not silently drop recorded data.
+                # suppress: save_episode raises if the buffer is empty (nothing to lose).
+                logger.info("Episodic control loop ended — saving any in-progress episode")
+                with contextlib.suppress(Exception):
+                    dataset.save_episode()
+
+    def _policy_loop(
+        self,
+        ctx: RolloutContext,
+        robot,
+        events: dict,
+        features: dict,
+        fps: float,
+        control_time_s: float,
+        dataset,
+        single_task: str,
+    ) -> None:
+        """Policy-driven recording loop for a single episode."""
+        interpolator = self._interpolator
+        control_interval = interpolator.get_control_interval(fps)
+
+        timestamp = 0.0
+        start_t = time.perf_counter()
+
+        while timestamp < control_time_s:
+            loop_start = time.perf_counter()
+
+            if events["exit_early"]:
+                events["exit_early"] = False
+                break
+
+            if ctx.runtime.shutdown_event.is_set():
+                break
+
+            obs = robot.get_observation()
+            obs_processed = self._process_observation_and_notify(ctx.processors, obs)
+
+            if self._handle_warmup(ctx.runtime.cfg.use_torch_compile, loop_start, control_interval):
+                continue
+
+            action_dict = send_next_action(obs_processed, obs, ctx, interpolator)
+
+            if action_dict is not None:
+                obs_frame = build_dataset_frame(features, obs_processed, prefix=OBS_STR)
+                action_frame = build_dataset_frame(features, action_dict, prefix=ACTION)
+                dataset.add_frame({**obs_frame, **action_frame, "task": single_task})
+                self._log_telemetry(obs_processed, action_dict, ctx.runtime)
+
+            dt = time.perf_counter() - loop_start
+            sleep_t = control_interval - dt
+            if sleep_t < 0:
+                logger.warning(
+                    f"Record loop is running slower ({1 / dt:.1f} Hz) than the target FPS ({fps} Hz). "
+                    "Dataset frames might be dropped and robot control might be unstable. "
+                    "Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long "
+                    "3) CPU starvation"
+                )
+            precise_sleep(max(sleep_t, 0.0))
+            timestamp = time.perf_counter() - start_t
+
+    def _reset_loop(
+        self,
+        ctx: RolloutContext,
+        robot,
+        teleop,
+        events: dict,
+        fps: float,
+        control_time_s: float,
+        display_data: bool,
+        display_compressed: bool,
+    ) -> None:
+        """Reset-phase loop: teleop drives the robot if available, no recording."""
+        processors = ctx.processors
+        control_interval = 1.0 / fps
+
+        timestamp = 0.0
+        start_t = time.perf_counter()
+
+        while timestamp < control_time_s:
+            loop_start = time.perf_counter()
+
+            if events["exit_early"]:
+                events["exit_early"] = False
+                break
+
+            if ctx.runtime.shutdown_event.is_set():
+                break
+
+            obs = robot.get_observation()
+
+            if teleop is not None:
+                act = teleop.get_action()
+                act_teleop = processors.teleop_action_processor((act, obs))
+                robot_action = processors.robot_action_processor((act_teleop, obs))
+                robot.send_action(robot_action)
+
+                if display_data:
+                    obs_processed = processors.robot_observation_processor(obs)
+                    log_rerun_data(
+                        observation=obs_processed,
+                        action=act_teleop,
+                        compress_images=display_compressed,
+                    )
+
+            dt = time.perf_counter() - loop_start
+            sleep_t = control_interval - dt
+            precise_sleep(max(sleep_t, 0.0))
+            timestamp = time.perf_counter() - start_t
+
+    def teardown(self, ctx: RolloutContext) -> None:
+        """Finalise dataset, stop listener, push to hub, and disconnect hardware."""
+        cfg = ctx.runtime.cfg
+        play_sounds = cfg.play_sounds
+
+        log_say("Stop recording", play_sounds, blocking=True)
+
+        if not is_headless() and self._listener is not None:
+            self._listener.stop()
+
+        if ctx.data.dataset is not None:
+            logger.info("Finalizing dataset...")
+            ctx.data.dataset.finalize()
+
+        if (
+            cfg.dataset is not None
+            and cfg.dataset.push_to_hub
+            and ctx.data.dataset is not None
+            and safe_push_to_hub(
+                ctx.data.dataset,
+                tags=cfg.dataset.tags,
+                private=cfg.dataset.private,
+            )
+        ):
+            logger.info("Dataset uploaded to hub")
+            log_say("Dataset uploaded to hub", play_sounds)
+
+        self._teardown_hardware(
+            ctx.hardware,
+            return_to_initial_position=cfg.return_to_initial_position,
+        )
+        log_say("Exiting", play_sounds)
+        logger.info("Episodic strategy teardown complete")

src/lerobot/rollout/strategies/factory.py

@@ -21,6 +21,7 @@ from typing import TYPE_CHECKING
 from .base import BaseStrategy
 from .core import RolloutStrategy
 from .dagger import DAggerStrategy
+from .episodic import EpisodicStrategy
 from .highlight import HighlightStrategy
 from .sentry import SentryStrategy
 
@@ -42,4 +43,8 @@ def create_strategy(config: RolloutStrategyConfig) -> RolloutStrategy:
         return HighlightStrategy(config)
     if config.type == "dagger":
         return DAggerStrategy(config)
-    raise ValueError(f"Unknown strategy type '{config.type}'. Available: base, sentry, highlight, dagger")
+    if config.type == "episodic":
+        return EpisodicStrategy(config)
+    raise ValueError(
+        f"Unknown strategy type '{config.type}'. Available: base, sentry, highlight, dagger, episodic"
+    )

src/lerobot/scripts/lerobot_rollout.py

@@ -25,6 +25,7 @@ Strategies
     --strategy.type=sentry     Continuous recording with auto-upload
     --strategy.type=highlight  Ring buffer + keystroke save
     --strategy.type=dagger     Human-in-the-loop (DAgger / RaC)
+    --strategy.type=episodic   Episode-oriented recording with reset phases
 
 Inference backends
 ------------------
@@ -111,6 +112,18 @@ Usage examples
         --display_data=true \\
         --use_torch_compile=true
 
+    # Episodic mode — episode-oriented recording with reset phases
+    lerobot-rollout \\
+        --strategy.type=episodic \\
+        --policy.path=user/my_policy \\
+        --robot.type=so100_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --teleop.type=so100_leader \\
+        --teleop.port=/dev/ttyACM1 \\
+        --dataset.repo_id=user/rollout_episodic_data \\
+        --dataset.num_episodes=20 \\
+        --dataset.single_task="Grab the cube"
+
     # Resume a previous sentry recording session
     lerobot-rollout \\
         --strategy.type=sentry \\

src/lerobot/scripts/lerobot_train.py

@@ -36,6 +36,8 @@ from tqdm import tqdm
 from lerobot.common.train_utils import (
     get_step_checkpoint_dir,
     get_step_identifier,
+    load_training_batch_size,
+    load_training_num_processes,
     load_training_state,
     save_checkpoint,
     update_last_checkpoint,
@@ -43,7 +45,7 @@ from lerobot.common.train_utils import (
 from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets import EpisodeAwareSampler, make_dataset
+from lerobot.datasets import EpisodeAwareSampler, compute_sampler_state, make_dataset
 from lerobot.envs import close_envs, make_env, make_env_pre_post_processors
 from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
@@ -232,14 +234,16 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
 
-    # Dataset loading synchronization: main process downloads first to avoid race conditions
+    # Dataset loading synchronization: the global main process downloads once to the shared
+    # dataset root, then a barrier lets every other rank read the already-populated copy.
+    # LeRobotDataset skips its snapshot_download when try_load() succeeds, so no rank re-downloads.
     if is_main_process:
         logging.info("Creating dataset")
         dataset = make_dataset(cfg)
 
     accelerator.wait_for_everyone()
 
-    # Now all other processes can safely load the dataset
+    # Other ranks read from the shared copy populated by the main process.
     if not is_main_process:
         dataset = make_dataset(cfg)
 
@@ -384,15 +388,47 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         logging.info(f"{num_total_params=} ({format_big_number(num_total_params)})")
 
     # create dataloader for offline training
-    if hasattr(active_cfg, "drop_n_last_frames"):
+    if not cfg.dataset.streaming:
+        # All non-streaming (map-style) datasets use EpisodeAwareSampler.
+        # The order is a pure function of (seed, epoch), so every rank independently produces the
+        # same permutation. accelerate then shards it disjointly across ranks via BatchSamplerShard
+        # without needing a `generator` attribute to synchronize an RNG, and resume is sample-exact.
         shuffle = False
         sampler = EpisodeAwareSampler(
             dataset.meta.episodes["dataset_from_index"],
             dataset.meta.episodes["dataset_to_index"],
             episode_indices_to_use=dataset.episodes,
-            drop_n_last_frames=active_cfg.drop_n_last_frames,
+            drop_n_last_frames=getattr(active_cfg, "drop_n_last_frames", 0),
             shuffle=True,
+            seed=cfg.seed if cfg.seed is not None else 0,
         )
+        if cfg.resume and step > 0:
+            # The resume offset depends on the (num_processes, batch_size) that produced `step`, so
+            # use the values recorded in the checkpoint (falling back to the current ones for older
+            # ckpts that did not store them).
+            saved_num_processes = load_training_num_processes(cfg.checkpoint_path)
+            saved_batch_size = load_training_batch_size(cfg.checkpoint_path)
+            ckpt_num_processes = saved_num_processes or accelerator.num_processes
+            ckpt_batch_size = saved_batch_size or cfg.batch_size
+            if is_main_process and saved_num_processes not in (None, accelerator.num_processes):
+                logging.warning(
+                    f"Resuming with num_processes={accelerator.num_processes} but the checkpoint was "
+                    f"written with num_processes={saved_num_processes}. The data order resumes at the "
+                    "right epoch/offset, but per-rank sample-exactness requires the same world size."
+                )
+            if is_main_process and saved_batch_size not in (None, cfg.batch_size):
+                logging.warning(
+                    f"Resuming with batch_size={cfg.batch_size} but the checkpoint was written with "
+                    f"batch_size={saved_batch_size}. The data order resumes at the right epoch/offset, "
+                    "but per-rank sample-exactness requires the same batch size."
+                )
+            sampler_state = compute_sampler_state(step, len(sampler), ckpt_batch_size, ckpt_num_processes)
+            sampler.load_state_dict(sampler_state)
+            if is_main_process:
+                logging.info(
+                    f"Resuming data order at epoch {sampler_state['epoch']}, "
+                    f"sample {sampler_state['start_index']}"
+                )
     else:
         shuffle = True
         sampler = None
@@ -511,6 +547,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     scheduler=lr_scheduler,
                     preprocessor=preprocessor,
                     postprocessor=postprocessor,
+                    num_processes=accelerator.num_processes,
+                    batch_size=cfg.batch_size,
                 )
                 update_last_checkpoint(checkpoint_dir)
                 if wandb_logger:

tests/datasets/test_sampler.py

@@ -114,6 +114,19 @@ def test_shuffle():
     assert set(sampler) == {0, 1, 2, 3, 4, 5}
 
 
+def test_shuffle_is_reproducible_across_instances():
+    # The order is a pure function of (seed, epoch), so two fresh samplers (e.g. two ranks)
+    # produce the same permutation without any generator synchronization.
+    sampler_a = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0
+    # Desyncing the global RNG must not affect the permutation.
+    sampler_c = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    torch.randperm(1000)  # consume global RNG, as rank-asymmetric code (e.g. eval) would
+    assert list(sampler_c) == epoch_0
+
+
 def test_negative_drop_first_frames_raises():
     with pytest.raises(ValueError, match="drop_n_first_frames must be >= 0"):
         EpisodeAwareSampler([0], [10], drop_n_first_frames=-1)
@@ -137,3 +150,87 @@ def test_partial_episode_drop_warns(caplog):
     # Episode 0 is skipped (1 frame, drop 1), Episode 1 keeps frames 2-5
     assert sampler.indices == [2, 3, 4, 5]
     assert "Episode 0" in caplog.text
+
+
+# --- seeded (seed, epoch) shuffling, resume, and state ---
+
+from lerobot.datasets.sampler import compute_sampler_state  # noqa: E402
+
+EPISODE_BOUNDS = ([0, 2, 3], [2, 3, 6])  # episodes of 2, 1 and 3 frames
+
+
+@pytest.mark.parametrize("num_frames", [1, 2, 3, 37, 64, 100])
+def test_deterministic_sampler_shuffle_is_permutation(num_frames):
+    for seed in (0, 1, 1234):
+        sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=seed)
+        assert sorted(sampler) == list(range(num_frames))
+
+
+def test_deterministic_sampler_epochs_reproduce_and_differ():
+    sampler_a = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0  # same (seed, epoch) -> same order on any process
+    epoch_1 = list(sampler_a)  # __iter__ auto-advances the epoch
+    assert epoch_1 != epoch_0
+    assert sorted(epoch_1) == sorted(epoch_0)
+    sampler_a.set_epoch(0)
+    assert list(sampler_a) == epoch_0
+    assert list(EpisodeAwareSampler([0], [100], shuffle=True, seed=7)) != epoch_0
+
+
+def test_deterministic_sampler_resume_mid_epoch():
+    reference = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+    epoch_0 = list(reference)
+    epoch_1 = list(reference)
+    for start in (0, 1, 4, len(epoch_0)):
+        resumed = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+        resumed.load_state_dict({"epoch": 0, "start_index": start})
+        assert list(resumed) == epoch_0[start:]
+        # the resumed sampler continues into the same epoch 1 as the uninterrupted one
+        assert list(resumed) == epoch_1
+
+
+def test_deterministic_sampler_construction_stores_only_boundaries():
+    # Construction is O(num_episodes), not O(num_frames): a million-frame single episode
+    # instantiates from just its boundaries without materializing a per-frame index list.
+    num_frames = 1_000_000
+    sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    assert len(sampler) == num_frames
+    assert sampler._starts.shape == (1,) and sampler._cum_lengths.shape == (1,)
+
+
+def test_deterministic_sampler_resume_is_exact_at_scale():
+    # Seeded randperm makes resume sample-exact at non-trivial sizes: regenerating the epoch's
+    # permutation and slicing from the saved offset reproduces the remaining order exactly.
+    num_frames = 100_000
+    reference = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    epoch_0 = list(reference)
+    assert sorted(epoch_0) == list(range(num_frames))
+    start = num_frames - 5
+    resumed = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    resumed.load_state_dict({"epoch": 0, "start_index": start})
+    assert list(resumed) == epoch_0[start:]
+
+
+def test_compute_sampler_state():
+    # 100 frames, batch 10, 2 ranks -> 10 underlying batches, 5 per rank per epoch.
+    assert compute_sampler_state(step=0, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 0,
+        "start_index": 0,
+    }
+    # step 7 -> epoch 1, 2 per-rank batches in = 2 * 10 * 2 = 40 samples in
+    assert compute_sampler_state(step=7, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 40,
+    }
+    # uneven epoch: 95 frames -> 10 underlying batches (last short), still 5 per rank
+    assert compute_sampler_state(step=12, num_frames=95, batch_size=10, num_processes=2) == {
+        "epoch": 2,
+        "start_index": 40,
+    }
+    # uneven sharding: 105 frames -> 11 underlying batches, 6 per rank (even_batches pads)
+    assert compute_sampler_state(step=11, num_frames=105, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 100,
+    }

tests/processor/test_pipeline.py

@@ -24,6 +24,7 @@ from typing import Any
 import pytest
 import torch
 import torch.nn as nn
+from safetensors.torch import load_file
 
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 
@@ -174,6 +175,53 @@ class MockStepWithTensorState(ProcessorStep):
         return features
 
 
+class MockLazyTensorStateStep(ProcessorStep):
+    """Mock step whose tensor state is not present in constructor config."""
+
+    def __init__(
+        self, name: str = "lazy_tensor_step", scale: float = 1.0, initial_value: float | None = None
+    ):
+        self.name = name
+        self.scale = scale
+        self.tensor_state: torch.Tensor | None = None
+
+        if initial_value is not None:
+            self.tensor_state = torch.tensor([initial_value], dtype=torch.float32)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Return the transition unchanged."""
+        return transition
+
+    def get_config(self) -> dict[str, Any]:
+        """Return constructor config while intentionally omitting tensor state."""
+        return {
+            "name": self.name,
+            "scale": self.scale,
+        }
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Return tensor state only after it has been initialized or loaded."""
+        if self.tensor_state is None:
+            return {}
+
+        return {"tensor_state": self.tensor_state}
+
+    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+        """Load tensor state."""
+        self.tensor_state = state["tensor_state"].clone()
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Return features unchanged."""
+        return features
+
+
+@ProcessorStepRegistry.register("registered_lazy_tensor_state_step")
+class RegisteredLazyTensorStateStep(MockLazyTensorStateStep):
+    """Registered lazy tensor state step for registry-based serialization tests."""
+
+
 def test_empty_pipeline():
     """Test pipeline with no steps."""
     pipeline = DataProcessorPipeline([], to_transition=identity_transition, to_output=identity_transition)
@@ -620,6 +668,178 @@ def test_mixed_json_and_tensor_state():
         assert torch.allclose(loaded_step.running_mean, step.running_mean)
 
 
+def test_get_config_matches_saved_json():
+    """Test that in-memory config matches the config written by save_pretrained."""
+    stateless_step = MockStep(name="stateless")
+    stateful_step = MockLazyTensorStateStep(name="stateful", initial_value=4.0)
+    pipeline = DataProcessorPipeline([stateless_step, stateful_step], name="Memory Pipeline")
+
+    in_memory_config = pipeline.get_config()
+
+    assert pipeline.get_config() == in_memory_config
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+
+        config_path = Path(tmp_dir) / "memory_pipeline.json"
+        with open(config_path) as file_pointer:
+            saved_config = json.load(file_pointer)
+
+    assert in_memory_config == saved_config
+    assert "state_file" not in in_memory_config["steps"][0]
+    assert in_memory_config["steps"][1]["state_file"] == "memory_pipeline_step_1.safetensors"
+
+
+def test_state_dict_matches_saved_safetensors():
+    """Test that in-memory state matches the safetensors written by save_pretrained."""
+    stateful_step = MockLazyTensorStateStep(initial_value=7.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Stateful Pipeline")
+
+    in_memory_state_dict = pipeline.state_dict()
+    state_filename = "stateful_pipeline_step_0.safetensors"
+    state_key = "stateful_pipeline_step_0"
+
+    assert set(in_memory_state_dict) == {state_key}
+    assert set(in_memory_state_dict[state_key]) == {"tensor_state"}
+
+    in_memory_state_dict[state_key]["tensor_state"].add_(1)
+    assert stateful_step.tensor_state is not None
+    assert torch.equal(stateful_step.tensor_state, torch.tensor([7.0]))
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+        saved_state_dict = load_file(Path(tmp_dir) / state_filename)
+
+    torch.testing.assert_close(saved_state_dict["tensor_state"], torch.tensor([7.0]))
+
+
+def test_save_pretrained_still_writes_expected_serialization_files():
+    """Test that save_pretrained keeps the existing config and state filenames."""
+    stateful_step = MockLazyTensorStateStep(initial_value=3.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Policy Preprocessor")
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+
+        save_path = Path(tmp_dir)
+        assert (save_path / "policy_preprocessor.json").exists()
+        assert (save_path / "policy_preprocessor_step_0.safetensors").exists()
+
+
+def test_from_config_round_trips_stateful_pipeline():
+    """Test that from_config rebuilds a stateful pipeline from in-memory artifacts."""
+    stateful_step = MockLazyTensorStateStep(name="roundtrip", initial_value=11.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Roundtrip Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config, state_dict=pipeline_state_dict)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert len(loaded_pipeline) == 1
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([11.0]))
+
+
+def test_from_config_round_trips_registered_stateful_pipeline():
+    """Test that from_config resolves registry steps and loads their named tensor state."""
+    stateful_step = RegisteredLazyTensorStateStep(name="registered", initial_value=29.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Registry Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+    state_filename = "registry_pipeline_step_0_registered_lazy_tensor_state_step.safetensors"
+    state_key = "registry_pipeline_step_0_registered_lazy_tensor_state_step"
+
+    assert config["steps"][0]["registry_name"] == "registered_lazy_tensor_state_step"
+    assert config["steps"][0]["state_file"] == state_filename
+    assert set(pipeline_state_dict) == {state_key}
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config, state_dict=pipeline_state_dict)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, RegisteredLazyTensorStateStep)
+    assert loaded_step.tensor_state is not None
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([29.0]))
+
+
+def test_from_config_preserves_state_metadata_for_empty_initial_state():
+    """Test in-memory loading when rebuilt steps start without tensor state."""
+    stateful_step = MockLazyTensorStateStep(name="lazy", initial_value=13.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Lazy Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    assert loaded_step.state_dict() == {}
+    assert "state_file" not in loaded_pipeline.get_config()["steps"][0]
+
+    loaded_pipeline.load_state_dict(pipeline_state_dict)
+
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([13.0]))
+
+
+def test_from_config_applies_overrides_before_state_loading():
+    """Test that constructor overrides and tensor state loading are separate operations."""
+    stateful_step = MockLazyTensorStateStep(name="override", scale=1.0, initial_value=17.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Override Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(
+        config,
+        state_dict=pipeline_state_dict,
+        overrides={"MockLazyTensorStateStep": {"scale": 5.0}},
+    )
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    assert loaded_step.scale == 5.0
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([17.0]))
+
+
+def test_load_state_dict_raises_on_missing_expected_state():
+    """Test loading raises when serialized config expects missing state."""
+    stateful_step = MockLazyTensorStateStep(initial_value=19.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Missing Pipeline")
+    loaded_pipeline = DataProcessorPipeline.from_config(pipeline.get_config())
+
+    with pytest.raises(KeyError, match="missing_pipeline_step_0"):
+        loaded_pipeline.load_state_dict({})
+
+
+def test_load_state_dict_raises_on_unexpected_extra_state():
+    """Test loading raises on unexpected top-level state keys."""
+    pipeline = DataProcessorPipeline([MockStep(name="stateless")], name="Unexpected Pipeline")
+
+    with pytest.raises(KeyError, match="extra"):
+        pipeline.load_state_dict({"extra": {"tensor_state": torch.tensor([1.0])}})
+
+
+def test_stateless_pipeline_in_memory_serialization_returns_empty_state():
+    """Test stateless in-memory serialization and loading."""
+    pipeline = DataProcessorPipeline([MockStep(name="stateless")], name="Stateless Pipeline")
+    config = pipeline.get_config()
+    config_without_name = {"steps": config["steps"]}
+
+    assert pipeline.state_dict() == {}
+    assert all("state_file" not in step_entry for step_entry in config["steps"])
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config_without_name, state_dict={})
+
+    assert loaded_pipeline.name == "DataProcessorPipeline"
+    assert loaded_pipeline.state_dict() == {}
+
+
+@pytest.mark.parametrize("invalid_config", [None, [], "not config"])
+def test_from_config_rejects_non_dict_config(invalid_config):
+    """Test from_config reports invalid top-level config values cleanly."""
+    with pytest.raises(ValueError, match="not a valid processor configuration"):
+        DataProcessorPipeline.from_config(invalid_config)  # type: ignore[arg-type]
+
+
 class MockModuleStep(ProcessorStep, nn.Module):
     """Mock step that inherits from nn.Module to test state_dict handling of module parameters."""
 

tests/test_rollout.py

@@ -59,6 +59,7 @@ def test_strategy_config_types():
     from lerobot.rollout import (
         BaseStrategyConfig,
         DAggerStrategyConfig,
+        EpisodicStrategyConfig,
         HighlightStrategyConfig,
         SentryStrategyConfig,
     )
@@ -67,6 +68,7 @@ def test_strategy_config_types():
     assert SentryStrategyConfig().type == "sentry"
     assert HighlightStrategyConfig().type == "highlight"
     assert DAggerStrategyConfig().type == "dagger"
+    assert EpisodicStrategyConfig().type == "episodic"
 
 
 def test_dagger_config_invalid_input_device():
@@ -203,6 +205,8 @@ def test_create_strategy_dispatches():
         BaseStrategyConfig,
         DAggerStrategy,
         DAggerStrategyConfig,
+        EpisodicStrategy,
+        EpisodicStrategyConfig,
         SentryStrategy,
         SentryStrategyConfig,
         create_strategy,
@@ -211,6 +215,7 @@ def test_create_strategy_dispatches():
     assert isinstance(create_strategy(BaseStrategyConfig()), BaseStrategy)
     assert isinstance(create_strategy(SentryStrategyConfig()), SentryStrategy)
     assert isinstance(create_strategy(DAggerStrategyConfig()), DAggerStrategy)
+    assert isinstance(create_strategy(EpisodicStrategyConfig()), EpisodicStrategy)
 
 
 def test_create_strategy_unknown_raises():

tests/utils/test_train_utils.py

@@ -20,6 +20,8 @@ from unittest.mock import Mock, patch
 from lerobot.common.train_utils import (
     get_step_checkpoint_dir,
     get_step_identifier,
+    load_training_batch_size,
+    load_training_num_processes,
     load_training_state,
     load_training_step,
     save_checkpoint,
@@ -63,6 +65,28 @@ def test_load_training_step(tmp_path):
     assert loaded_step == step
 
 
+def test_save_training_state_records_num_processes(tmp_path, optimizer, scheduler):
+    save_training_state(tmp_path, 10, optimizer, scheduler, num_processes=4)
+    assert load_training_num_processes(tmp_path) == 4
+
+
+def test_load_training_num_processes_absent_returns_none(tmp_path, optimizer, scheduler):
+    # Checkpoints written before the world size was recorded must still load (back-compat).
+    save_training_state(tmp_path, 10, optimizer, scheduler)
+    assert load_training_num_processes(tmp_path) is None
+
+
+def test_save_training_state_records_batch_size(tmp_path, optimizer, scheduler):
+    save_training_state(tmp_path, 10, optimizer, scheduler, batch_size=32)
+    assert load_training_batch_size(tmp_path) == 32
+
+
+def test_load_training_batch_size_absent_returns_none(tmp_path, optimizer, scheduler):
+    # Checkpoints written before the batch size was recorded must still load (back-compat).
+    save_training_state(tmp_path, 10, optimizer, scheduler)
+    assert load_training_batch_size(tmp_path) is None
+
+
 def test_update_last_checkpoint(tmp_path):
     checkpoint = tmp_path / "0005"
     checkpoint.mkdir()

uv.lock

@@ -1764,7 +1764,7 @@ wheels = [
 
 [[package]]
 name = "gym-aloha"
-version = "0.1.3"
+version = "0.1.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "dm-control" },
@@ -1772,14 +1772,14 @@ dependencies = [
     { name = "imageio", extra = ["ffmpeg"] },
     { name = "mujoco" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/b5/5e/4bb7204730501c2f645e0532a2df4339206948b2882f77cbf0eaf75bc5fe/gym_aloha-0.1.3.tar.gz", hash = "sha256:b794b246a2e6da6ce5f75e152f553fbd4412704bc217fe6311d0ede3bb72a75e", size = 443468, upload-time = "2025-10-09T14:02:35.024Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/4a/c5/a5b8bdbddfcadec0b52b50e6d1a70325e09e6b594e5f55929d67d9122e2c/gym_aloha-0.1.4.tar.gz", hash = "sha256:0dc4e645045aeb3e74e3c320872d28df6dc93a8751d6ab2f266a2ca11323131f", size = 443466, upload-time = "2026-06-10T09:13:25.525Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/57/6c/10da397177c48ce360efa66ec21b10b10ef5fa2766256fcd8d7d9b5fa6fc/gym_aloha-0.1.3-py3-none-any.whl", hash = "sha256:a94e5747e71307897ded7ae17ed97fab05e814dcb714a16d320f110444f9d0c3", size = 447908, upload-time = "2025-10-09T14:02:33.253Z" },
+    { url = "https://files.pythonhosted.org/packages/35/e3/3afd0e517a503aabe255bf65f5136490acb79c43189e8d56a3aa63081a10/gym_aloha-0.1.4-py3-none-any.whl", hash = "sha256:d9044290fbccddf0be4246b5287cf0eb6b9ddee545a3d222ce8d78c93ce7125e", size = 447908, upload-time = "2026-06-10T09:13:23.868Z" },
 ]
 
 [[package]]
 name = "gym-hil"
-version = "0.1.13"
+version = "0.1.14"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "gymnasium" },
@@ -1789,9 +1789,9 @@ dependencies = [
     { name = "pygame" },
     { name = "pynput" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/f3/41/e89c87b3c66fb2f8ab5818bff4aa552977911eabaee7c12a8a336dcc406f/gym_hil-0.1.13.tar.gz", hash = "sha256:b9eab7a0acc811f181254e3ad72865830fdbb292c236895f374135d3d62f1b27", size = 5668001, upload-time = "2025-10-21T09:57:24.01Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/0c/64/b5cfe59d6a69d20497218f01ad2bdaa2a5a72b850bdb1a445d804ecc9948/gym_hil-0.1.14.tar.gz", hash = "sha256:aeee688dcb3ec72e7bcbe604df4a3f990cce49c8a2da469dd67c3a4eeb4c6bbb", size = 5667991, upload-time = "2026-06-10T09:16:38.98Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c2/8d/9e3ab53f9aac7bd542f339efd0a9283fa76e034474987e0705379274dfcf/gym_hil-0.1.13-py3-none-any.whl", hash = "sha256:b6444fc43ce1a68ce403df14f99100d9c903ae05d822959e9cd0b76a50b93320", size = 5750805, upload-time = "2025-10-21T09:57:22.068Z" },
+    { url = "https://files.pythonhosted.org/packages/72/97/a7a9c3886306a89046ba5c989bc8b79008e7ec973228bad1fa20d7a94bba/gym_hil-0.1.14-py3-none-any.whl", hash = "sha256:9a2799d47a4561e0b0bb8d37fb3d84934657240be328d13991ea06758726533d", size = 5750805, upload-time = "2026-06-10T09:16:36.827Z" },
 ]
 
 [[package]]
@@ -1881,7 +1881,7 @@ sdist = { url = "https://files.pythonhosted.org/packages/e6/3e/ffad88145b342d5a9
 
 [[package]]
 name = "hf-libero"
-version = "0.1.3"
+version = "0.1.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "bddl", marker = "sys_platform == 'linux'" },
@@ -1902,7 +1902,10 @@ dependencies = [
     { name = "transformers", marker = "sys_platform == 'linux'" },
     { name = "wandb", marker = "sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/7e/ca/7f1c90aedcd067d608681cf03469ae548990ba0806f68a67927dcc801f04/hf_libero-0.1.3.tar.gz", hash = "sha256:0d6b9a215a658db86f66c03d063d6d877d2e9f96d2d326cfa9f43ba4da4a6d5a", size = 2960521, upload-time = "2025-11-03T17:58:00.003Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/af/aa/4e9eb8715e0bff9cb6553db563a35d253393097d446f82bd53575e8b253d/hf_libero-0.1.4.tar.gz", hash = "sha256:c058d67ad5a2b589529c14d614282ef4cca3a7763dafa134f58a6c9039657e34", size = 2961319, upload-time = "2026-06-10T09:56:13.994Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2a/79/c286b894c051988d062241682834df915c945bcf51009ffdffbe5ecf69bf/hf_libero-0.1.4-py3-none-any.whl", hash = "sha256:207f76e2f28bff30f78132223d8592fe8f64b1f8fd90ce7024948ada0d7e2c27", size = 3169084, upload-time = "2026-06-10T09:56:12.441Z" },
+]
 
 [[package]]
 name = "hf-xet"
@@ -3090,12 +3093,12 @@ requires-dist = [
     { name = "flash-attn", marker = "sys_platform != 'darwin' and extra == 'groot'", specifier = ">=2.5.9,<3.0.0" },
     { name = "grpcio", marker = "extra == 'grpcio-dep'", specifier = "==1.73.1" },
     { name = "grpcio-tools", marker = "extra == 'dev'", specifier = "==1.73.1" },
-    { name = "gym-aloha", marker = "extra == 'aloha'", specifier = ">=0.1.2,<0.2.0" },
-    { name = "gym-hil", marker = "extra == 'hilserl'", specifier = ">=0.1.13,<0.2.0" },
+    { name = "gym-aloha", marker = "extra == 'aloha'", specifier = ">=0.1.4,<0.2.0" },
+    { name = "gym-hil", marker = "extra == 'hilserl'", specifier = ">=0.1.14,<0.2.0" },
     { name = "gym-pusht", marker = "extra == 'pusht'", specifier = ">=0.1.5,<0.2.0" },
     { name = "gymnasium", specifier = ">=1.1.1,<2.0.0" },
     { name = "hebi-py", marker = "extra == 'phone'", specifier = ">=2.8.0,<2.12.0" },
-    { name = "hf-libero", marker = "sys_platform == 'linux' and extra == 'libero'", specifier = ">=0.1.3,<0.2.0" },
+    { name = "hf-libero", marker = "sys_platform == 'linux' and extra == 'libero'", specifier = ">=0.1.4,<0.2.0" },
     { name = "hidapi", marker = "extra == 'gamepad'", specifier = ">=0.14.0,<0.15.0" },
     { name = "huggingface-hub", specifier = ">=1.0.0,<2.0.0" },
     { name = "ipykernel", marker = "extra == 'notebook'", specifier = ">=6.0.0,<7.0.0" },