Merge remote-tracking branch 'origin/main' into user/rcadene/2024_07_16_control_robot_v2

.github/workflows/quality.yml

@@ -54,3 +54,31 @@ jobs:
 
       - name: Poetry check
         run: poetry check
+
+
+  poetry_relax:
+    name: Poetry relax
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Install poetry-relax
+        run: poetry self add poetry-relax
+
+      - name: Poetry relax
+        id: poetry_relax
+        run: |
+          output=$(poetry relax --check 2>&1)
+          if echo "$output" | grep -q "Proposing updates"; then
+            echo "$output"
+            echo ""
+            echo "Some dependencies have caret '^' version requirement added by poetry by default."
+            echo "Please replace them with '>='. You can do this by hand or use poetry-relax to do this."
+            exit 1
+          else
+            echo "$output"
+          fi

.github/workflows/trufflehog.yml

@@ -16,3 +16,5 @@ jobs:
         fetch-depth: 0
     - name: Secret Scanning
       uses: trufflesecurity/trufflehog@main
+      with:
+        extra_args: --only-verified

.pre-commit-config.yaml

@@ -14,11 +14,11 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.2
+    rev: v3.16.0
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.4.3
+    rev: v0.5.2
     hooks:
       - id: ruff
         args: [--fix]
@@ -31,3 +31,7 @@ repos:
         args:
           - "--check"
           - "--no-update"
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.18.4
+    hooks:
+      - id: gitleaks

Makefile

@@ -26,6 +26,7 @@ test-end-to-end:
 	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
 	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
 	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train-with-online
 	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
 	${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
 	${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
@@ -113,7 +114,6 @@ test-diffusion-ete-eval:
 		env.episode_length=8 \
 		device=$(DEVICE) \
 
-# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
@@ -133,6 +133,28 @@ test-tdmpc-ete-train:
 		training.image_transforms.enable=true \
 		hydra.run.dir=tests/outputs/tdmpc/
 
+test-tdmpc-ete-train-with-online:
+	python lerobot/scripts/train.py \
+		env=pusht \
+		env.gym.obs_type=environment_state_agent_pos \
+		policy=tdmpc_pusht_keypoints \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=10 \
+		device=$(DEVICE) \
+		training.offline_steps=2 \
+		training.online_steps=20 \
+		training.save_checkpoint=false \
+		training.save_freq=10 \
+		training.batch_size=2 \
+		training.online_rollout_n_episodes=2 \
+		training.online_rollout_batch_size=2 \
+		training.online_steps_between_rollouts=10 \
+		training.online_buffer_capacity=15 \
+		eval.use_async_envs=true \
+		hydra.run.dir=tests/outputs/tdmpc_online/
+
+
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
 		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \

README.md

@@ -180,8 +180,10 @@ dataset attributes:
   │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
   │  ...
   ├ info: a dictionary of metadata on the dataset
+  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
   │  ├ fps (float): frame per second the dataset is recorded/synchronized to
-  │  └ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  ├ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  └ encoding (dict): if video, this documents the main options that were used with ffmpeg to encode the videos
   ├ videos_dir (Path): where the mp4 videos or png images are stored/accessed
   └ camera_keys (list of string): the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
 ```

benchmarks/video/run_video_benchmark.py

@@ -257,10 +257,10 @@ def benchmark_encoding_decoding(
             imgs_dir=imgs_dir,
             video_path=video_path,
             fps=fps,
-            video_codec=encoding_cfg["vcodec"],
-            pixel_format=encoding_cfg["pix_fmt"],
-            group_of_pictures_size=encoding_cfg.get("g"),
-            constant_rate_factor=encoding_cfg.get("crf"),
+            vcodec=encoding_cfg["vcodec"],
+            pix_fmt=encoding_cfg["pix_fmt"],
+            g=encoding_cfg.get("g"),
+            crf=encoding_cfg.get("crf"),
             # fast_decode=encoding_cfg.get("fastdecode"),
             overwrite=True,
         )

lerobot/__init__.py

@@ -125,6 +125,10 @@ available_real_world_datasets = [
     "lerobot/aloha_static_vinh_cup_left",
     "lerobot/aloha_static_ziploc_slide",
     "lerobot/umi_cup_in_the_wild",
+    "lerobot/unitreeh1_fold_clothes",
+    "lerobot/unitreeh1_rearrange_objects",
+    "lerobot/unitreeh1_two_robot_greeting",
+    "lerobot/unitreeh1_warehouse",
 ]
 
 available_datasets = list(

lerobot/common/datasets/lerobot_dataset.py

@@ -35,9 +35,8 @@ from lerobot.common.datasets.utils import (
 )
 from lerobot.common.datasets.video_utils import VideoFrame, load_from_videos
 
-# For maintainers, see lerobot/common/datasets/push_dataset_to_hub/codebase_version.md
-CODEBASE_VERSION = "v1.5"
-
+# For maintainers, see lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md
+CODEBASE_VERSION = "v1.6"
 DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
 
 

lerobot/common/datasets/online_buffer.py

@@ -0,0 +1,384 @@
+#!/usr/bin/env python
+
+# Copyright 2024 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.
+"""An online buffer for the online training loop in train.py
+
+Note to maintainers: This duplicates some logic from LeRobotDataset and EpisodeAwareSampler. We should
+consider converging to one approach. Here we have opted to use numpy.memmap to back the data buffer. It's much
+faster than using HuggingFace Datasets as there's no conversion to an intermediate non-python object. Also it
+supports in-place slicing and mutation which is very handy for a dynamic buffer.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+def _make_memmap_safe(**kwargs) -> np.memmap:
+    """Make a numpy memmap with checks on available disk space first.
+
+    Expected kwargs are: "filename", "dtype" (must by np.dtype), "mode" and "shape"
+
+    For information on dtypes:
+    https://numpy.org/doc/stable/reference/arrays.dtypes.html#arrays-dtypes-constructing
+    """
+    if kwargs["mode"].startswith("w"):
+        required_space = kwargs["dtype"].itemsize * np.prod(kwargs["shape"])  # bytes
+        stats = os.statvfs(Path(kwargs["filename"]).parent)
+        available_space = stats.f_bavail * stats.f_frsize  # bytes
+        if required_space >= available_space * 0.8:
+            raise RuntimeError(
+                f"You're about to take up {required_space} of {available_space} bytes available."
+            )
+    return np.memmap(**kwargs)
+
+
+class OnlineBuffer(torch.utils.data.Dataset):
+    """FIFO data buffer for the online training loop in train.py.
+
+    Follows the protocol of LeRobotDataset as much as is required to have it be used by the online training
+    loop in the same way that a LeRobotDataset would be used.
+
+    The underlying data structure will have data inserted in a circular fashion. Always insert after the
+    last index, and when you reach the end, wrap around to the start.
+
+    The data is stored in a numpy memmap.
+    """
+
+    NEXT_INDEX_KEY = "_next_index"
+    OCCUPANCY_MASK_KEY = "_occupancy_mask"
+    INDEX_KEY = "index"
+    FRAME_INDEX_KEY = "frame_index"
+    EPISODE_INDEX_KEY = "episode_index"
+    TIMESTAMP_KEY = "timestamp"
+    IS_PAD_POSTFIX = "_is_pad"
+
+    def __init__(
+        self,
+        write_dir: str | Path,
+        data_spec: dict[str, Any] | None,
+        buffer_capacity: int | None,
+        fps: float | None = None,
+        delta_timestamps: dict[str, list[float]] | dict[str, np.ndarray] | None = None,
+    ):
+        """
+        The online buffer can be provided from scratch or you can load an existing online buffer by passing
+        a `write_dir` associated with an existing buffer.
+
+        Args:
+            write_dir: Where to keep the numpy memmap files. One memmap file will be stored for each data key.
+                Note that if the files already exist, they are opened in read-write mode (used for training
+                resumption.)
+            data_spec: A mapping from data key to data specification, like {data_key: {"shape": tuple[int],
+                "dtype": np.dtype}}. This should include all the data that you wish to record into the buffer,
+                but note that "index", "frame_index" and "episode_index" are already accounted for by this
+                class, so you don't need to include them.
+            buffer_capacity: How many frames should be stored in the buffer as a maximum. Be aware of your
+                system's available disk space when choosing this.
+            fps: Same as the fps concept in LeRobot dataset. Here it needs to be provided for the
+                 delta_timestamps logic. You can pass None if you are not using delta_timestamps.
+            delta_timestamps: Same as the delta_timestamps concept in LeRobotDataset. This is internally
+                converted to dict[str, np.ndarray] for optimization purposes.
+
+        """
+        self.set_delta_timestamps(delta_timestamps)
+        self._fps = fps
+        # Tolerance in seconds used to discard loaded frames when their timestamps are not close enough from
+        # the requested frames. It is only used when `delta_timestamps` is provided.
+        # minus 1e-4 to account for possible numerical error
+        self.tolerance_s = 1 / self.fps - 1e-4 if fps is not None else None
+        self._buffer_capacity = buffer_capacity
+        data_spec = self._make_data_spec(data_spec, buffer_capacity)
+        Path(write_dir).mkdir(parents=True, exist_ok=True)
+        self._data = {}
+        for k, v in data_spec.items():
+            self._data[k] = _make_memmap_safe(
+                filename=Path(write_dir) / k,
+                dtype=v["dtype"] if v is not None else None,
+                mode="r+" if (Path(write_dir) / k).exists() else "w+",
+                shape=tuple(v["shape"]) if v is not None else None,
+            )
+
+    @property
+    def delta_timestamps(self) -> dict[str, np.ndarray] | None:
+        return self._delta_timestamps
+
+    def set_delta_timestamps(self, value: dict[str, list[float]] | None):
+        """Set delta_timestamps converting the values to numpy arrays.
+
+        The conversion is for an optimization in the __getitem__. The loop is much slower if the arrays
+        need to be converted into numpy arrays.
+        """
+        if value is not None:
+            self._delta_timestamps = {k: np.array(v) for k, v in value.items()}
+        else:
+            self._delta_timestamps = None
+
+    def _make_data_spec(self, data_spec: dict[str, Any], buffer_capacity: int) -> dict[str, dict[str, Any]]:
+        """Makes the data spec for np.memmap."""
+        if any(k.startswith("_") for k in data_spec):
+            raise ValueError(
+                "data_spec keys should not start with '_'. This prefix is reserved for internal logic."
+            )
+        preset_keys = {
+            OnlineBuffer.INDEX_KEY,
+            OnlineBuffer.FRAME_INDEX_KEY,
+            OnlineBuffer.EPISODE_INDEX_KEY,
+            OnlineBuffer.TIMESTAMP_KEY,
+        }
+        if len(intersection := set(data_spec).intersection(preset_keys)) > 0:
+            raise ValueError(
+                f"data_spec should not contain any of {preset_keys} as these are handled internally. "
+                f"The provided data_spec has {intersection}."
+            )
+        complete_data_spec = {
+            # _next_index will be a pointer to the next index that we should start filling from when we add
+            # more data.
+            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
+            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
+            # with real data rather than the dummy initialization.
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
+            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
+        }
+        for k, v in data_spec.items():
+            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
+        return complete_data_spec
+
+    def add_data(self, data: dict[str, np.ndarray]):
+        """Add new data to the buffer, which could potentially mean shifting old data out.
+
+        The new data should contain all the frames (in order) of any number of episodes. The indices should
+        start from 0 (note to the developer: this can easily be generalized). See the `rollout` and
+        `eval_policy` functions in `eval.py` for more information on how the data is constructed.
+
+        Shift the incoming data index and episode_index to continue on from the last frame. Note that this
+        will be done in place!
+        """
+        if len(missing_keys := (set(self.data_keys).difference(set(data)))) > 0:
+            raise ValueError(f"Missing data keys: {missing_keys}")
+        new_data_length = len(data[self.data_keys[0]])
+        if not all(len(data[k]) == new_data_length for k in self.data_keys):
+            raise ValueError("All data items should have the same length")
+
+        next_index = self._data[OnlineBuffer.NEXT_INDEX_KEY]
+
+        # Sanity check to make sure that the new data indices start from 0.
+        assert data[OnlineBuffer.EPISODE_INDEX_KEY][0].item() == 0
+        assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
+
+        # Shift the incoming indices if necessary.
+        if self.num_samples > 0:
+            last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
+            last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
+            data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
+            data[OnlineBuffer.INDEX_KEY] += last_data_index + 1
+
+        # Insert the new data starting from next_index. It may be necessary to wrap around to the start.
+        n_surplus = max(0, new_data_length - (self._buffer_capacity - next_index))
+        for k in self.data_keys:
+            if n_surplus == 0:
+                slc = slice(next_index, next_index + new_data_length)
+                self._data[k][slc] = data[k]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][slc] = True
+            else:
+                self._data[k][next_index:] = data[k][:-n_surplus]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][next_index:] = True
+                self._data[k][:n_surplus] = data[k][-n_surplus:]
+        if n_surplus == 0:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = next_index + new_data_length
+        else:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = n_surplus
+
+    @property
+    def data_keys(self) -> list[str]:
+        keys = set(self._data)
+        keys.remove(OnlineBuffer.OCCUPANCY_MASK_KEY)
+        keys.remove(OnlineBuffer.NEXT_INDEX_KEY)
+        return sorted(keys)
+
+    @property
+    def fps(self) -> float | None:
+        return self._fps
+
+    @property
+    def num_episodes(self) -> int:
+        return len(
+            np.unique(self._data[OnlineBuffer.EPISODE_INDEX_KEY][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+        )
+
+    @property
+    def num_samples(self) -> int:
+        return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
+
+    def __len__(self):
+        return self.num_samples
+
+    def _item_to_tensors(self, item: dict) -> dict:
+        item_ = {}
+        for k, v in item.items():
+            if isinstance(v, torch.Tensor):
+                item_[k] = v
+            elif isinstance(v, np.ndarray):
+                item_[k] = torch.from_numpy(v)
+            else:
+                item_[k] = torch.tensor(v)
+        return item_
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self) or idx < -len(self):
+            raise IndexError
+
+        item = {k: v[idx] for k, v in self._data.items() if not k.startswith("_")}
+
+        if self.delta_timestamps is None:
+            return self._item_to_tensors(item)
+
+        episode_index = item[OnlineBuffer.EPISODE_INDEX_KEY]
+        current_ts = item[OnlineBuffer.TIMESTAMP_KEY]
+        episode_data_indices = np.where(
+            np.bitwise_and(
+                self._data[OnlineBuffer.EPISODE_INDEX_KEY] == episode_index,
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY],
+            )
+        )[0]
+        episode_timestamps = self._data[OnlineBuffer.TIMESTAMP_KEY][episode_data_indices]
+
+        for data_key in self.delta_timestamps:
+            # Note: The logic in this loop is copied from `load_previous_and_future_frames`.
+            # Get timestamps used as query to retrieve data of previous/future frames.
+            query_ts = current_ts + self.delta_timestamps[data_key]
+
+            # Compute distances between each query timestamp and all timestamps of all the frames belonging to
+            # the episode.
+            dist = np.abs(query_ts[:, None] - episode_timestamps[None, :])
+            argmin_ = np.argmin(dist, axis=1)
+            min_ = dist[np.arange(dist.shape[0]), argmin_]
+
+            is_pad = min_ > self.tolerance_s
+
+            # Check violated query timestamps are all outside the episode range.
+            assert (
+                (query_ts[is_pad] < episode_timestamps[0]) | (episode_timestamps[-1] < query_ts[is_pad])
+            ).all(), (
+                f"One or several timestamps unexpectedly violate the tolerance ({min_} > {self.tolerance_s=}"
+                ") inside the episode range."
+            )
+
+            # Load frames for this data key.
+            item[data_key] = self._data[data_key][episode_data_indices[argmin_]]
+
+            item[f"{data_key}{OnlineBuffer.IS_PAD_POSTFIX}"] = is_pad
+
+        return self._item_to_tensors(item)
+
+    def get_data_by_key(self, key: str) -> torch.Tensor:
+        """Returns all data for a given data key as a Tensor."""
+        return torch.from_numpy(self._data[key][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+
+
+def compute_sampler_weights(
+    offline_dataset: LeRobotDataset,
+    offline_drop_n_last_frames: int = 0,
+    online_dataset: OnlineBuffer | None = None,
+    online_sampling_ratio: float | None = None,
+    online_drop_n_last_frames: int = 0,
+) -> torch.Tensor:
+    """Compute the sampling weights for the online training dataloader in train.py.
+
+    Args:
+        offline_dataset: The LeRobotDataset used for offline pre-training.
+        online_drop_n_last_frames: Number of frames to drop from the end of each offline dataset episode.
+        online_dataset: The OnlineBuffer used in online training.
+        online_sampling_ratio: The proportion of data that should be sampled from the online dataset. If an
+            online dataset is provided, this value must also be provided.
+        online_drop_n_first_frames: See `offline_drop_n_last_frames`. This is the same, but for the online
+            dataset.
+    Returns:
+        Tensor of weights for [offline_dataset; online_dataset], normalized to 1.
+
+    Notes to maintainers:
+        - This duplicates some logic from EpisodeAwareSampler. We should consider converging to one approach.
+        - When used with `torch.utils.data.WeightedRandomSampler`, it could completely replace
+          `EpisodeAwareSampler` as the online dataset related arguments are optional. The only missing feature
+          is the ability to turn shuffling off.
+        - Options `drop_first_n_frames` and `episode_indices_to_use` can be added easily. They were not
+          included here to avoid adding complexity.
+    """
+    if len(offline_dataset) == 0 and (online_dataset is None or len(online_dataset) == 0):
+        raise ValueError("At least one of `offline_dataset` or `online_dataset` should be contain data.")
+    if (online_dataset is None) ^ (online_sampling_ratio is None):
+        raise ValueError(
+            "`online_dataset` and `online_sampling_ratio` must be provided together or not at all."
+        )
+    offline_sampling_ratio = 0 if online_sampling_ratio is None else 1 - online_sampling_ratio
+
+    weights = []
+
+    if len(offline_dataset) > 0:
+        offline_data_mask_indices = []
+        for start_index, end_index in zip(
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
+            strict=True,
+        ):
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
+        offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
+        offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(offline_dataset),),
+                fill_value=offline_sampling_ratio / offline_data_mask.sum(),
+            )
+            * offline_data_mask
+        )
+
+    if online_dataset is not None and len(online_dataset) > 0:
+        online_data_mask_indices = []
+        episode_indices = online_dataset.get_data_by_key("episode_index")
+        for episode_idx in torch.unique(episode_indices):
+            where_episode = torch.where(episode_indices == episode_idx)
+            start_index = where_episode[0][0]
+            end_index = where_episode[0][-1] + 1
+            online_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - online_drop_n_last_frames)
+            )
+        online_data_mask = torch.zeros(len(online_dataset), dtype=torch.bool)
+        online_data_mask[torch.tensor(online_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(online_dataset),),
+                fill_value=online_sampling_ratio / online_data_mask.sum(),
+            )
+            * online_data_mask
+        )
+
+    weights = torch.cat(weights)
+
+    if weights.sum() == 0:
+        weights += 1 / len(weights)
+    else:
+        weights /= weights.sum()
+
+    return weights

lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md

@@ -10,7 +10,8 @@ For instance, [`lerobot/pusht`](https://huggingface.co/datasets/lerobot/pusht) h
 - [v1.2](https://huggingface.co/datasets/lerobot/pusht/tree/v1.2)
 - [v1.3](https://huggingface.co/datasets/lerobot/pusht/tree/v1.3)
 - [v1.4](https://huggingface.co/datasets/lerobot/pusht/tree/v1.4)
-- [v1.5](https://huggingface.co/datasets/lerobot/pusht/tree/v1.5) <-- last version
+- [v1.5](https://huggingface.co/datasets/lerobot/pusht/tree/v1.5)
+- [v1.6](https://huggingface.co/datasets/lerobot/pusht/tree/v1.6) <-- last version
 - [main](https://huggingface.co/datasets/lerobot/pusht/tree/main) <-- points to the last version
 
 Starting with v1.6, every dataset pushed to the hub or saved locally also have this version number in their
@@ -45,13 +46,11 @@ for repo_id in available_datasets:
     dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
     branches = [b.name for b in dataset_info.branches]
     if CODEBASE_VERSION in branches:
-        # First check if the newer version already exists.
-        print(f"Found existing branch for {repo_id}. Please contact a member of the core LeRobot team.")
-        print("Exiting early")
-        break
+        print(f"{repo_id} already @{CODEBASE_VERSION}, skipping.")
+        continue
     else:
         # Now create a branch named after the new version by branching out from "main"
         # which is expected to be the preceding version
         api.create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION, revision="main")
-        print(f"{repo_id} successfully updated")
+        print(f"{repo_id} successfully updated @{CODEBASE_VERSION}")
 ```

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -19,8 +19,8 @@ This file contains download scripts for raw datasets.
 Example of usage:
 ```
 python lerobot/common/datasets/push_dataset_to_hub/_download_raw.py \
---raw-dir data/cadene/pusht_raw \
---repo-id cadene/pusht_raw
+--raw-dir data/lerobot-raw/pusht_raw \
+--repo-id lerobot-raw/pusht_raw
 ```
 """
 
@@ -31,63 +31,65 @@ from pathlib import Path
 
 from huggingface_hub import snapshot_download
 
-AVAILABLE_RAW_REPO_IDS = [
-    "lerobot-raw/aloha_mobile_cabinet_raw",
-    "lerobot-raw/aloha_mobile_chair_raw",
-    "lerobot-raw/aloha_mobile_elevator_raw",
-    "lerobot-raw/aloha_mobile_shrimp_raw",
-    "lerobot-raw/aloha_mobile_wash_pan_raw",
-    "lerobot-raw/aloha_mobile_wipe_wine_raw",
-    "lerobot-raw/aloha_sim_insertion_human_raw",
-    "lerobot-raw/aloha_sim_insertion_scripted_raw",
-    "lerobot-raw/aloha_sim_transfer_cube_human_raw",
-    "lerobot-raw/aloha_sim_transfer_cube_scripted_raw",
-    "lerobot-raw/aloha_static_battery_raw",
-    "lerobot-raw/aloha_static_candy_raw",
-    "lerobot-raw/aloha_static_coffee_new_raw",
-    "lerobot-raw/aloha_static_coffee_raw",
-    "lerobot-raw/aloha_static_cups_open_raw",
-    "lerobot-raw/aloha_static_fork_pick_up_raw",
-    "lerobot-raw/aloha_static_pingpong_test_raw",
-    "lerobot-raw/aloha_static_pro_pencil_raw",
-    "lerobot-raw/aloha_static_screw_driver_raw",
-    "lerobot-raw/aloha_static_tape_raw",
-    "lerobot-raw/aloha_static_thread_velcro_raw",
-    "lerobot-raw/aloha_static_towel_raw",
-    "lerobot-raw/aloha_static_vinh_cup_left_raw",
-    "lerobot-raw/aloha_static_vinh_cup_raw",
-    "lerobot-raw/aloha_static_ziploc_slide_raw",
-    "lerobot-raw/pusht_raw",
-    "lerobot-raw/umi_cup_in_the_wild_raw",
-    "lerobot-raw/unitreeh1_fold_clothes_raw",
-    "lerobot-raw/unitreeh1_rearrange_objects_raw",
-    "lerobot-raw/unitreeh1_two_robot_greeting_raw",
-    "lerobot-raw/unitreeh1_warehouse_raw",
-    "lerobot-raw/xarm_lift_medium_raw",
-    "lerobot-raw/xarm_lift_medium_replay_raw",
-    "lerobot-raw/xarm_push_medium_raw",
-    "lerobot-raw/xarm_push_medium_replay_raw",
-]
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+
+# {raw_repo_id: raw_format}
+AVAILABLE_RAW_REPO_IDS = {
+    "lerobot-raw/aloha_mobile_cabinet_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_chair_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_elevator_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_shrimp_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wash_pan_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wipe_wine_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_battery_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_candy_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_new_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_cups_open_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_fork_pick_up_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pingpong_test_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pro_pencil_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_screw_driver_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_tape_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_thread_velcro_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_towel_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
+    "lerobot-raw/pusht_raw": "pusht_zarr",
+    "lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
+    "lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_warehouse_raw": "aloha_hdf5",
+    "lerobot-raw/xarm_lift_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
+}
 
 
 def download_raw(raw_dir: Path, repo_id: str):
-    # Check repo_id is well formated
-    if len(repo_id.split("/")) != 2:
-        raise ValueError(
-            f"`repo_id` is expected to contain a community or user id `/` the name of the dataset (e.g. 'lerobot/pusht'), but contains '{repo_id}'."
-        )
+    check_repo_id(repo_id)
     user_id, dataset_id = repo_id.split("/")
 
     if not dataset_id.endswith("_raw"):
         warnings.warn(
-            f"`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this naming convention by renaming your repository is advised, but not mandatory.",
+            f"""`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this
+             naming convention by renaming your repository is advised, but not mandatory.""",
             stacklevel=1,
         )
 
     # Send warning if raw_dir isn't well formated
     if raw_dir.parts[-2] != user_id or raw_dir.parts[-1] != dataset_id:
         warnings.warn(
-            f"`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised, but not mandatory.",
+            f"""`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that
+             match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised,
+             but not mandatory.""",
             stacklevel=1,
         )
     raw_dir.mkdir(parents=True, exist_ok=True)
@@ -97,8 +99,9 @@ def download_raw(raw_dir: Path, repo_id: str):
     logging.info(f"Finish downloading from huggingface.co/{user_id} for {dataset_id}")
 
 
-def download_all_raw_datasets():
-    data_dir = Path("data")
+def download_all_raw_datasets(data_dir: Path | None = None):
+    if data_dir is None:
+        data_dir = Path("data")
     for repo_id in AVAILABLE_RAW_REPO_IDS:
         raw_dir = data_dir / repo_id
         download_raw(raw_dir, repo_id)
@@ -106,7 +109,8 @@ def download_all_raw_datasets():
 
 def main():
     parser = argparse.ArgumentParser(
-        description=f"A script to download raw datasets from Hugging Face hub to a local directory. Here is a non exhaustive list of available repositories to use in `--repo-id`: {AVAILABLE_RAW_REPO_IDS}",
+        description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
+            non exhaustive list of available repositories to use in `--repo-id`: {AVAILABLE_RAW_REPO_IDS}""",
     )
 
     parser.add_argument(
@@ -119,7 +123,8 @@ def main():
         "--repo-id",
         type=str,
         required=True,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).",
+        help="""Repositery identifier on Hugging Face: a community or a user name `/` the name of
+        the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).""",
     )
     args = parser.parse_args()
     download_raw(**vars(args))

lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py

@@ -0,0 +1,184 @@
+#!/usr/bin/env python
+
+# Copyright 2024 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.
+"""
+Use this script to batch encode lerobot dataset from their raw format to LeRobotDataset and push their updated
+version to the hub. Under the hood, this script reuses 'push_dataset_to_hub.py'. It assumes that you already
+downloaded raw datasets, which you can do with the related '_download_raw.py' script.
+
+For instance, for codebase_version = 'v1.6', the following command was run, assuming raw datasets from
+lerobot-raw were downloaded in 'raw/datasets/directory':
+```bash
+python lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py \
+  --raw-dir raw/datasets/directory \
+  --raw-repo-ids lerobot-raw \
+  --local-dir push/datasets/directory \
+  --tests-data-dir tests/data \
+  --push-repo lerobot \
+  --vcodec libsvtav1 \
+  --pix-fmt yuv420p \
+  --g 2 \
+  --crf 30
+```
+"""
+
+import argparse
+from pathlib import Path
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import AVAILABLE_RAW_REPO_IDS
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+from lerobot.scripts.push_dataset_to_hub import push_dataset_to_hub
+
+
+def get_push_repo_id_from_raw(raw_repo_id: str, push_repo: str) -> str:
+    dataset_id_raw = raw_repo_id.split("/")[1]
+    dataset_id = dataset_id_raw.removesuffix("_raw")
+    return f"{push_repo}/{dataset_id}"
+
+
+def encode_datasets(
+    raw_dir: Path,
+    raw_repo_ids: list[str],
+    push_repo: str,
+    vcodec: str,
+    pix_fmt: str,
+    g: int,
+    crf: int,
+    local_dir: Path | None = None,
+    tests_data_dir: Path | None = None,
+    raw_format: str | None = None,
+    dry_run: bool = False,
+) -> None:
+    if len(raw_repo_ids) == 1 and raw_repo_ids[0].lower() == "lerobot-raw":
+        raw_repo_ids_format = AVAILABLE_RAW_REPO_IDS
+    else:
+        if raw_format is None:
+            raise ValueError(raw_format)
+        raw_repo_ids_format = {id_: raw_format for id_ in raw_repo_ids}
+
+    for raw_repo_id, repo_raw_format in raw_repo_ids_format.items():
+        check_repo_id(raw_repo_id)
+        dataset_repo_id_push = get_push_repo_id_from_raw(raw_repo_id, push_repo)
+        dataset_raw_dir = raw_dir / raw_repo_id
+        dataset_dir = local_dir / dataset_repo_id_push if local_dir is not None else None
+        encoding = {
+            "vcodec": vcodec,
+            "pix_fmt": pix_fmt,
+            "g": g,
+            "crf": crf,
+        }
+
+        if not (dataset_raw_dir).is_dir():
+            raise NotADirectoryError(dataset_raw_dir)
+
+        if not dry_run:
+            push_dataset_to_hub(
+                dataset_raw_dir,
+                raw_format=repo_raw_format,
+                repo_id=dataset_repo_id_push,
+                local_dir=dataset_dir,
+                resume=True,
+                encoding=encoding,
+                tests_data_dir=tests_data_dir,
+            )
+        else:
+            print(
+                f"DRY RUN: {dataset_raw_dir}  -->  {dataset_dir}  -->  {dataset_repo_id_push}@{CODEBASE_VERSION}"
+            )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        default=Path("data"),
+        help="Directory where raw datasets are located.",
+    )
+    parser.add_argument(
+        "--raw-repo-ids",
+        type=str,
+        nargs="*",
+        default=["lerobot-raw"],
+        help="""Raw dataset repo ids. if 'lerobot-raw', the keys from `AVAILABLE_RAW_REPO_IDS` will be
+            used and raw datasets will be fetched from the 'lerobot-raw/' repo and pushed with their
+            associated format. It is assumed that each dataset is located at `raw_dir / raw_repo_id` """,
+    )
+    parser.add_argument(
+        "--raw-format",
+        type=str,
+        default=None,
+        help="""Raw format to use for the raw repo-ids. Must be specified if --raw-repo-ids is not
+            'lerobot-raw'""",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="""When provided, writes the dataset converted to LeRobotDataset format in this directory
+        (e.g. `data/lerobot/aloha_mobile_chair`).""",
+    )
+    parser.add_argument(
+        "--push-repo",
+        type=str,
+        default="lerobot",
+        help="Repo to upload datasets to",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        default="libsvtav1",
+        help="Codec to use for encoding videos",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        default="yuv420p",
+        help="Pixel formats (chroma subsampling) to be used for encoding",
+    )
+    parser.add_argument(
+        "--g",
+        type=int,
+        default=2,
+        help="Group of pictures sizes to be used for encoding.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=int,
+        default=30,
+        help="Constant rate factors to be used for encoding.",
+    )
+    parser.add_argument(
+        "--tests-data-dir",
+        type=Path,
+        default=None,
+        help=(
+            "When provided, save tests artifacts into the given directory "
+            "(e.g. `--tests-data-dir tests/data` will save to tests/data/{--repo-id})."
+        ),
+    )
+    parser.add_argument(
+        "--dry-run",
+        type=int,
+        default=0,
+        help="If not set to 0, this script won't download or upload anything.",
+    )
+    args = parser.parse_args()
+    encode_datasets(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -29,7 +29,11 @@ from datasets import Dataset, Features, Image, Sequence, Value
 from PIL import Image as PILImage
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
 from lerobot.common.datasets.utils import (
     calculate_episode_data_index,
     hf_transform_to_torch,
@@ -72,7 +76,14 @@ def check_format(raw_dir) -> bool:
                     assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
 
 
-def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
     # only frames from simulation are uncompressed
     compressed_images = "sim" not in raw_dir.name
 
@@ -123,7 +134,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
                     # encode images to a mp4 video
                     fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
                     video_path = videos_dir / fname
-                    encode_video_frames(tmp_imgs_dir, video_path, fps)
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
 
                     # clean temporary images directory
                     shutil.rmtree(tmp_imgs_dir)
@@ -200,6 +211,7 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
     # sanity check
     check_format(raw_dir)
@@ -207,7 +219,7 @@ def from_raw_to_lerobot_format(
     if fps is None:
         fps = 50
 
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
     hf_dataset = to_hf_dataset(data_dict, video)
     episode_data_index = calculate_episode_data_index(hf_dataset)
     info = {
@@ -215,4 +227,7 @@ def from_raw_to_lerobot_format(
         "fps": fps,
         "video": video,
     }
+    if video:
+        info["encoding"] = get_default_encoding()
+
     return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -81,8 +81,9 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
-    if video or episodes is not None:
+    if video or episodes or encoding is not None:
         # TODO(aliberts): support this
         raise NotImplementedError
 

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -18,6 +18,7 @@ Contains utilities to process raw data format from dora-record
 """
 
 import re
+import warnings
 from pathlib import Path
 
 import pandas as pd
@@ -199,6 +200,7 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
     # sanity check
     check_format(raw_dir)
@@ -211,6 +213,12 @@ def from_raw_to_lerobot_format(
     if not video:
         raise NotImplementedError()
 
+    if encoding is not None:
+        warnings.warn(
+            "Video encoding is currently done outside of LeRobot for the dora_parquet format.",
+            stacklevel=1,
+        )
+
     data_df = load_from_raw(raw_dir, videos_dir, fps, episodes)
     hf_dataset = to_hf_dataset(data_df, video)
     episode_data_index = calculate_episode_data_index(hf_dataset)
@@ -219,4 +227,7 @@ def from_raw_to_lerobot_format(
         "fps": fps,
         "video": video,
     }
+    if video:
+        info["encoding"] = "unknown"
+
     return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -26,7 +26,11 @@ from datasets import Dataset, Features, Image, Sequence, Value
 from PIL import Image as PILImage
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
 from lerobot.common.datasets.utils import (
     calculate_episode_data_index,
     hf_transform_to_torch,
@@ -62,6 +66,7 @@ def load_from_raw(
     video: bool,
     episodes: list[int] | None = None,
     keypoints_instead_of_image: bool = False,
+    encoding: dict | None = None,
 ):
     try:
         import pymunk
@@ -172,7 +177,7 @@ def load_from_raw(
                 # encode images to a mp4 video
                 fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
                 video_path = videos_dir / fname
-                encode_video_frames(tmp_imgs_dir, video_path, fps)
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
 
                 # clean temporary images directory
                 shutil.rmtree(tmp_imgs_dir)
@@ -244,6 +249,7 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
     # Manually change this to True to use keypoints of the T instead of an image observation (but don't merge
     # with True). Also make sure to use video = 0 in the `push_dataset_to_hub.py` script.
@@ -255,7 +261,7 @@ def from_raw_to_lerobot_format(
     if fps is None:
         fps = 10
 
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image)
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image, encoding)
     hf_dataset = to_hf_dataset(data_dict, video, keypoints_instead_of_image)
     episode_data_index = calculate_episode_data_index(hf_dataset)
     info = {
@@ -263,4 +269,7 @@ def from_raw_to_lerobot_format(
         "fps": fps,
         "video": video if not keypoints_instead_of_image else 0,
     }
+    if video:
+        info["encoding"] = get_default_encoding()
+
     return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -27,7 +27,11 @@ from PIL import Image as PILImage
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
 from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
 from lerobot.common.datasets.utils import (
     calculate_episode_data_index,
     hf_transform_to_torch,
@@ -60,7 +64,14 @@ def check_format(raw_dir) -> bool:
     assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
 
 
-def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
     zarr_path = raw_dir / "cup_in_the_wild.zarr"
     zarr_data = zarr.open(zarr_path, mode="r")
 
@@ -88,49 +99,61 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
         to_ids.append(to_idx)
         from_idx = to_idx
 
+    ep_dicts_dir = videos_dir / "ep_dicts"
+    ep_dicts_dir.mkdir(exist_ok=True, parents=True)
     ep_dicts = []
+
     ep_ids = episodes if episodes else range(num_episodes)
     for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
-        from_idx = from_ids[selected_ep_idx]
-        to_idx = to_ids[selected_ep_idx]
-        num_frames = to_idx - from_idx
+        ep_dict_path = ep_dicts_dir / f"{ep_idx}"
+        if not ep_dict_path.is_file():
+            from_idx = from_ids[selected_ep_idx]
+            to_idx = to_ids[selected_ep_idx]
+            num_frames = to_idx - from_idx
 
-        # TODO(rcadene): save temporary images of the episode?
+            # TODO(rcadene): save temporary images of the episode?
 
-        state = states[from_idx:to_idx]
+            state = states[from_idx:to_idx]
 
-        ep_dict = {}
+            ep_dict = {}
 
-        # load 57MB of images in RAM (400x224x224x3 uint8)
-        imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
-        img_key = "observation.image"
-        if video:
-            # save png images in temporary directory
-            tmp_imgs_dir = videos_dir / "tmp_images"
-            save_images_concurrently(imgs_array, tmp_imgs_dir)
+            # load 57MB of images in RAM (400x224x224x3 uint8)
+            imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
+            img_key = "observation.image"
+            if video:
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                if not video_path.is_file():
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
 
-            # encode images to a mp4 video
-            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
-            video_path = videos_dir / fname
-            encode_video_frames(tmp_imgs_dir, video_path, fps)
+                    # encode images to a mp4 video
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
 
-            # clean temporary images directory
-            shutil.rmtree(tmp_imgs_dir)
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
 
-            # store the reference to the video frame
-            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["episode_data_index_from"] = torch.tensor([from_idx] * num_frames)
+            ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
+            ep_dict["end_pose"] = end_pose[from_idx:to_idx]
+            ep_dict["start_pos"] = start_pos[from_idx:to_idx]
+            ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
+            torch.save(ep_dict, ep_dict_path)
         else:
-            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+            ep_dict = torch.load(ep_dict_path)
 
-        ep_dict["observation.state"] = state
-        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
-        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
-        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
-        ep_dict["episode_data_index_from"] = torch.tensor([from_idx] * num_frames)
-        ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
-        ep_dict["end_pose"] = end_pose[from_idx:to_idx]
-        ep_dict["start_pos"] = start_pos[from_idx:to_idx]
-        ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
         ep_dicts.append(ep_dict)
 
     data_dict = concatenate_episodes(ep_dicts)
@@ -183,6 +206,7 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
     # sanity check
     check_format(raw_dir)
@@ -196,7 +220,7 @@ def from_raw_to_lerobot_format(
             "Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
         )
 
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
     hf_dataset = to_hf_dataset(data_dict, video)
     episode_data_index = calculate_episode_data_index(hf_dataset)
     info = {
@@ -204,4 +228,7 @@ def from_raw_to_lerobot_format(
         "fps": fps,
         "video": video,
     }
+    if video:
+        info["encoding"] = get_default_encoding()
+
     return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -13,6 +13,7 @@
 # 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.
+import inspect
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
@@ -20,6 +21,8 @@ import numpy
 import PIL
 import torch
 
+from lerobot.common.datasets.video_utils import encode_video_frames
+
 
 def concatenate_episodes(ep_dicts):
     data_dict = {}
@@ -51,3 +54,21 @@ def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers
     num_images = len(imgs_array)
     with ThreadPoolExecutor(max_workers=max_workers) as executor:
         [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
+
+
+def get_default_encoding() -> dict:
+    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
+    signature = inspect.signature(encode_video_frames)
+    return {
+        k: v.default
+        for k, v in signature.parameters.items()
+        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
+    }
+
+
+def check_repo_id(repo_id: str) -> None:
+    if len(repo_id.split("/")) != 2:
+        raise ValueError(
+            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
+            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
+        )

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -26,7 +26,11 @@ from datasets import Dataset, Features, Image, Sequence, Value
 from PIL import Image as PILImage
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
 from lerobot.common.datasets.utils import (
     calculate_episode_data_index,
     hf_transform_to_torch,
@@ -56,7 +60,14 @@ def check_format(raw_dir):
         assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
 
 
-def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
     pkl_path = raw_dir / "buffer.pkl"
 
     with open(pkl_path, "rb") as f:
@@ -105,7 +116,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
             # encode images to a mp4 video
             fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
             video_path = videos_dir / fname
-            encode_video_frames(tmp_imgs_dir, video_path, fps)
+            encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
 
             # clean temporary images directory
             shutil.rmtree(tmp_imgs_dir)
@@ -167,6 +178,7 @@ def from_raw_to_lerobot_format(
     fps: int | None = None,
     video: bool = True,
     episodes: list[int] | None = None,
+    encoding: dict | None = None,
 ):
     # sanity check
     check_format(raw_dir)
@@ -174,7 +186,7 @@ def from_raw_to_lerobot_format(
     if fps is None:
         fps = 15
 
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
     hf_dataset = to_hf_dataset(data_dict, video)
     episode_data_index = calculate_episode_data_index(hf_dataset)
     info = {
@@ -182,4 +194,7 @@ def from_raw_to_lerobot_format(
         "fps": fps,
         "video": video,
     }
+    if video:
+        info["encoding"] = get_default_encoding()
+
     return hf_dataset, episode_data_index, info

lerobot/common/datasets/video_utils.py

@@ -166,10 +166,10 @@ def encode_video_frames(
     imgs_dir: Path,
     video_path: Path,
     fps: int,
-    video_codec: str = "libsvtav1",
-    pixel_format: str = "yuv420p",
-    group_of_pictures_size: int | None = 2,
-    constant_rate_factor: int | None = 30,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int | None = 2,
+    crf: int | None = 30,
     fast_decode: int = 0,
     log_level: str | None = "error",
     overwrite: bool = False,
@@ -183,20 +183,20 @@ def encode_video_frames(
             ("-f", "image2"),
             ("-r", str(fps)),
             ("-i", str(imgs_dir / "frame_%06d.png")),
-            ("-vcodec", video_codec),
-            ("-pix_fmt", pixel_format),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
         ]
     )
 
-    if group_of_pictures_size is not None:
-        ffmpeg_args["-g"] = str(group_of_pictures_size)
+    if g is not None:
+        ffmpeg_args["-g"] = str(g)
 
-    if constant_rate_factor is not None:
-        ffmpeg_args["-crf"] = str(constant_rate_factor)
+    if crf is not None:
+        ffmpeg_args["-crf"] = str(crf)
 
     if fast_decode:
-        key = "-svtav1-params" if video_codec == "libsvtav1" else "-tune"
-        value = f"fast-decode={fast_decode}" if video_codec == "libsvtav1" else "fastdecode"
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
+        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         ffmpeg_args[key] = value
 
     if log_level is not None:

lerobot/common/policies/act/modeling_act.py

@@ -101,6 +101,7 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
 
         batch = self.normalize_inputs(batch)
         if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
 
         # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
@@ -128,6 +129,7 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
         if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
@@ -467,10 +469,9 @@ class ACT(nn.Module):
         if self.use_images:
             all_cam_features = []
             all_cam_pos_embeds = []
-            images = batch["observation.images"]
 
-            for cam_index in range(images.shape[-4]):
-                cam_features = self.backbone(images[:, cam_index])["feature_map"]
+            for cam_index in range(batch["observation.images"].shape[-4]):
+                cam_features = self.backbone(batch["observation.images"][:, cam_index])["feature_map"]
                 # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use
                 # buffer
                 cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -111,17 +111,18 @@ class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
         Schematically this looks like:
             ----------------------------------------------------------------------------------------------
             (legend: o = n_obs_steps, h = horizon, a = n_action_steps)
-            |timestep            | n-o+1 | n-o+2 | ..... | n     | ..... | n+a-1 | n+a   | ..... |n-o+1+h|
-            |observation is used | YES   | YES   | YES   | NO    | NO    | NO    | NO    | NO    | NO    |
+            |timestep            | n-o+1 | n-o+2 | ..... | n     | ..... | n+a-1 | n+a   | ..... | n-o+h |
+            |observation is used | YES   | YES   | YES   | YES   | NO    | NO    | NO    | NO    | NO    |
             |action is generated | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   |
             |action is used      | NO    | NO    | NO    | YES   | YES   | YES   | NO    | NO    | NO    |
             ----------------------------------------------------------------------------------------------
-        Note that this means we require: `n_action_steps < horizon - n_obs_steps + 1`. Also, note that
+        Note that this means we require: `n_action_steps <= horizon - n_obs_steps + 1`. Also, note that
         "horizon" may not the best name to describe what the variable actually means, because this period is
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
         """
         batch = self.normalize_inputs(batch)
         if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
@@ -143,6 +144,7 @@ class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
         if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)

lerobot/common/policies/normalize.py

@@ -132,6 +132,7 @@ class Normalize(nn.Module):
     # TODO(rcadene): should we remove torch.no_grad?
     @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, mode in self.modes.items():
             buffer = getattr(self, "buffer_" + key.replace(".", "_"))
 
@@ -197,6 +198,7 @@ class Unnormalize(nn.Module):
     # TODO(rcadene): should we remove torch.no_grad?
     @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, mode in self.modes.items():
             buffer = getattr(self, "buffer_" + key.replace(".", "_"))
 

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -25,12 +25,16 @@ class TDMPCConfig:
     camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift`.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
 
     Args:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
             action repeats in Q-learning or ask your favorite chatbot)
         horizon: Horizon for model predictive control.
+        n_action_steps: Number of action steps to take from the plan given by model predictive control. This
+            is an alternative to using action repeats. If this is set to more than 1, then we require
+            `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
+            approach of using multiple steps from the plan is not in the original implementation.
         input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
             the input data name, and the value is a list indicating the dimensions of the corresponding data.
             For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
@@ -100,6 +104,7 @@ class TDMPCConfig:
     # Input / output structure.
     n_action_repeats: int = 2
     horizon: int = 5
+    n_action_steps: int = 1
 
     input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
@@ -158,17 +163,18 @@ class TDMPCConfig:
         """Input validation (not exhaustive)."""
         # There should only be one image key.
         image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
-        if len(image_keys) != 1:
+        if len(image_keys) > 1:
             raise ValueError(
-                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
-            )
-        image_key = next(iter(image_keys))
-        if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
-            # TODO(alexander-soare): This limitation is solely because of code in the random shift
-            # augmentation. It should be able to be removed.
-            raise ValueError(
-                f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                f"{self.__class__.__name__} handles at most one image for now. Got image keys {image_keys}."
             )
+        if len(image_keys) > 0:
+            image_key = next(iter(image_keys))
+            if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
+                # TODO(alexander-soare): This limitation is solely because of code in the random shift
+                # augmentation. It should be able to be removed.
+                raise ValueError(
+                    f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                )
         if self.n_gaussian_samples <= 0:
             raise ValueError(
                 f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
@@ -179,3 +185,12 @@ class TDMPCConfig:
                 f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
                 "information."
             )
+        if self.n_action_steps > 1:
+            if self.n_action_repeats != 1:
+                raise ValueError(
+                    "If `n_action_steps > 1`, `n_action_repeats` must be left to its default value of 1."
+                )
+            if not self.use_mpc:
+                raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
+            if self.n_action_steps > self.horizon:
+                raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -19,14 +19,10 @@
 The comments in this code may sometimes refer to these references:
     TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://arxiv.org/abs/2203.04955)
     FOWM paper: Finetuning Offline World Models in the Real World (https://arxiv.org/abs/2310.16029)
-
-TODO(alexander-soare): Make rollout work for batch sizes larger than 1.
-TODO(alexander-soare): Use batch-first throughout.
 """
 
 # ruff: noqa: N806
 
-import logging
 from collections import deque
 from copy import deepcopy
 from functools import partial
@@ -56,9 +52,11 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             process communication to use the xarm environment from FOWM. This is because our xarm
             environment uses newer dependencies and does not match the environment in FOWM. See
             https://github.com/huggingface/lerobot/pull/103 for implementation details.
-        - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+        - We have NOT checked that training on LeRobot reproduces the results from FOWM.
+        - Nevertheless, we have verified that we can train TD-MPC for PushT. See
+          `lerobot/configs/policy/tdmpc_pusht_keypoints.yaml`.
         - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
-            match our xarm environment.
+          match our xarm environment.
     """
 
     name = "tdmpc"
@@ -74,22 +72,6 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__()
-        logging.warning(
-            """
-            Please note several warnings for this policy.
-
-            - Evaluation of pretrained weights created with the original FOWM code
-              (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
-              model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
-              to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
-              process communication to use the xarm environment from FOWM. This is because our xarm
-              environment uses newer dependencies and does not match the environment in FOWM. See
-              https://github.com/huggingface/lerobot/pull/103 for implementation details.
-            - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
-            - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
-              match our xarm environment.
-            """
-        )
 
         if config is None:
             config = TDMPCConfig()
@@ -114,8 +96,14 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
         # Note: This check is covered in the post-init of the config but have a sanity check just in case.
-        assert len(image_keys) == 1
-        self.input_image_key = image_keys[0]
+        self._use_image = False
+        self._use_env_state = False
+        if len(image_keys) > 0:
+            assert len(image_keys) == 1
+            self._use_image = True
+            self.input_image_key = image_keys[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
 
         self.reset()
 
@@ -125,10 +113,13 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         called on `env.reset()`
         """
         self._queues = {
-            "observation.image": deque(maxlen=1),
             "observation.state": deque(maxlen=1),
-            "action": deque(maxlen=self.config.n_action_repeats),
+            "action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
         }
+        if self._use_image:
+            self._queues["observation.image"] = deque(maxlen=1)
+        if self._use_env_state:
+            self._queues["observation.environment_state"] = deque(maxlen=1)
         # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
         # CEM for the next step.
         self._prev_mean: torch.Tensor | None = None
@@ -137,7 +128,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations."""
         batch = self.normalize_inputs(batch)
-        batch["observation.image"] = batch[self.input_image_key]
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -151,49 +144,57 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
                 batch[key] = batch[key][:, 0]
 
             # NOTE: Order of observations matters here.
-            z = self.model.encode({k: batch[k] for k in ["observation.image", "observation.state"]})
-            if self.config.use_mpc:
-                batch_size = batch["observation.image"].shape[0]
-                # Batch processing is not handled in MPC mode, so process the batch in a loop.
-                action = []  # will be a batch of actions for one step
-                for i in range(batch_size):
-                    # Note: self.plan does not handle batches, hence the squeeze.
-                    action.append(self.plan(z[i]))
-                action = torch.stack(action)
+            encode_keys = []
+            if self._use_image:
+                encode_keys.append("observation.image")
+            if self._use_env_state:
+                encode_keys.append("observation.environment_state")
+            encode_keys.append("observation.state")
+            z = self.model.encode({k: batch[k] for k in encode_keys})
+            if self.config.use_mpc:  # noqa: SIM108
+                actions = self.plan(z)  # (horizon, batch, action_dim)
             else:
-                # Plan with the policy (π) alone.
-                action = self.model.pi(z)
+                # Plan with the policy (π) alone. This always returns one action so unsqueeze to get a
+                # sequence dimension like in the MPC branch.
+                actions = self.model.pi(z).unsqueeze(0)
 
-            self.unnormalize_outputs({"action": action})["action"]
+            actions = torch.clamp(actions, -1, +1)
 
-            for _ in range(self.config.n_action_repeats):
-                self._queues["action"].append(action)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.n_action_repeats > 1:
+                for _ in range(self.config.n_action_repeats):
+                    self._queues["action"].append(actions[0])
+            else:
+                # Action queue is (n_action_steps, batch_size, action_dim), so we transpose the action.
+                self._queues["action"].extend(actions[: self.config.n_action_steps])
 
         action = self._queues["action"].popleft()
-        return torch.clamp(action, -1, 1)
+        return action
 
     @torch.no_grad()
     def plan(self, z: Tensor) -> Tensor:
-        """Plan next action using TD-MPC inference.
+        """Plan sequence of actions using TD-MPC inference.
 
         Args:
-            z: (latent_dim,) tensor for the initial state.
+            z: (batch, latent_dim,) tensor for the initial state.
         Returns:
-            (action_dim,) tensor for the next action.
-
-        TODO(alexander-soare) Extend this to be able to work with batches.
+            (horizon, batch, action_dim,) tensor for the planned trajectory of actions.
         """
         device = get_device_from_parameters(self)
 
+        batch_size = z.shape[0]
+
         # Sample Nπ trajectories from the policy.
         pi_actions = torch.empty(
             self.config.horizon,
             self.config.n_pi_samples,
+            batch_size,
             self.config.output_shapes["action"][0],
             device=device,
         )
         if self.config.n_pi_samples > 0:
-            _z = einops.repeat(z, "d -> n d", n=self.config.n_pi_samples)
+            _z = einops.repeat(z, "b d -> n b d", n=self.config.n_pi_samples)
             for t in range(self.config.horizon):
                 # Note: Adding a small amount of noise here doesn't hurt during inference and may even be
                 # helpful for CEM.
@@ -202,12 +203,14 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
         # trajectories.
-        z = einops.repeat(z, "d -> n d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+        z = einops.repeat(z, "b d -> n b d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
 
         # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
         # algorithm.
         # The initial mean and standard deviation for the cross-entropy method (CEM).
-        mean = torch.zeros(self.config.horizon, self.config.output_shapes["action"][0], device=device)
+        mean = torch.zeros(
+            self.config.horizon, batch_size, self.config.output_shapes["action"][0], device=device
+        )
         # Maybe warm start CEM with the mean from the previous step.
         if self._prev_mean is not None:
             mean[:-1] = self._prev_mean[1:]
@@ -218,6 +221,7 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             std_normal_noise = torch.randn(
                 self.config.horizon,
                 self.config.n_gaussian_samples,
+                batch_size,
                 self.config.output_shapes["action"][0],
                 device=std.device,
             )
@@ -226,21 +230,24 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             # Compute elite actions.
             actions = torch.cat([gaussian_actions, pi_actions], dim=1)
             value = self.estimate_value(z, actions).nan_to_num_(0)
-            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices
-            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
+            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices  # (n_elites, batch)
+            elite_value = value.take_along_dim(elite_idxs, dim=0)  # (n_elites, batch)
+            # (horizon, n_elites, batch, action_dim)
+            elite_actions = actions.take_along_dim(einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1)
 
-            # Update guassian PDF parameters to be the (weighted) mean and standard deviation of the elites.
-            max_value = elite_value.max(0)[0]
+            # Update gaussian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0, keepdim=True)[0]  # (1, batch)
             # The weighting is a softmax over trajectory values. Note that this is not the same as the usage
             # of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
             # makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
             score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
-            score /= score.sum()
-            _mean = torch.sum(einops.rearrange(score, "n -> n 1") * elite_actions, dim=1)
+            score /= score.sum(axis=0, keepdim=True)
+            # (horizon, batch, action_dim)
+            _mean = torch.sum(einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1)
             _std = torch.sqrt(
                 torch.sum(
-                    einops.rearrange(score, "n -> n 1")
-                    * (elite_actions - einops.rearrange(_mean, "h d -> h 1 d")) ** 2,
+                    einops.rearrange(score, "n b -> n b 1")
+                    * (elite_actions - einops.rearrange(_mean, "h b d -> h 1 b d")) ** 2,
                     dim=1,
                 )
             )
@@ -255,11 +262,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         # Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
         # scores from the last iteration.
-        actions = elite_actions[:, torch.multinomial(score, 1).item()]
+        actions = elite_actions[:, torch.multinomial(score.T, 1).squeeze(), torch.arange(batch_size)]
 
-        # Select only the first action
-        action = actions[0]
-        return action
+        return actions
 
     @torch.no_grad()
     def estimate_value(self, z: Tensor, actions: Tensor):
@@ -311,12 +316,17 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
         return G
 
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """Run the batch through the model and compute the loss."""
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
+        """Run the batch through the model and compute the loss.
+
+        Returns a dictionary with loss as a tensor, and other information as native floats.
+        """
         device = get_device_from_parameters(self)
 
         batch = self.normalize_inputs(batch)
-        batch["observation.image"] = batch[self.input_image_key]
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
         batch = self.normalize_targets(batch)
 
         info = {}
@@ -326,12 +336,12 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             if batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
-        action = batch["action"]  # (t, b)
-        reward = batch["next.reward"]  # (t,)
+        action = batch["action"]  # (t, b, action_dim)
+        reward = batch["next.reward"]  # (t, b)
         observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
 
         # Apply random image augmentations.
-        if self.config.max_random_shift_ratio > 0:
+        if self._use_image and self.config.max_random_shift_ratio > 0:
             observations["observation.image"] = flatten_forward_unflatten(
                 partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
                 observations["observation.image"],
@@ -343,7 +353,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         for k in observations:
             current_observation[k] = observations[k][0]
             next_observations[k] = observations[k][1:]
-        horizon = next_observations["observation.image"].shape[0]
+        horizon, batch_size = next_observations[
+            "observation.image" if self._use_image else "observation.environment_state"
+        ].shape[:2]
 
         # Run latent rollout using the latent dynamics model and policy model.
         # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
@@ -413,7 +425,8 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
         q_value_loss = (
             (
-                F.mse_loss(
+                temporal_loss_coeffs
+                * F.mse_loss(
                     q_preds_ensemble,
                     einops.repeat(q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]),
                     reduction="none",
@@ -462,10 +475,11 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         action_preds = self.model.pi(z_preds[:-1])  # (t, b, a)
         # Calculate the MSE between the actions and the action predictions.
         # Note: FOWM's original code calculates the log probability (wrt to a unit standard deviation
-        # gaussian) and sums over the action dimension. Computing the log probability amounts to multiplying
-        # the MSE by 0.5 and adding a constant offset (the log(2*pi) term) . Here we drop the constant offset
-        # as it doesn't change the optimization step, and we drop the 0.5 as we instead make a configuration
-        # parameter for it (see below where we compute the total loss).
+        # gaussian) and sums over the action dimension. Computing the (negative) log probability amounts to
+        # multiplying the MSE by 0.5 and adding a constant offset (the log(2*pi)/2 term, times the action
+        # dimension). Here we drop the constant offset as it doesn't change the optimization step, and we drop
+        # the 0.5 as we instead make a configuration parameter for it (see below where we compute the total
+        # loss).
         mse = F.mse_loss(action_preds, action, reduction="none").sum(-1)  # (t, b)
         # NOTE: The original implementation does not take the sum over the temporal dimension like with the
         # other losses.
@@ -726,6 +740,16 @@ class TDMPCObservationEncoder(nn.Module):
                 nn.LayerNorm(config.latent_dim),
                 nn.Sigmoid(),
             )
+        if "observation.environment_state" in config.input_shapes:
+            self.env_state_enc_layers = nn.Sequential(
+                nn.Linear(
+                    config.input_shapes["observation.environment_state"][0], config.state_encoder_hidden_dim
+                ),
+                nn.ELU(),
+                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
+                nn.LayerNorm(config.latent_dim),
+                nn.Sigmoid(),
+            )
 
     def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
         """Encode the image and/or state vector.
@@ -734,8 +758,11 @@ class TDMPCObservationEncoder(nn.Module):
         over all features.
         """
         feat = []
+        # NOTE: Order of observations matters here.
         if "observation.image" in self.config.input_shapes:
             feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.environment_state" in self.config.input_shapes:
+            feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
         if "observation.state" in self.config.input_shapes:
             feat.append(self.state_enc_layers(obs_dict["observation.state"]))
         return torch.stack(feat, dim=0).mean(0)

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -98,6 +98,7 @@ class VQBeTPolicy(nn.Module, PyTorchModelHubMixin):
         """
 
         batch = self.normalize_inputs(batch)
+        batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
@@ -123,6 +124,7 @@ class VQBeTPolicy(nn.Module, PyTorchModelHubMixin):
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
+        batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://arxiv.org/pdf/2403.03181)

lerobot/configs/default.yaml

@@ -32,19 +32,54 @@ video_backend: pyav
 
 training:
   offline_steps: ???
-  # NOTE: `online_steps` is not implemented yet. It's here as a placeholder.
-  online_steps: ???
-  online_steps_between_rollouts: ???
-  online_sampling_ratio: 0.5
-  # `online_env_seed` is used for environments for online training data rollouts.
-  online_env_seed: ???
+
+  # Number of workers for the offline training dataloader.
+  num_workers: 4
+
+  batch_size: ???
+
   eval_freq: ???
   log_freq: 200
   save_checkpoint: true
   # Checkpoint is saved every `save_freq` training iterations and after the last training step.
   save_freq: ???
-  num_workers: 4
-  batch_size: ???
+
+  # Online training. Note that the online training loop adopts most of the options above apart from the
+  # dataloader options. Unless otherwise specified.
+  # The online training look looks something like:
+  #
+  # for i in range(online_steps):
+  #     do_online_rollout_and_update_online_buffer()
+  #     for j in range(online_steps_between_rollouts):
+  #         batch = next(dataloader_with_offline_and_online_data)
+  #         loss = policy(batch)
+  #         loss.backward()
+  #         optimizer.step()
+  #
+  online_steps: ???
+  # How many episodes to collect at once when we reach the online rollout part of the training loop.
+  online_rollout_n_episodes: 1
+  # The number of environments to use in the gym.vector.VectorEnv. This ends up also being the batch size for
+  # the policy. Ideally you should set this to by an even divisor or online_rollout_n_episodes.
+  online_rollout_batch_size: 1
+  # How many optimization steps (forward, backward, optimizer step) to do between running rollouts.
+  online_steps_between_rollouts: null
+  # The proportion of online samples (vs offline samples) to include in the online training batches.
+  online_sampling_ratio: 0.5
+  # First seed to use for the online rollout environment. Seeds for subsequent rollouts are incremented by 1.
+  online_env_seed: null
+  # Sets the maximum number of frames that are stored in the online buffer for online training. The buffer is
+  # FIFO.
+  online_buffer_capacity: null
+  # The minimum number of frames to have in the online buffer before commencing online training.
+  # If online_buffer_seed_size > online_rollout_n_episodes, the rollout will be run multiple times until the
+  # seed size condition is satisfied.
+  online_buffer_seed_size: 0
+  # Whether to run the online rollouts asynchronously. This means we can run the online training steps in
+  # parallel with the rollouts. This might be advised if your GPU has the bandwidth to handle training
+  # + eval + environment rendering simultaneously.
+  do_online_rollout_async: false
+
   image_transforms:
   # These transforms are all using standard torchvision.transforms.v2
   # You can find out how these transformations affect images here:

lerobot/configs/env/xarm.yaml

@@ -9,7 +9,7 @@ env:
   state_dim: 4
   action_dim: 4
   fps: ${fps}
-  episode_length: 25
+  episode_length: 200
   gym:
     obs_type: pixels_agent_pos
     render_mode: rgb_array

lerobot/configs/policy/tdmpc.yaml

@@ -4,19 +4,30 @@ seed: 1
 dataset_repo_id: lerobot/xarm_lift_medium
 
 training:
-  offline_steps: 25000
-  # TODO(alexander-soare): uncomment when online training gets reinstated
-  online_steps: 0  # 25000 not implemented yet
-  eval_freq: 5000
-  online_steps_between_rollouts: 1
-  online_sampling_ratio: 0.5
-  online_env_seed: 10000
-  log_freq: 100
+  offline_steps: 50000
+
+  num_workers: 4
 
   batch_size: 256
   grad_clip_norm: 10.0
   lr: 3e-4
 
+  eval_freq: 5000
+  log_freq: 100
+
+  online_steps: 50000
+  online_rollout_n_episodes: 1
+  online_rollout_batch_size: 1
+  # Note: in FOWM `online_steps_between_rollouts` is actually dynamically set to match exactly the length of
+  # the last sampled episode.
+  online_steps_between_rollouts: 50
+  online_sampling_ratio: 0.5
+  online_env_seed: 10000
+  # FOWM Push uses 10000 for `online_buffer_capacity`. Given that their maximum episode length for this task
+  # is 25, 10000 is approx 400 of their episodes worth. Since our episodes are about 8 times longer, we'll use
+  # 80000.
+  online_buffer_capacity: 80000
+
   delta_timestamps:
     observation.image: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
     observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
@@ -31,6 +42,7 @@ policy:
   # Input / output structure.
   n_action_repeats: 2
   horizon: 5
+  n_action_steps: 1
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?

lerobot/configs/policy/tdmpc_pusht_keypoints.yaml

@@ -0,0 +1,105 @@
+# @package _global_
+
+# Train with:
+#
+# python lerobot/scripts/train.py \
+#   env=pusht \
+#   env.gym.obs_type=environment_state_agent_pos \
+#   policy=tdmpc_pusht_keypoints \
+#   eval.batch_size=50 \
+#   eval.n_episodes=50 \
+#   eval.use_async_envs=true \
+#   device=cuda \
+#   use_amp=true
+
+seed: 1
+dataset_repo_id: lerobot/pusht_keypoints
+
+training:
+  offline_steps: 0
+
+  # Offline training dataloader
+  num_workers: 4
+
+  batch_size: 256
+  grad_clip_norm: 10.0
+  lr: 3e-4
+
+  eval_freq: 10000
+  log_freq: 500
+  save_freq: 50000
+
+  online_steps: 1000000
+  online_rollout_n_episodes: 10
+  online_rollout_batch_size: 10
+  online_steps_between_rollouts: 1000
+  online_sampling_ratio: 1.0
+  online_env_seed: 10000
+  online_buffer_capacity: 40000
+  online_buffer_seed_size: 0
+  do_online_rollout_async: false
+
+  delta_timestamps:
+    observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    action: "[i / ${fps} for i in range(${policy.horizon})]"
+    next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
+
+policy:
+  name: tdmpc
+
+  pretrained_model_path:
+
+  # Input / output structure.
+  n_action_repeats: 1
+  horizon: 5
+  n_action_steps: 5
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.environment_state: [16]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.environment_state: min_max
+    observation.state: min_max
+  output_normalization_modes:
+    action: min_max
+
+  # Architecture / modeling.
+  # Neural networks.
+  image_encoder_hidden_dim: 32
+  state_encoder_hidden_dim: 256
+  latent_dim: 50
+  q_ensemble_size: 5
+  mlp_dim: 512
+  # Reinforcement learning.
+  discount: 0.98
+
+  # Inference.
+  use_mpc: true
+  cem_iterations: 6
+  max_std: 2.0
+  min_std: 0.05
+  n_gaussian_samples: 512
+  n_pi_samples: 51
+  uncertainty_regularizer_coeff: 1.0
+  n_elites: 50
+  elite_weighting_temperature: 0.5
+  gaussian_mean_momentum: 0.1
+
+  # Training and loss computation.
+  max_random_shift_ratio: 0.0476
+  # Loss coefficients.
+  reward_coeff: 0.5
+  expectile_weight: 0.9
+  value_coeff: 0.1
+  consistency_coeff: 20.0
+  advantage_scaling: 3.0
+  pi_coeff: 0.5
+  temporal_decay_coeff: 0.5
+  # Target model.
+  target_model_momentum: 0.995

lerobot/scripts/control_robot.py

@@ -112,7 +112,7 @@ from termcolor import colored
 from lerobot.common.datasets.compute_stats import compute_stats
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import to_hf_dataset
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, get_default_encoding
 from lerobot.common.datasets.utils import calculate_episode_data_index
 from lerobot.common.datasets.video_utils import encode_video_frames
 from lerobot.common.policies.factory import make_policy
@@ -557,6 +557,8 @@ def record(
         "fps": fps,
         "video": video,
     }
+    if video:
+        info["encoding"] = get_default_encoding()
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
         repo_id=repo_id,

lerobot/scripts/eval.py

@@ -56,16 +56,13 @@ import einops
 import gymnasium as gym
 import numpy as np
 import torch
-from datasets import Dataset, Features, Image, Sequence, Value, concatenate_datasets
 from huggingface_hub import snapshot_download
 from huggingface_hub.utils._errors import RepositoryNotFoundError
 from huggingface_hub.utils._validators import HFValidationError
-from PIL import Image as PILImage
 from torch import Tensor, nn
 from tqdm import trange
 
 from lerobot.common.datasets.factory import make_dataset
-from lerobot.common.datasets.utils import hf_transform_to_torch
 from lerobot.common.envs.factory import make_env
 from lerobot.common.envs.utils import preprocess_observation
 from lerobot.common.logger import log_output_dir
@@ -318,41 +315,17 @@ def eval_policy(
                 rollout_data,
                 done_indices,
                 start_episode_index=batch_ix * env.num_envs,
-                start_data_index=(
-                    0 if episode_data is None else (episode_data["episode_data_index"]["to"][-1].item())
-                ),
+                start_data_index=(0 if episode_data is None else (episode_data["index"][-1].item() + 1)),
                 fps=env.unwrapped.metadata["render_fps"],
             )
             if episode_data is None:
                 episode_data = this_episode_data
             else:
-                # Some sanity checks to make sure we are not correctly compiling the data.
-                assert (
-                    episode_data["hf_dataset"]["episode_index"][-1] + 1
-                    == this_episode_data["hf_dataset"]["episode_index"][0]
-                )
-                assert (
-                    episode_data["hf_dataset"]["index"][-1] + 1 == this_episode_data["hf_dataset"]["index"][0]
-                )
-                assert torch.equal(
-                    episode_data["episode_data_index"]["to"][-1],
-                    this_episode_data["episode_data_index"]["from"][0],
-                )
+                # Some sanity checks to make sure we are correctly compiling the data.
+                assert episode_data["episode_index"][-1] + 1 == this_episode_data["episode_index"][0]
+                assert episode_data["index"][-1] + 1 == this_episode_data["index"][0]
                 # Concatenate the episode data.
-                episode_data = {
-                    "hf_dataset": concatenate_datasets(
-                        [episode_data["hf_dataset"], this_episode_data["hf_dataset"]]
-                    ),
-                    "episode_data_index": {
-                        k: torch.cat(
-                            [
-                                episode_data["episode_data_index"][k],
-                                this_episode_data["episode_data_index"][k],
-                            ]
-                        )
-                        for k in ["from", "to"]
-                    },
-                }
+                episode_data = {k: torch.cat([episode_data[k], this_episode_data[k]]) for k in episode_data}
 
         # Maybe render video for visualization.
         if max_episodes_rendered > 0 and len(ep_frames) > 0:
@@ -434,89 +407,39 @@ def _compile_episode_data(
     Similar logic is implemented when datasets are pushed to hub (see: `push_to_hub`).
     """
     ep_dicts = []
-    episode_data_index = {"from": [], "to": []}
     total_frames = 0
-    data_index_from = start_data_index
     for ep_ix in range(rollout_data["action"].shape[0]):
-        num_frames = done_indices[ep_ix].item() + 1  # + 1 to include the first done frame
+        # + 2 to include the first done frame and the last observation frame.
+        num_frames = done_indices[ep_ix].item() + 2
         total_frames += num_frames
 
-        # TODO(rcadene): We need to add a missing last frame which is the observation
-        # of a done state. it is critical to have this frame for tdmpc to predict a "done observation/state"
+        # Here we do `num_frames - 1` as we don't want to include the last observation frame just yet.
         ep_dict = {
-            "action": rollout_data["action"][ep_ix, :num_frames],
-            "episode_index": torch.tensor([start_episode_index + ep_ix] * num_frames),
-            "frame_index": torch.arange(0, num_frames, 1),
-            "timestamp": torch.arange(0, num_frames, 1) / fps,
-            "next.done": rollout_data["done"][ep_ix, :num_frames],
-            "next.reward": rollout_data["reward"][ep_ix, :num_frames].type(torch.float32),
+            "action": rollout_data["action"][ep_ix, : num_frames - 1],
+            "episode_index": torch.tensor([start_episode_index + ep_ix] * (num_frames - 1)),
+            "frame_index": torch.arange(0, num_frames - 1, 1),
+            "timestamp": torch.arange(0, num_frames - 1, 1) / fps,
+            "next.done": rollout_data["done"][ep_ix, : num_frames - 1],
+            "next.success": rollout_data["success"][ep_ix, : num_frames - 1],
+            "next.reward": rollout_data["reward"][ep_ix, : num_frames - 1].type(torch.float32),
         }
+
+        # For the last observation frame, all other keys will just be copy padded.
+        for k in ep_dict:
+            ep_dict[k] = torch.cat([ep_dict[k], ep_dict[k][-1:]])
+
         for key in rollout_data["observation"]:
-            ep_dict[key] = rollout_data["observation"][key][ep_ix][:num_frames]
+            ep_dict[key] = rollout_data["observation"][key][ep_ix, :num_frames]
+
         ep_dicts.append(ep_dict)
 
-        episode_data_index["from"].append(data_index_from)
-        episode_data_index["to"].append(data_index_from + num_frames)
-
-        data_index_from += num_frames
-
     data_dict = {}
     for key in ep_dicts[0]:
-        if "image" not in key:
-            data_dict[key] = torch.cat([x[key] for x in ep_dicts])
-        else:
-            if key not in data_dict:
-                data_dict[key] = []
-            for ep_dict in ep_dicts:
-                for img in ep_dict[key]:
-                    # sanity check that images are channel first
-                    c, h, w = img.shape
-                    assert c < h and c < w, f"expect channel first images, but instead {img.shape}"
-
-                    # sanity check that images are float32 in range [0,1]
-                    assert img.dtype == torch.float32, f"expect torch.float32, but instead {img.dtype=}"
-                    assert img.max() <= 1, f"expect pixels lower than 1, but instead {img.max()=}"
-                    assert img.min() >= 0, f"expect pixels greater than 1, but instead {img.min()=}"
-
-                    # from float32 in range [0,1] to uint8 in range [0,255]
-                    img *= 255
-                    img = img.type(torch.uint8)
-
-                    # convert to channel last and numpy as expected by PIL
-                    img = PILImage.fromarray(img.permute(1, 2, 0).numpy())
-
-                    data_dict[key].append(img)
+        data_dict[key] = torch.cat([x[key] for x in ep_dicts])
 
     data_dict["index"] = torch.arange(start_data_index, start_data_index + total_frames, 1)
-    episode_data_index["from"] = torch.tensor(episode_data_index["from"])
-    episode_data_index["to"] = torch.tensor(episode_data_index["to"])
 
-    # TODO(rcadene): clean this
-    features = {}
-    for key in rollout_data["observation"]:
-        if "image" in key:
-            features[key] = Image()
-        else:
-            features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
-    features.update(
-        {
-            "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-            "episode_index": Value(dtype="int64", id=None),
-            "frame_index": Value(dtype="int64", id=None),
-            "timestamp": Value(dtype="float32", id=None),
-            "next.reward": Value(dtype="float32", id=None),
-            "next.done": Value(dtype="bool", id=None),
-            #'next.success': Value(dtype='bool', id=None),
-            "index": Value(dtype="int64", id=None),
-        }
-    )
-    features = Features(features)
-    hf_dataset = Dataset.from_dict(data_dict, features=features)
-    hf_dataset.set_transform(hf_transform_to_torch)
-    return {
-        "hf_dataset": hf_dataset,
-        "episode_data_index": episode_data_index,
-    }
+    return data_dict
 
 
 def main(

lerobot/scripts/push_dataset_to_hub.py

@@ -55,6 +55,7 @@ from safetensors.torch import save_file
 
 from lerobot.common.datasets.compute_stats import compute_stats
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
 from lerobot.common.datasets.utils import flatten_dict
 
 
@@ -140,14 +141,12 @@ def push_dataset_to_hub(
     num_workers: int = 8,
     episodes: list[int] | None = None,
     force_override: bool = False,
+    resume: bool = False,
     cache_dir: Path = Path("/tmp"),
     tests_data_dir: Path | None = None,
+    encoding: dict | None = None,
 ):
-    # Check repo_id is well formated
-    if len(repo_id.split("/")) != 2:
-        raise ValueError(
-            f"`repo_id` is expected to contain a community or user id `/` the name of the dataset (e.g. 'lerobot/pusht'), but instead contains '{repo_id}'."
-        )
+    check_repo_id(repo_id)
     user_id, dataset_id = repo_id.split("/")
 
     # Robustify when `raw_dir` is str instead of Path
@@ -173,7 +172,7 @@ def push_dataset_to_hub(
         if local_dir.exists():
             if force_override:
                 shutil.rmtree(local_dir)
-            else:
+            elif not resume:
                 raise ValueError(f"`local_dir` already exists ({local_dir}). Use `--force-override 1`.")
 
         meta_data_dir = local_dir / "meta_data"
@@ -191,7 +190,7 @@ def push_dataset_to_hub(
     # convert dataset from original raw format to LeRobot format
     from_raw_to_lerobot_format = get_from_raw_to_lerobot_format_fn(raw_format)
     hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(
-        raw_dir, videos_dir, fps, video, episodes
+        raw_dir, videos_dir, fps, video, episodes, encoding
     )
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
@@ -315,6 +314,12 @@ def main():
         default=0,
         help="When set to 1, removes provided output directory if it already exists. By default, raises a ValueError exception.",
     )
+    parser.add_argument(
+        "--resume",
+        type=int,
+        default=0,
+        help="When set to 1, resumes a previous run.",
+    )
     parser.add_argument(
         "--tests-data-dir",
         type=Path,

lerobot/scripts/train.py

@@ -15,20 +15,25 @@
 # limitations under the License.
 import logging
 import time
+from concurrent.futures import ThreadPoolExecutor
 from contextlib import nullcontext
+from copy import deepcopy
 from pathlib import Path
 from pprint import pformat
+from threading import Lock
 
 import hydra
+import numpy as np
 import torch
 from deepdiff import DeepDiff
-from omegaconf import DictConfig, OmegaConf
+from omegaconf import DictConfig, ListConfig, OmegaConf
 from termcolor import colored
 from torch import nn
 from torch.cuda.amp import GradScaler
 
 from lerobot.common.datasets.factory import make_dataset, resolve_delta_timestamps
 from lerobot.common.datasets.lerobot_dataset import MultiLeRobotDataset
+from lerobot.common.datasets.online_buffer import OnlineBuffer, compute_sampler_weights
 from lerobot.common.datasets.sampler import EpisodeAwareSampler
 from lerobot.common.datasets.utils import cycle
 from lerobot.common.envs.factory import make_env
@@ -107,6 +112,7 @@ def update_policy(
     grad_scaler: GradScaler,
     lr_scheduler=None,
     use_amp: bool = False,
+    lock=None,
 ):
     """Returns a dictionary of items for logging."""
     start_time = time.perf_counter()
@@ -129,7 +135,8 @@ def update_policy(
 
     # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
     # although it still skips optimizer.step() if the gradients contain infs or NaNs.
-    grad_scaler.step(optimizer)
+    with lock if lock is not None else nullcontext():
+        grad_scaler.step(optimizer)
     # Updates the scale for next iteration.
     grad_scaler.update()
 
@@ -149,11 +156,12 @@ def update_policy(
         "update_s": time.perf_counter() - start_time,
         **{k: v for k, v in output_dict.items() if k != "loss"},
     }
+    info.update({k: v for k, v in output_dict.items() if k not in info})
 
     return info
 
 
-def log_train_info(logger: Logger, info, step, cfg, dataset, is_offline):
+def log_train_info(logger: Logger, info, step, cfg, dataset, is_online):
     loss = info["loss"]
     grad_norm = info["grad_norm"]
     lr = info["lr"]
@@ -187,12 +195,12 @@ def log_train_info(logger: Logger, info, step, cfg, dataset, is_offline):
     info["num_samples"] = num_samples
     info["num_episodes"] = num_episodes
     info["num_epochs"] = num_epochs
-    info["is_offline"] = is_offline
+    info["is_online"] = is_online
 
     logger.log_dict(info, step, mode="train")
 
 
-def log_eval_info(logger, info, step, cfg, dataset, is_offline):
+def log_eval_info(logger, info, step, cfg, dataset, is_online):
     eval_s = info["eval_s"]
     avg_sum_reward = info["avg_sum_reward"]
     pc_success = info["pc_success"]
@@ -221,7 +229,7 @@ def log_eval_info(logger, info, step, cfg, dataset, is_offline):
     info["num_samples"] = num_samples
     info["num_episodes"] = num_episodes
     info["num_epochs"] = num_epochs
-    info["is_offline"] = is_offline
+    info["is_online"] = is_online
 
     logger.log_dict(info, step, mode="eval")
 
@@ -234,6 +242,9 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     init_logging()
 
+    if cfg.training.online_steps > 0 and isinstance(cfg.dataset_repo_id, ListConfig):
+        raise NotImplementedError("Online training with LeRobotMultiDataset is not implemented.")
+
     # If we are resuming a run, we need to check that a checkpoint exists in the log directory, and we need
     # to check for any differences between the provided config and the checkpoint's config.
     if cfg.resume:
@@ -279,9 +290,6 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     # log metrics to terminal and wandb
     logger = Logger(cfg, out_dir, wandb_job_name=job_name)
 
-    if cfg.training.online_steps > 0:
-        raise NotImplementedError("Online training is not implemented yet.")
-
     set_global_seed(cfg.seed)
 
     # Check device is available
@@ -336,7 +344,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     logging.info(f"{num_total_params=} ({format_big_number(num_total_params)})")
 
     # Note: this helper will be used in offline and online training loops.
-    def evaluate_and_checkpoint_if_needed(step):
+    def evaluate_and_checkpoint_if_needed(step, is_online):
         _num_digits = max(6, len(str(cfg.training.offline_steps + cfg.training.online_steps)))
         step_identifier = f"{step:0{_num_digits}d}"
 
@@ -352,7 +360,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
                     max_episodes_rendered=4,
                     start_seed=cfg.seed,
                 )
-            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline=True)
+            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_online=is_online)
             if cfg.wandb.enable:
                 logger.log_video(eval_info["video_paths"][0], step, mode="eval")
             logging.info("Resume training")
@@ -396,8 +404,9 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     dl_iter = cycle(dataloader)
 
     policy.train()
+    offline_step = 0
     for _ in range(step, cfg.training.offline_steps):
-        if step == 0:
+        if offline_step == 0:
             logging.info("Start offline training on a fixed dataset")
 
         start_time = time.perf_counter()
@@ -420,13 +429,207 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         train_info["dataloading_s"] = dataloading_s
 
         if step % cfg.training.log_freq == 0:
-            log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline=True)
+            log_train_info(logger, train_info, step, cfg, offline_dataset, is_online=False)
 
         # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
         # so we pass in step + 1.
-        evaluate_and_checkpoint_if_needed(step + 1)
+        evaluate_and_checkpoint_if_needed(step + 1, is_online=False)
 
         step += 1
+        offline_step += 1  # noqa: SIM113
+
+    if cfg.training.online_steps == 0:
+        if eval_env:
+            eval_env.close()
+        logging.info("End of training")
+        return
+
+    # Online training.
+
+    # Create an env dedicated to online episodes collection from policy rollout.
+    online_env = make_env(cfg, n_envs=cfg.training.online_rollout_batch_size)
+    resolve_delta_timestamps(cfg)
+    online_buffer_path = logger.log_dir / "online_buffer"
+    if cfg.resume and not online_buffer_path.exists():
+        # If we are resuming a run, we default to the data shapes and buffer capacity from the saved online
+        # buffer.
+        logging.warning(
+            "When online training is resumed, we load the latest online buffer from the prior run, "
+            "and this might not coincide with the state of the buffer as it was at the moment the checkpoint "
+            "was made. This is because the online buffer is updated on disk during training, independently "
+            "of our explicit checkpointing mechanisms."
+        )
+    online_dataset = OnlineBuffer(
+        online_buffer_path,
+        data_spec={
+            **{k: {"shape": v, "dtype": np.dtype("float32")} for k, v in policy.config.input_shapes.items()},
+            **{k: {"shape": v, "dtype": np.dtype("float32")} for k, v in policy.config.output_shapes.items()},
+            "next.reward": {"shape": (), "dtype": np.dtype("float32")},
+            "next.done": {"shape": (), "dtype": np.dtype("?")},
+            "next.success": {"shape": (), "dtype": np.dtype("?")},
+        },
+        buffer_capacity=cfg.training.online_buffer_capacity,
+        fps=online_env.unwrapped.metadata["render_fps"],
+        delta_timestamps=cfg.training.delta_timestamps,
+    )
+
+    # If we are doing online rollouts asynchronously, deepcopy the policy to use for online rollouts (this
+    # makes it possible to do online rollouts in parallel with training updates).
+    online_rollout_policy = deepcopy(policy) if cfg.training.do_online_rollout_async else policy
+
+    # Create dataloader for online training.
+    concat_dataset = torch.utils.data.ConcatDataset([offline_dataset, online_dataset])
+    sampler_weights = compute_sampler_weights(
+        offline_dataset,
+        offline_drop_n_last_frames=cfg.training.get("drop_n_last_frames", 0),
+        online_dataset=online_dataset,
+        # +1 because online rollouts return an extra frame for the "final observation". Note: we don't have
+        # this final observation in the offline datasets, but we might add them in future.
+        online_drop_n_last_frames=cfg.training.get("drop_n_last_frames", 0) + 1,
+        online_sampling_ratio=cfg.training.online_sampling_ratio,
+    )
+    sampler = torch.utils.data.WeightedRandomSampler(
+        sampler_weights,
+        num_samples=len(concat_dataset),
+        replacement=True,
+    )
+    dataloader = torch.utils.data.DataLoader(
+        concat_dataset,
+        batch_size=cfg.training.batch_size,
+        num_workers=cfg.training.num_workers,
+        sampler=sampler,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
+    dl_iter = cycle(dataloader)
+
+    # Lock and thread pool executor for asynchronous online rollouts. When asynchronous mode is disabled,
+    # these are still used but effectively do nothing.
+    lock = Lock()
+    # Note: 1 worker because we only ever want to run one set of online rollouts at a time. Batch
+    # parallelization of rollouts is handled within the job.
+    executor = ThreadPoolExecutor(max_workers=1)
+
+    online_step = 0
+    online_rollout_s = 0  # time take to do online rollout
+    update_online_buffer_s = 0  # time taken to update the online buffer with the online rollout data
+    # Time taken waiting for the online buffer to finish being updated. This is relevant when using the async
+    # online rollout option.
+    await_update_online_buffer_s = 0
+    rollout_start_seed = cfg.training.online_env_seed
+
+    while True:
+        if online_step == cfg.training.online_steps:
+            break
+
+        if online_step == 0:
+            logging.info("Start online training by interacting with environment")
+
+        def sample_trajectory_and_update_buffer():
+            nonlocal rollout_start_seed
+            with lock:
+                online_rollout_policy.load_state_dict(policy.state_dict())
+            online_rollout_policy.eval()
+            start_rollout_time = time.perf_counter()
+            with torch.no_grad():
+                eval_info = eval_policy(
+                    online_env,
+                    online_rollout_policy,
+                    n_episodes=cfg.training.online_rollout_n_episodes,
+                    max_episodes_rendered=min(10, cfg.training.online_rollout_n_episodes),
+                    videos_dir=logger.log_dir / "online_rollout_videos",
+                    return_episode_data=True,
+                    start_seed=(
+                        rollout_start_seed := (rollout_start_seed + cfg.training.batch_size) % 1000000
+                    ),
+                )
+            online_rollout_s = time.perf_counter() - start_rollout_time
+
+            with lock:
+                start_update_buffer_time = time.perf_counter()
+                online_dataset.add_data(eval_info["episodes"])
+
+                # Update the concatenated dataset length used during sampling.
+                concat_dataset.cumulative_sizes = concat_dataset.cumsum(concat_dataset.datasets)
+
+                # Update the sampling weights.
+                sampler.weights = compute_sampler_weights(
+                    offline_dataset,
+                    offline_drop_n_last_frames=cfg.training.get("drop_n_last_frames", 0),
+                    online_dataset=online_dataset,
+                    # +1 because online rollouts return an extra frame for the "final observation". Note: we don't have
+                    # this final observation in the offline datasets, but we might add them in future.
+                    online_drop_n_last_frames=cfg.training.get("drop_n_last_frames", 0) + 1,
+                    online_sampling_ratio=cfg.training.online_sampling_ratio,
+                )
+                sampler.num_samples = len(concat_dataset)
+
+                update_online_buffer_s = time.perf_counter() - start_update_buffer_time
+
+            return online_rollout_s, update_online_buffer_s
+
+        future = executor.submit(sample_trajectory_and_update_buffer)
+        # If we aren't doing async rollouts, or if we haven't yet gotten enough examples in our buffer, wait
+        # here until the rollout and buffer update is done, before proceeding to the policy update steps.
+        if (
+            not cfg.training.do_online_rollout_async
+            or len(online_dataset) <= cfg.training.online_buffer_seed_size
+        ):
+            online_rollout_s, update_online_buffer_s = future.result()
+
+        if len(online_dataset) <= cfg.training.online_buffer_seed_size:
+            logging.info(
+                f"Seeding online buffer: {len(online_dataset)}/{cfg.training.online_buffer_seed_size}"
+            )
+            continue
+
+        policy.train()
+        for _ in range(cfg.training.online_steps_between_rollouts):
+            with lock:
+                start_time = time.perf_counter()
+                batch = next(dl_iter)
+                dataloading_s = time.perf_counter() - start_time
+
+            for key in batch:
+                batch[key] = batch[key].to(cfg.device, non_blocking=True)
+
+            train_info = update_policy(
+                policy,
+                batch,
+                optimizer,
+                cfg.training.grad_clip_norm,
+                grad_scaler=grad_scaler,
+                lr_scheduler=lr_scheduler,
+                use_amp=cfg.use_amp,
+                lock=lock,
+            )
+
+            train_info["dataloading_s"] = dataloading_s
+            train_info["online_rollout_s"] = online_rollout_s
+            train_info["update_online_buffer_s"] = update_online_buffer_s
+            train_info["await_update_online_buffer_s"] = await_update_online_buffer_s
+            with lock:
+                train_info["online_buffer_size"] = len(online_dataset)
+
+            if step % cfg.training.log_freq == 0:
+                log_train_info(logger, train_info, step, cfg, online_dataset, is_online=True)
+
+            # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
+            # so we pass in step + 1.
+            evaluate_and_checkpoint_if_needed(step + 1, is_online=True)
+
+            step += 1
+            online_step += 1
+
+        # If we're doing async rollouts, we should now wait until we've completed them before proceeding
+        # to do the next batch of rollouts.
+        if future.running():
+            start = time.perf_counter()
+            online_rollout_s, update_online_buffer_s = future.result()
+            await_update_online_buffer_s = time.perf_counter() - start
+
+        if online_step >= cfg.training.online_steps:
+            break
 
     if eval_env:
         eval_env.close()

lerobot/scripts/visualize_image_transforms.py

@@ -25,7 +25,7 @@ Increase hue jitter
 ```
 python lerobot/scripts/visualize_image_transforms.py \
     dataset_repo_id=lerobot/aloha_mobile_shrimp \
-    training.image_transforms.hue.min_max=[-0.25,0.25]
+    training.image_transforms.hue.min_max="[-0.25,0.25]"
 ```
 
 Increase brightness & brightness weight
@@ -33,7 +33,7 @@ Increase brightness & brightness weight
 python lerobot/scripts/visualize_image_transforms.py \
     dataset_repo_id=lerobot/aloha_mobile_shrimp \
     training.image_transforms.brightness.weight=10.0 \
-    training.image_transforms.brightness.min_max=[1.0,2.0]
+    training.image_transforms.brightness.min_max="[1.0,2.0]"
 ```
 
 Blur images and disable saturation & hue
@@ -41,7 +41,7 @@ Blur images and disable saturation & hue
 python lerobot/scripts/visualize_image_transforms.py \
     dataset_repo_id=lerobot/aloha_mobile_shrimp \
     training.image_transforms.sharpness.weight=10.0 \
-    training.image_transforms.sharpness.min_max=[0.0,1.0] \
+    training.image_transforms.sharpness.min_max="[0.0,1.0]" \
     training.image_transforms.saturation.weight=0.0 \
     training.image_transforms.hue.weight=0.0
 ```

poetry.lock

@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.8.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.8.3 and should not be changed by hand.
 
 [[package]]
 name = "absl-py"
@@ -444,63 +444,63 @@ files = [
 
 [[package]]
 name = "coverage"
-version = "7.5.4"
+version = "7.6.0"
 description = "Code coverage measurement for Python"
 optional = true
 python-versions = ">=3.8"
 files = [
-    {file = "coverage-7.5.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:6cfb5a4f556bb51aba274588200a46e4dd6b505fb1a5f8c5ae408222eb416f99"},
-    {file = "coverage-7.5.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2174e7c23e0a454ffe12267a10732c273243b4f2d50d07544a91198f05c48f47"},
-    {file = "coverage-7.5.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2214ee920787d85db1b6a0bd9da5f8503ccc8fcd5814d90796c2f2493a2f4d2e"},
-    {file = "coverage-7.5.4-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:1137f46adb28e3813dec8c01fefadcb8c614f33576f672962e323b5128d9a68d"},
-    {file = "coverage-7.5.4-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b385d49609f8e9efc885790a5a0e89f2e3ae042cdf12958b6034cc442de428d3"},
-    {file = "coverage-7.5.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:b4a474f799456e0eb46d78ab07303286a84a3140e9700b9e154cfebc8f527016"},
-    {file = "coverage-7.5.4-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5cd64adedf3be66f8ccee418473c2916492d53cbafbfcff851cbec5a8454b136"},
-    {file = "coverage-7.5.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:e564c2cf45d2f44a9da56f4e3a26b2236504a496eb4cb0ca7221cd4cc7a9aca9"},
-    {file = "coverage-7.5.4-cp310-cp310-win32.whl", hash = "sha256:7076b4b3a5f6d2b5d7f1185fde25b1e54eb66e647a1dfef0e2c2bfaf9b4c88c8"},
-    {file = "coverage-7.5.4-cp310-cp310-win_amd64.whl", hash = "sha256:018a12985185038a5b2bcafab04ab833a9a0f2c59995b3cec07e10074c78635f"},
-    {file = "coverage-7.5.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:db14f552ac38f10758ad14dd7b983dbab424e731588d300c7db25b6f89e335b5"},
-    {file = "coverage-7.5.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:3257fdd8e574805f27bb5342b77bc65578e98cbc004a92232106344053f319ba"},
-    {file = "coverage-7.5.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3a6612c99081d8d6134005b1354191e103ec9705d7ba2754e848211ac8cacc6b"},
-    {file = "coverage-7.5.4-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d45d3cbd94159c468b9b8c5a556e3f6b81a8d1af2a92b77320e887c3e7a5d080"},
-    {file = "coverage-7.5.4-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ed550e7442f278af76d9d65af48069f1fb84c9f745ae249c1a183c1e9d1b025c"},
-    {file = "coverage-7.5.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:7a892be37ca35eb5019ec85402c3371b0f7cda5ab5056023a7f13da0961e60da"},
-    {file = "coverage-7.5.4-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:8192794d120167e2a64721d88dbd688584675e86e15d0569599257566dec9bf0"},
-    {file = "coverage-7.5.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:820bc841faa502e727a48311948e0461132a9c8baa42f6b2b84a29ced24cc078"},
-    {file = "coverage-7.5.4-cp311-cp311-win32.whl", hash = "sha256:6aae5cce399a0f065da65c7bb1e8abd5c7a3043da9dceb429ebe1b289bc07806"},
-    {file = "coverage-7.5.4-cp311-cp311-win_amd64.whl", hash = "sha256:d2e344d6adc8ef81c5a233d3a57b3c7d5181f40e79e05e1c143da143ccb6377d"},
-    {file = "coverage-7.5.4-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:54317c2b806354cbb2dc7ac27e2b93f97096912cc16b18289c5d4e44fc663233"},
-    {file = "coverage-7.5.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:042183de01f8b6d531e10c197f7f0315a61e8d805ab29c5f7b51a01d62782747"},
-    {file = "coverage-7.5.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a6bb74ed465d5fb204b2ec41d79bcd28afccf817de721e8a807d5141c3426638"},
-    {file = "coverage-7.5.4-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b3d45ff86efb129c599a3b287ae2e44c1e281ae0f9a9bad0edc202179bcc3a2e"},
-    {file = "coverage-7.5.4-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5013ed890dc917cef2c9f765c4c6a8ae9df983cd60dbb635df8ed9f4ebc9f555"},
-    {file = "coverage-7.5.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:1014fbf665fef86cdfd6cb5b7371496ce35e4d2a00cda501cf9f5b9e6fced69f"},
-    {file = "coverage-7.5.4-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:3684bc2ff328f935981847082ba4fdc950d58906a40eafa93510d1b54c08a66c"},
-    {file = "coverage-7.5.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:581ea96f92bf71a5ec0974001f900db495488434a6928a2ca7f01eee20c23805"},
-    {file = "coverage-7.5.4-cp312-cp312-win32.whl", hash = "sha256:73ca8fbc5bc622e54627314c1a6f1dfdd8db69788f3443e752c215f29fa87a0b"},
-    {file = "coverage-7.5.4-cp312-cp312-win_amd64.whl", hash = "sha256:cef4649ec906ea7ea5e9e796e68b987f83fa9a718514fe147f538cfeda76d7a7"},
-    {file = "coverage-7.5.4-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:cdd31315fc20868c194130de9ee6bfd99755cc9565edff98ecc12585b90be882"},
-    {file = "coverage-7.5.4-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:02ff6e898197cc1e9fa375581382b72498eb2e6d5fc0b53f03e496cfee3fac6d"},
-    {file = "coverage-7.5.4-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d05c16cf4b4c2fc880cb12ba4c9b526e9e5d5bb1d81313d4d732a5b9fe2b9d53"},
-    {file = "coverage-7.5.4-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c5986ee7ea0795a4095ac4d113cbb3448601efca7f158ec7f7087a6c705304e4"},
-    {file = "coverage-7.5.4-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5df54843b88901fdc2f598ac06737f03d71168fd1175728054c8f5a2739ac3e4"},
-    {file = "coverage-7.5.4-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:ab73b35e8d109bffbda9a3e91c64e29fe26e03e49addf5b43d85fc426dde11f9"},
-    {file = "coverage-7.5.4-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:aea072a941b033813f5e4814541fc265a5c12ed9720daef11ca516aeacd3bd7f"},
-    {file = "coverage-7.5.4-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:16852febd96acd953b0d55fc842ce2dac1710f26729b31c80b940b9afcd9896f"},
-    {file = "coverage-7.5.4-cp38-cp38-win32.whl", hash = "sha256:8f894208794b164e6bd4bba61fc98bf6b06be4d390cf2daacfa6eca0a6d2bb4f"},
-    {file = "coverage-7.5.4-cp38-cp38-win_amd64.whl", hash = "sha256:e2afe743289273209c992075a5a4913e8d007d569a406ffed0bd080ea02b0633"},
-    {file = "coverage-7.5.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:b95c3a8cb0463ba9f77383d0fa8c9194cf91f64445a63fc26fb2327e1e1eb088"},
-    {file = "coverage-7.5.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:3d7564cc09dd91b5a6001754a5b3c6ecc4aba6323baf33a12bd751036c998be4"},
-    {file = "coverage-7.5.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:44da56a2589b684813f86d07597fdf8a9c6ce77f58976727329272f5a01f99f7"},
-    {file = "coverage-7.5.4-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e16f3d6b491c48c5ae726308e6ab1e18ee830b4cdd6913f2d7f77354b33f91c8"},
-    {file = "coverage-7.5.4-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:dbc5958cb471e5a5af41b0ddaea96a37e74ed289535e8deca404811f6cb0bc3d"},
-    {file = "coverage-7.5.4-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:a04e990a2a41740b02d6182b498ee9796cf60eefe40cf859b016650147908029"},
-    {file = "coverage-7.5.4-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:ddbd2f9713a79e8e7242d7c51f1929611e991d855f414ca9996c20e44a895f7c"},
-    {file = "coverage-7.5.4-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:b1ccf5e728ccf83acd313c89f07c22d70d6c375a9c6f339233dcf792094bcbf7"},
-    {file = "coverage-7.5.4-cp39-cp39-win32.whl", hash = "sha256:56b4eafa21c6c175b3ede004ca12c653a88b6f922494b023aeb1e836df953ace"},
-    {file = "coverage-7.5.4-cp39-cp39-win_amd64.whl", hash = "sha256:65e528e2e921ba8fd67d9055e6b9f9e34b21ebd6768ae1c1723f4ea6ace1234d"},
-    {file = "coverage-7.5.4-pp38.pp39.pp310-none-any.whl", hash = "sha256:79b356f3dd5b26f3ad23b35c75dbdaf1f9e2450b6bcefc6d0825ea0aa3f86ca5"},
-    {file = "coverage-7.5.4.tar.gz", hash = "sha256:a44963520b069e12789d0faea4e9fdb1e410cdc4aab89d94f7f55cbb7fef0353"},
+    {file = "coverage-7.6.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:dff044f661f59dace805eedb4a7404c573b6ff0cdba4a524141bc63d7be5c7fd"},
+    {file = "coverage-7.6.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a8659fd33ee9e6ca03950cfdcdf271d645cf681609153f218826dd9805ab585c"},
+    {file = "coverage-7.6.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7792f0ab20df8071d669d929c75c97fecfa6bcab82c10ee4adb91c7a54055463"},
+    {file = "coverage-7.6.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d4b3cd1ca7cd73d229487fa5caca9e4bc1f0bca96526b922d61053ea751fe791"},
+    {file = "coverage-7.6.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e7e128f85c0b419907d1f38e616c4f1e9f1d1b37a7949f44df9a73d5da5cd53c"},
+    {file = "coverage-7.6.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:a94925102c89247530ae1dab7dc02c690942566f22e189cbd53579b0693c0783"},
+    {file = "coverage-7.6.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:dcd070b5b585b50e6617e8972f3fbbee786afca71b1936ac06257f7e178f00f6"},
+    {file = "coverage-7.6.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:d50a252b23b9b4dfeefc1f663c568a221092cbaded20a05a11665d0dbec9b8fb"},
+    {file = "coverage-7.6.0-cp310-cp310-win32.whl", hash = "sha256:0e7b27d04131c46e6894f23a4ae186a6a2207209a05df5b6ad4caee6d54a222c"},
+    {file = "coverage-7.6.0-cp310-cp310-win_amd64.whl", hash = "sha256:54dece71673b3187c86226c3ca793c5f891f9fc3d8aa183f2e3653da18566169"},
+    {file = "coverage-7.6.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:c7b525ab52ce18c57ae232ba6f7010297a87ced82a2383b1afd238849c1ff933"},
+    {file = "coverage-7.6.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:4bea27c4269234e06f621f3fac3925f56ff34bc14521484b8f66a580aacc2e7d"},
+    {file = "coverage-7.6.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ed8d1d1821ba5fc88d4a4f45387b65de52382fa3ef1f0115a4f7a20cdfab0e94"},
+    {file = "coverage-7.6.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:01c322ef2bbe15057bc4bf132b525b7e3f7206f071799eb8aa6ad1940bcf5fb1"},
+    {file = "coverage-7.6.0-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:03cafe82c1b32b770a29fd6de923625ccac3185a54a5e66606da26d105f37dac"},
+    {file = "coverage-7.6.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:0d1b923fc4a40c5832be4f35a5dab0e5ff89cddf83bb4174499e02ea089daf57"},
+    {file = "coverage-7.6.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:4b03741e70fb811d1a9a1d75355cf391f274ed85847f4b78e35459899f57af4d"},
+    {file = "coverage-7.6.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:a73d18625f6a8a1cbb11eadc1d03929f9510f4131879288e3f7922097a429f63"},
+    {file = "coverage-7.6.0-cp311-cp311-win32.whl", hash = "sha256:65fa405b837060db569a61ec368b74688f429b32fa47a8929a7a2f9b47183713"},
+    {file = "coverage-7.6.0-cp311-cp311-win_amd64.whl", hash = "sha256:6379688fb4cfa921ae349c76eb1a9ab26b65f32b03d46bb0eed841fd4cb6afb1"},
+    {file = "coverage-7.6.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:f7db0b6ae1f96ae41afe626095149ecd1b212b424626175a6633c2999eaad45b"},
+    {file = "coverage-7.6.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:bbdf9a72403110a3bdae77948b8011f644571311c2fb35ee15f0f10a8fc082e8"},
+    {file = "coverage-7.6.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9cc44bf0315268e253bf563f3560e6c004efe38f76db03a1558274a6e04bf5d5"},
+    {file = "coverage-7.6.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:da8549d17489cd52f85a9829d0e1d91059359b3c54a26f28bec2c5d369524807"},
+    {file = "coverage-7.6.0-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0086cd4fc71b7d485ac93ca4239c8f75732c2ae3ba83f6be1c9be59d9e2c6382"},
+    {file = "coverage-7.6.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:1fad32ee9b27350687035cb5fdf9145bc9cf0a094a9577d43e909948ebcfa27b"},
+    {file = "coverage-7.6.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:044a0985a4f25b335882b0966625270a8d9db3d3409ddc49a4eb00b0ef5e8cee"},
+    {file = "coverage-7.6.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:76d5f82213aa78098b9b964ea89de4617e70e0d43e97900c2778a50856dac605"},
+    {file = "coverage-7.6.0-cp312-cp312-win32.whl", hash = "sha256:3c59105f8d58ce500f348c5b56163a4113a440dad6daa2294b5052a10db866da"},
+    {file = "coverage-7.6.0-cp312-cp312-win_amd64.whl", hash = "sha256:ca5d79cfdae420a1d52bf177de4bc2289c321d6c961ae321503b2ca59c17ae67"},
+    {file = "coverage-7.6.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:d39bd10f0ae453554798b125d2f39884290c480f56e8a02ba7a6ed552005243b"},
+    {file = "coverage-7.6.0-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:beb08e8508e53a568811016e59f3234d29c2583f6b6e28572f0954a6b4f7e03d"},
+    {file = "coverage-7.6.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b2e16f4cd2bc4d88ba30ca2d3bbf2f21f00f382cf4e1ce3b1ddc96c634bc48ca"},
+    {file = "coverage-7.6.0-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6616d1c9bf1e3faea78711ee42a8b972367d82ceae233ec0ac61cc7fec09fa6b"},
+    {file = "coverage-7.6.0-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ad4567d6c334c46046d1c4c20024de2a1c3abc626817ae21ae3da600f5779b44"},
+    {file = "coverage-7.6.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:d17c6a415d68cfe1091d3296ba5749d3d8696e42c37fca5d4860c5bf7b729f03"},
+    {file = "coverage-7.6.0-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:9146579352d7b5f6412735d0f203bbd8d00113a680b66565e205bc605ef81bc6"},
+    {file = "coverage-7.6.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:cdab02a0a941af190df8782aafc591ef3ad08824f97850b015c8c6a8b3877b0b"},
+    {file = "coverage-7.6.0-cp38-cp38-win32.whl", hash = "sha256:df423f351b162a702c053d5dddc0fc0ef9a9e27ea3f449781ace5f906b664428"},
+    {file = "coverage-7.6.0-cp38-cp38-win_amd64.whl", hash = "sha256:f2501d60d7497fd55e391f423f965bbe9e650e9ffc3c627d5f0ac516026000b8"},
+    {file = "coverage-7.6.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7221f9ac9dad9492cecab6f676b3eaf9185141539d5c9689d13fd6b0d7de840c"},
+    {file = "coverage-7.6.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:ddaaa91bfc4477d2871442bbf30a125e8fe6b05da8a0015507bfbf4718228ab2"},
+    {file = "coverage-7.6.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c4cbe651f3904e28f3a55d6f371203049034b4ddbce65a54527a3f189ca3b390"},
+    {file = "coverage-7.6.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:831b476d79408ab6ccfadaaf199906c833f02fdb32c9ab907b1d4aa0713cfa3b"},
+    {file = "coverage-7.6.0-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:46c3d091059ad0b9c59d1034de74a7f36dcfa7f6d3bde782c49deb42438f2450"},
+    {file = "coverage-7.6.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:4d5fae0a22dc86259dee66f2cc6c1d3e490c4a1214d7daa2a93d07491c5c04b6"},
+    {file = "coverage-7.6.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:07ed352205574aad067482e53dd606926afebcb5590653121063fbf4e2175166"},
+    {file = "coverage-7.6.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:49c76cdfa13015c4560702574bad67f0e15ca5a2872c6a125f6327ead2b731dd"},
+    {file = "coverage-7.6.0-cp39-cp39-win32.whl", hash = "sha256:482855914928c8175735a2a59c8dc5806cf7d8f032e4820d52e845d1f731dca2"},
+    {file = "coverage-7.6.0-cp39-cp39-win_amd64.whl", hash = "sha256:543ef9179bc55edfd895154a51792b01c017c87af0ebaae092720152e19e42ca"},
+    {file = "coverage-7.6.0-pp38.pp39.pp310-none-any.whl", hash = "sha256:6fe885135c8a479d3e37a7aae61cbd3a0fb2deccb4dda3c25f92a49189f766d6"},
+    {file = "coverage-7.6.0.tar.gz", hash = "sha256:289cc803fa1dc901f84701ac10c9ee873619320f2f9aff38794db4a4a0268d51"},
 ]
 
 [package.dependencies]
@@ -615,18 +615,18 @@ optimize = ["orjson"]
 
 [[package]]
 name = "diffusers"
-version = "0.27.2"
+version = "0.29.2"
 description = "State-of-the-art diffusion in PyTorch and JAX."
 optional = false
 python-versions = ">=3.8.0"
 files = [
-    {file = "diffusers-0.27.2-py3-none-any.whl", hash = "sha256:85da5cd1098ab428535d592136973ec0c3f12f78148c94b379cb9f02d2414e75"},
-    {file = "diffusers-0.27.2.tar.gz", hash = "sha256:6cefd7770d7fc1d139614233aa17cdcd639c138d0c3517b8d8bbc8cf573050a0"},
+    {file = "diffusers-0.29.2-py3-none-any.whl", hash = "sha256:d5e9bb13c8097b4eed10df23d1294d2e5a418f53e3f89c7ef228b5b982970428"},
+    {file = "diffusers-0.29.2.tar.gz", hash = "sha256:b85f277668e22089cf68b40dd9b76940db7d24ba9cdac107533ed10ab8e4e9db"},
 ]
 
 [package.dependencies]
 filelock = "*"
-huggingface-hub = ">=0.20.2"
+huggingface-hub = ">=0.23.2"
 importlib-metadata = "*"
 numpy = "*"
 Pillow = "*"
@@ -635,13 +635,13 @@ requests = "*"
 safetensors = ">=0.3.1"
 
 [package.extras]
-dev = ["GitPython (<3.1.19)", "Jinja2", "accelerate (>=0.11.0)", "compel (==0.1.8)", "datasets", "flax (>=0.4.1)", "hf-doc-builder (>=0.3.0)", "invisible-watermark (>=0.2.0)", "isort (>=5.5.4)", "jax (>=0.4.1)", "jaxlib (>=0.4.1)", "k-diffusion (>=0.0.12)", "librosa", "parameterized", "peft (>=0.6.0)", "protobuf (>=3.20.3,<4)", "pytest", "pytest-timeout", "pytest-xdist", "requests-mock (==1.10.0)", "ruff (==0.1.5)", "safetensors (>=0.3.1)", "scipy", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "torch (>=1.4)", "torchvision", "transformers (>=4.25.1)", "urllib3 (<=2.0.0)"]
+dev = ["GitPython (<3.1.19)", "Jinja2", "accelerate (>=0.29.3)", "compel (==0.1.8)", "datasets", "flax (>=0.4.1)", "hf-doc-builder (>=0.3.0)", "invisible-watermark (>=0.2.0)", "isort (>=5.5.4)", "jax (>=0.4.1)", "jaxlib (>=0.4.1)", "k-diffusion (>=0.0.12)", "librosa", "parameterized", "peft (>=0.6.0)", "protobuf (>=3.20.3,<4)", "pytest", "pytest-timeout", "pytest-xdist", "requests-mock (==1.10.0)", "ruff (==0.1.5)", "safetensors (>=0.3.1)", "scipy", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "torch (>=1.4)", "torchvision", "transformers (>=4.25.1)", "urllib3 (<=2.0.0)"]
 docs = ["hf-doc-builder (>=0.3.0)"]
 flax = ["flax (>=0.4.1)", "jax (>=0.4.1)", "jaxlib (>=0.4.1)"]
 quality = ["hf-doc-builder (>=0.3.0)", "isort (>=5.5.4)", "ruff (==0.1.5)", "urllib3 (<=2.0.0)"]
 test = ["GitPython (<3.1.19)", "Jinja2", "compel (==0.1.8)", "datasets", "invisible-watermark (>=0.2.0)", "k-diffusion (>=0.0.12)", "librosa", "parameterized", "pytest", "pytest-timeout", "pytest-xdist", "requests-mock (==1.10.0)", "safetensors (>=0.3.1)", "scipy", "sentencepiece (>=0.1.91,!=0.1.92)", "torchvision", "transformers (>=4.25.1)"]
-torch = ["accelerate (>=0.11.0)", "torch (>=1.4)"]
-training = ["Jinja2", "accelerate (>=0.11.0)", "datasets", "peft (>=0.6.0)", "protobuf (>=3.20.3,<4)", "tensorboard"]
+torch = ["accelerate (>=0.29.3)", "torch (>=1.4)"]
+training = ["Jinja2", "accelerate (>=0.29.3)", "datasets", "peft (>=0.6.0)", "protobuf (>=3.20.3,<4)", "tensorboard"]
 
 [[package]]
 name = "dill"
@@ -795,7 +795,6 @@ files = [
     {file = "dora_rs-0.3.5-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:01f811d0c6722f74743c153a7be0144686daeafa968c473e60f6b6c5dc8f5bff"},
     {file = "dora_rs-0.3.5-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:a36e97d31eeb66e6d5913130695d188ceee1248029961012a8b4f59fd3f58670"},
     {file = "dora_rs-0.3.5-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:25d620123a733661dc740ef2b456601ddbaa69ae2b50d8141daa3c684bda385c"},
-    {file = "dora_rs-0.3.5-cp37-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:a9fdc4e73578bebb1c8d0f8bea2243a5a9e179f08c74d98576123b59b75e5cac"},
     {file = "dora_rs-0.3.5-cp37-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e65830634c58158557f0ab90e5d1f492bcbc6b74587b05825ba4c20b634dc1bd"},
     {file = "dora_rs-0.3.5-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c01f9ab8f93295341aeab2d606d484d9cff9d05f57581e2180433ec8e0d38307"},
     {file = "dora_rs-0.3.5-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:5d6d46a49a34cd7e4f74496a1089b9a1b78282c219a28d98fe031a763e92d530"},
@@ -843,13 +842,13 @@ files = [
 
 [[package]]
 name = "exceptiongroup"
-version = "1.2.1"
+version = "1.2.2"
 description = "Backport of PEP 654 (exception groups)"
 optional = true
 python-versions = ">=3.7"
 files = [
-    {file = "exceptiongroup-1.2.1-py3-none-any.whl", hash = "sha256:5258b9ed329c5bbdd31a309f53cbfb0b155341807f6ff7606a1e801a891b29ad"},
-    {file = "exceptiongroup-1.2.1.tar.gz", hash = "sha256:a4785e48b045528f5bfe627b6ad554ff32def154f42372786903b7abcfe1aa16"},
+    {file = "exceptiongroup-1.2.2-py3-none-any.whl", hash = "sha256:3111b9d131c238bec2f8f516e123e14ba243563fb135d3fe885990585aa7795b"},
+    {file = "exceptiongroup-1.2.2.tar.gz", hash = "sha256:47c2edf7c6738fafb49fd34290706d1a1a2f4d1c6df275526b62cbb4aa5393cc"},
 ]
 
 [package.extras]
@@ -1132,7 +1131,7 @@ pyarrow = ">=12.0.0"
 type = "git"
 url = "https://github.com/dora-rs/dora-lerobot.git"
 reference = "HEAD"
-resolved_reference = "2addd1131a3c94f7b70b805577901b7967853e98"
+resolved_reference = "fda22deba84c46695369736edd34dc740aef45eb"
 subdirectory = "gym_dora"
 
 [[package]]
@@ -1338,13 +1337,13 @@ files = [
 
 [[package]]
 name = "huggingface-hub"
-version = "0.23.4"
+version = "0.23.5"
 description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub"
 optional = false
 python-versions = ">=3.8.0"
 files = [
-    {file = "huggingface_hub-0.23.4-py3-none-any.whl", hash = "sha256:3a0b957aa87150addf0cc7bd71b4d954b78e749850e1e7fb29ebbd2db64ca037"},
-    {file = "huggingface_hub-0.23.4.tar.gz", hash = "sha256:35d99016433900e44ae7efe1c209164a5a81dbbcd53a52f99c281dcd7ce22431"},
+    {file = "huggingface_hub-0.23.5-py3-none-any.whl", hash = "sha256:d7a7d337615e11a45cc14a0ce5a605db6b038dc24af42866f731684825226e90"},
+    {file = "huggingface_hub-0.23.5.tar.gz", hash = "sha256:67a9caba79b71235be3752852ca27da86bd54311d2424ca8afdb8dda056edf98"},
 ]
 
 [package.dependencies]
@@ -1389,13 +1388,13 @@ packaging = "*"
 
 [[package]]
 name = "identify"
-version = "2.5.36"
+version = "2.6.0"
 description = "File identification library for Python"
 optional = true
 python-versions = ">=3.8"
 files = [
-    {file = "identify-2.5.36-py2.py3-none-any.whl", hash = "sha256:37d93f380f4de590500d9dba7db359d0d3da95ffe7f9de1753faa159e71e7dfa"},
-    {file = "identify-2.5.36.tar.gz", hash = "sha256:e5e00f54165f9047fbebeb4a560f9acfb8af4c88232be60a488e9b68d122745d"},
+    {file = "identify-2.6.0-py2.py3-none-any.whl", hash = "sha256:e79ae4406387a9d300332b5fd366d8994f1525e8414984e1a59e058b2eda2dd0"},
+    {file = "identify-2.6.0.tar.gz", hash = "sha256:cb171c685bdc31bcc4c1734698736a7d5b6c8bf2e0c15117f4d469c8640ae5cf"},
 ]
 
 [package.extras]
@@ -1742,13 +1741,9 @@ files = [
     {file = "lxml-5.2.2-cp36-cp36m-win_amd64.whl", hash = "sha256:edcfa83e03370032a489430215c1e7783128808fd3e2e0a3225deee278585196"},
     {file = "lxml-5.2.2-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:28bf95177400066596cdbcfc933312493799382879da504633d16cf60bba735b"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3a745cc98d504d5bd2c19b10c79c61c7c3df9222629f1b6210c0368177589fb8"},
-    {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1b590b39ef90c6b22ec0be925b211298e810b4856909c8ca60d27ffbca6c12e6"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b336b0416828022bfd5a2e3083e7f5ba54b96242159f83c7e3eebaec752f1716"},
-    {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_28_aarch64.whl", hash = "sha256:c2faf60c583af0d135e853c86ac2735ce178f0e338a3c7f9ae8f622fd2eb788c"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_28_x86_64.whl", hash = "sha256:4bc6cb140a7a0ad1f7bc37e018d0ed690b7b6520ade518285dc3171f7a117905"},
-    {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:7ff762670cada8e05b32bf1e4dc50b140790909caa8303cfddc4d702b71ea184"},
     {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:57f0a0bbc9868e10ebe874e9f129d2917750adf008fe7b9c1598c0fbbfdde6a6"},
-    {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:a6d2092797b388342c1bc932077ad232f914351932353e2e8706851c870bca1f"},
     {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:60499fe961b21264e17a471ec296dcbf4365fbea611bf9e303ab69db7159ce61"},
     {file = "lxml-5.2.2-cp37-cp37m-win32.whl", hash = "sha256:d9b342c76003c6b9336a80efcc766748a333573abf9350f4094ee46b006ec18f"},
     {file = "lxml-5.2.2-cp37-cp37m-win_amd64.whl", hash = "sha256:b16db2770517b8799c79aa80f4053cd6f8b716f21f8aca962725a9565ce3ee40"},
@@ -2175,43 +2170,36 @@ numpy = ">=1.22,<2.1"
 
 [[package]]
 name = "numcodecs"
-version = "0.12.1"
+version = "0.13.0"
 description = "A Python package providing buffer compression and transformation codecs for use in data storage and communication applications."
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.10"
 files = [
-    {file = "numcodecs-0.12.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d37f628fe92b3699e65831d5733feca74d2e33b50ef29118ffd41c13c677210e"},
-    {file = "numcodecs-0.12.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:941b7446b68cf79f089bcfe92edaa3b154533dcbcd82474f994b28f2eedb1c60"},
-    {file = "numcodecs-0.12.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0e79bf9d1d37199ac00a60ff3adb64757523291d19d03116832e600cac391c51"},
-    {file = "numcodecs-0.12.1-cp310-cp310-win_amd64.whl", hash = "sha256:82d7107f80f9307235cb7e74719292d101c7ea1e393fe628817f0d635b7384f5"},
-    {file = "numcodecs-0.12.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:eeaf42768910f1c6eebf6c1bb00160728e62c9343df9e2e315dc9fe12e3f6071"},
-    {file = "numcodecs-0.12.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:135b2d47563f7b9dc5ee6ce3d1b81b0f1397f69309e909f1a35bb0f7c553d45e"},
-    {file = "numcodecs-0.12.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a191a8e347ecd016e5c357f2bf41fbcb026f6ffe78fff50c77ab12e96701d155"},
-    {file = "numcodecs-0.12.1-cp311-cp311-win_amd64.whl", hash = "sha256:21d8267bd4313f4d16f5b6287731d4c8ebdab236038f29ad1b0e93c9b2ca64ee"},
-    {file = "numcodecs-0.12.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:2f84df6b8693206365a5b37c005bfa9d1be486122bde683a7b6446af4b75d862"},
-    {file = "numcodecs-0.12.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:760627780a8b6afdb7f942f2a0ddaf4e31d3d7eea1d8498cf0fd3204a33c4618"},
-    {file = "numcodecs-0.12.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c258bd1d3dfa75a9b708540d23b2da43d63607f9df76dfa0309a7597d1de3b73"},
-    {file = "numcodecs-0.12.1-cp312-cp312-win_amd64.whl", hash = "sha256:e04649ea504aff858dbe294631f098fbfd671baf58bfc04fc48d746554c05d67"},
-    {file = "numcodecs-0.12.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:caf1a1e6678aab9c1e29d2109b299f7a467bd4d4c34235b1f0e082167846b88f"},
-    {file = "numcodecs-0.12.1-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:c17687b1fd1fef68af616bc83f896035d24e40e04e91e7e6dae56379eb59fe33"},
-    {file = "numcodecs-0.12.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:29dfb195f835a55c4d490fb097aac8c1bcb96c54cf1b037d9218492c95e9d8c5"},
-    {file = "numcodecs-0.12.1-cp38-cp38-win_amd64.whl", hash = "sha256:2f1ba2f4af3fd3ba65b1bcffb717fe65efe101a50a91c368f79f3101dbb1e243"},
-    {file = "numcodecs-0.12.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:2fbb12a6a1abe95926f25c65e283762d63a9bf9e43c0de2c6a1a798347dfcb40"},
-    {file = "numcodecs-0.12.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:f2207871868b2464dc11c513965fd99b958a9d7cde2629be7b2dc84fdaab013b"},
-    {file = "numcodecs-0.12.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:abff3554a6892a89aacf7b642a044e4535499edf07aeae2f2e6e8fc08c9ba07f"},
-    {file = "numcodecs-0.12.1-cp39-cp39-win_amd64.whl", hash = "sha256:ef964d4860d3e6b38df0633caf3e51dc850a6293fd8e93240473642681d95136"},
-    {file = "numcodecs-0.12.1.tar.gz", hash = "sha256:05d91a433733e7eef268d7e80ec226a0232da244289614a8f3826901aec1098e"},
+    {file = "numcodecs-0.13.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:56e49f68ce6aeba29f144992524c8897d94f846d02bbcc820dd29d7c5c2a073e"},
+    {file = "numcodecs-0.13.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:17bc4b568214582f4c623700592f633f3afd920848630049c584fa1e535253ad"},
+    {file = "numcodecs-0.13.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eed420a9c62d0a569aa94a387f93045f068ad3e7bbd787c6ce70bc5fefbaa7d9"},
+    {file = "numcodecs-0.13.0-cp310-cp310-win_amd64.whl", hash = "sha256:e7d3b9693df52eeaf978d2a56971d01cf9b4e284ae769ec764807f2087cce51d"},
+    {file = "numcodecs-0.13.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:f208a1b8b5e66c767ed043812ca74d9045e09b7b2e085d064a585c30b9efc8e7"},
+    {file = "numcodecs-0.13.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a68368d3ce625ec76fcacd84785f6110d30a232909d5c6093a7aa25628880477"},
+    {file = "numcodecs-0.13.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f5904216811f2e9d312c23ffaad3b3d4c7442a3583d3a8bf81ca8319e9f5deb5"},
+    {file = "numcodecs-0.13.0-cp311-cp311-win_amd64.whl", hash = "sha256:208cab0f4d9cf4409e9c4a4c935e165833786614822c81dee9d865af372da9df"},
+    {file = "numcodecs-0.13.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:f3cf462d2357998d7f6baaa0427657b0eeda3eb79fba2b146d2d04542912a513"},
+    {file = "numcodecs-0.13.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ac4dd5556fb126271e93bd1a02266e21b01d3617db448d70d00eec8e034506b4"},
+    {file = "numcodecs-0.13.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:820be89729583c91601a6b35c052008cdd2665b25bfedb91b367cc155fb34ba0"},
+    {file = "numcodecs-0.13.0-cp312-cp312-win_amd64.whl", hash = "sha256:d67a859dd8a7f026829e91cb1799c26720cc9d29ee4ae0060cc7a581670abc06"},
+    {file = "numcodecs-0.13.0.tar.gz", hash = "sha256:ba4fac7036ea5a078c7afe1d4dffeb9685080d42f19c9c16b12dad866703aa2e"},
 ]
 
 [package.dependencies]
 numpy = ">=1.7"
 
 [package.extras]
-docs = ["mock", "numpydoc", "sphinx (<7.0.0)", "sphinx-issues"]
+docs = ["mock", "numpydoc", "pydata-sphinx-theme", "sphinx (<7.0.0)", "sphinx-issues"]
 msgpack = ["msgpack"]
-test = ["coverage", "flake8", "pytest", "pytest-cov"]
+pcodec = ["pcodec (>=0.2.0)"]
+test = ["coverage", "pytest", "pytest-cov"]
 test-extras = ["importlib-metadata"]
-zfpy = ["zfpy (>=1.0.0)"]
+zfpy = ["numpy (<2.0.0)", "zfpy (>=1.0.0)"]
 
 [[package]]
 name = "numpy"
@@ -2773,52 +2761,55 @@ test = ["enum34", "ipaddress", "mock", "pywin32", "wmi"]
 
 [[package]]
 name = "pyarrow"
-version = "16.1.0"
+version = "17.0.0"
 description = "Python library for Apache Arrow"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "pyarrow-16.1.0-cp310-cp310-macosx_10_15_x86_64.whl", hash = "sha256:17e23b9a65a70cc733d8b738baa6ad3722298fa0c81d88f63ff94bf25eaa77b9"},
-    {file = "pyarrow-16.1.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4740cc41e2ba5d641071d0ab5e9ef9b5e6e8c7611351a5cb7c1d175eaf43674a"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:98100e0268d04e0eec47b73f20b39c45b4006f3c4233719c3848aa27a03c1aef"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f68f409e7b283c085f2da014f9ef81e885d90dcd733bd648cfba3ef265961848"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:a8914cd176f448e09746037b0c6b3a9d7688cef451ec5735094055116857580c"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:48be160782c0556156d91adbdd5a4a7e719f8d407cb46ae3bb4eaee09b3111bd"},
-    {file = "pyarrow-16.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:9cf389d444b0f41d9fe1444b70650fea31e9d52cfcb5f818b7888b91b586efff"},
-    {file = "pyarrow-16.1.0-cp311-cp311-macosx_10_15_x86_64.whl", hash = "sha256:d0ebea336b535b37eee9eee31761813086d33ed06de9ab6fc6aaa0bace7b250c"},
-    {file = "pyarrow-16.1.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2e73cfc4a99e796727919c5541c65bb88b973377501e39b9842ea71401ca6c1c"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bf9251264247ecfe93e5f5a0cd43b8ae834f1e61d1abca22da55b20c788417f6"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ddf5aace92d520d3d2a20031d8b0ec27b4395cab9f74e07cc95edf42a5cc0147"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:25233642583bf658f629eb230b9bb79d9af4d9f9229890b3c878699c82f7d11e"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:a33a64576fddfbec0a44112eaf844c20853647ca833e9a647bfae0582b2ff94b"},
-    {file = "pyarrow-16.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:185d121b50836379fe012753cf15c4ba9638bda9645183ab36246923875f8d1b"},
-    {file = "pyarrow-16.1.0-cp312-cp312-macosx_10_15_x86_64.whl", hash = "sha256:2e51ca1d6ed7f2e9d5c3c83decf27b0d17bb207a7dea986e8dc3e24f80ff7d6f"},
-    {file = "pyarrow-16.1.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:06ebccb6f8cb7357de85f60d5da50e83507954af617d7b05f48af1621d331c9a"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b04707f1979815f5e49824ce52d1dceb46e2f12909a48a6a753fe7cafbc44a0c"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0d32000693deff8dc5df444b032b5985a48592c0697cb6e3071a5d59888714e2"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:8785bb10d5d6fd5e15d718ee1d1f914fe768bf8b4d1e5e9bf253de8a26cb1628"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:e1369af39587b794873b8a307cc6623a3b1194e69399af0efd05bb202195a5a7"},
-    {file = "pyarrow-16.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:febde33305f1498f6df85e8020bca496d0e9ebf2093bab9e0f65e2b4ae2b3444"},
-    {file = "pyarrow-16.1.0-cp38-cp38-macosx_10_15_x86_64.whl", hash = "sha256:b5f5705ab977947a43ac83b52ade3b881eb6e95fcc02d76f501d549a210ba77f"},
-    {file = "pyarrow-16.1.0-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:0d27bf89dfc2576f6206e9cd6cf7a107c9c06dc13d53bbc25b0bd4556f19cf5f"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0d07de3ee730647a600037bc1d7b7994067ed64d0eba797ac74b2bc77384f4c2"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fbef391b63f708e103df99fbaa3acf9f671d77a183a07546ba2f2c297b361e83"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_28_aarch64.whl", hash = "sha256:19741c4dbbbc986d38856ee7ddfdd6a00fc3b0fc2d928795b95410d38bb97d15"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_28_x86_64.whl", hash = "sha256:f2c5fb249caa17b94e2b9278b36a05ce03d3180e6da0c4c3b3ce5b2788f30eed"},
-    {file = "pyarrow-16.1.0-cp38-cp38-win_amd64.whl", hash = "sha256:e6b6d3cd35fbb93b70ade1336022cc1147b95ec6af7d36906ca7fe432eb09710"},
-    {file = "pyarrow-16.1.0-cp39-cp39-macosx_10_15_x86_64.whl", hash = "sha256:18da9b76a36a954665ccca8aa6bd9f46c1145f79c0bb8f4f244f5f8e799bca55"},
-    {file = "pyarrow-16.1.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:99f7549779b6e434467d2aa43ab2b7224dd9e41bdde486020bae198978c9e05e"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f07fdffe4fd5b15f5ec15c8b64584868d063bc22b86b46c9695624ca3505b7b4"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ddfe389a08ea374972bd4065d5f25d14e36b43ebc22fc75f7b951f24378bf0b5"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_28_aarch64.whl", hash = "sha256:3b20bd67c94b3a2ea0a749d2a5712fc845a69cb5d52e78e6449bbd295611f3aa"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_28_x86_64.whl", hash = "sha256:ba8ac20693c0bb0bf4b238751d4409e62852004a8cf031c73b0e0962b03e45e3"},
-    {file = "pyarrow-16.1.0-cp39-cp39-win_amd64.whl", hash = "sha256:31a1851751433d89a986616015841977e0a188662fcffd1a5677453f1df2de0a"},
-    {file = "pyarrow-16.1.0.tar.gz", hash = "sha256:15fbb22ea96d11f0b5768504a3f961edab25eaf4197c341720c4a387f6c60315"},
+    {file = "pyarrow-17.0.0-cp310-cp310-macosx_10_15_x86_64.whl", hash = "sha256:a5c8b238d47e48812ee577ee20c9a2779e6a5904f1708ae240f53ecbee7c9f07"},
+    {file = "pyarrow-17.0.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:db023dc4c6cae1015de9e198d41250688383c3f9af8f565370ab2b4cb5f62655"},
+    {file = "pyarrow-17.0.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:da1e060b3876faa11cee287839f9cc7cdc00649f475714b8680a05fd9071d545"},
+    {file = "pyarrow-17.0.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:75c06d4624c0ad6674364bb46ef38c3132768139ddec1c56582dbac54f2663e2"},
+    {file = "pyarrow-17.0.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:fa3c246cc58cb5a4a5cb407a18f193354ea47dd0648194e6265bd24177982fe8"},
+    {file = "pyarrow-17.0.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:f7ae2de664e0b158d1607699a16a488de3d008ba99b3a7aa5de1cbc13574d047"},
+    {file = "pyarrow-17.0.0-cp310-cp310-win_amd64.whl", hash = "sha256:5984f416552eea15fd9cee03da53542bf4cddaef5afecefb9aa8d1010c335087"},
+    {file = "pyarrow-17.0.0-cp311-cp311-macosx_10_15_x86_64.whl", hash = "sha256:1c8856e2ef09eb87ecf937104aacfa0708f22dfeb039c363ec99735190ffb977"},
+    {file = "pyarrow-17.0.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2e19f569567efcbbd42084e87f948778eb371d308e137a0f97afe19bb860ccb3"},
+    {file = "pyarrow-17.0.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6b244dc8e08a23b3e352899a006a26ae7b4d0da7bb636872fa8f5884e70acf15"},
+    {file = "pyarrow-17.0.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0b72e87fe3e1db343995562f7fff8aee354b55ee83d13afba65400c178ab2597"},
+    {file = "pyarrow-17.0.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:dc5c31c37409dfbc5d014047817cb4ccd8c1ea25d19576acf1a001fe07f5b420"},
+    {file = "pyarrow-17.0.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:e3343cb1e88bc2ea605986d4b94948716edc7a8d14afd4e2c097232f729758b4"},
+    {file = "pyarrow-17.0.0-cp311-cp311-win_amd64.whl", hash = "sha256:a27532c38f3de9eb3e90ecab63dfda948a8ca859a66e3a47f5f42d1e403c4d03"},
+    {file = "pyarrow-17.0.0-cp312-cp312-macosx_10_15_x86_64.whl", hash = "sha256:9b8a823cea605221e61f34859dcc03207e52e409ccf6354634143e23af7c8d22"},
+    {file = "pyarrow-17.0.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f1e70de6cb5790a50b01d2b686d54aaf73da01266850b05e3af2a1bc89e16053"},
+    {file = "pyarrow-17.0.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0071ce35788c6f9077ff9ecba4858108eebe2ea5a3f7cf2cf55ebc1dbc6ee24a"},
+    {file = "pyarrow-17.0.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:757074882f844411fcca735e39aae74248a1531367a7c80799b4266390ae51cc"},
+    {file = "pyarrow-17.0.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:9ba11c4f16976e89146781a83833df7f82077cdab7dc6232c897789343f7891a"},
+    {file = "pyarrow-17.0.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:b0c6ac301093b42d34410b187bba560b17c0330f64907bfa4f7f7f2444b0cf9b"},
+    {file = "pyarrow-17.0.0-cp312-cp312-win_amd64.whl", hash = "sha256:392bc9feabc647338e6c89267635e111d71edad5fcffba204425a7c8d13610d7"},
+    {file = "pyarrow-17.0.0-cp38-cp38-macosx_10_15_x86_64.whl", hash = "sha256:af5ff82a04b2171415f1410cff7ebb79861afc5dae50be73ce06d6e870615204"},
+    {file = "pyarrow-17.0.0-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:edca18eaca89cd6382dfbcff3dd2d87633433043650c07375d095cd3517561d8"},
+    {file = "pyarrow-17.0.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7c7916bff914ac5d4a8fe25b7a25e432ff921e72f6f2b7547d1e325c1ad9d155"},
+    {file = "pyarrow-17.0.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f553ca691b9e94b202ff741bdd40f6ccb70cdd5fbf65c187af132f1317de6145"},
+    {file = "pyarrow-17.0.0-cp38-cp38-manylinux_2_28_aarch64.whl", hash = "sha256:0cdb0e627c86c373205a2f94a510ac4376fdc523f8bb36beab2e7f204416163c"},
+    {file = "pyarrow-17.0.0-cp38-cp38-manylinux_2_28_x86_64.whl", hash = "sha256:d7d192305d9d8bc9082d10f361fc70a73590a4c65cf31c3e6926cd72b76bc35c"},
+    {file = "pyarrow-17.0.0-cp38-cp38-win_amd64.whl", hash = "sha256:02dae06ce212d8b3244dd3e7d12d9c4d3046945a5933d28026598e9dbbda1fca"},
+    {file = "pyarrow-17.0.0-cp39-cp39-macosx_10_15_x86_64.whl", hash = "sha256:13d7a460b412f31e4c0efa1148e1d29bdf18ad1411eb6757d38f8fbdcc8645fb"},
+    {file = "pyarrow-17.0.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:9b564a51fbccfab5a04a80453e5ac6c9954a9c5ef2890d1bcf63741909c3f8df"},
+    {file = "pyarrow-17.0.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:32503827abbc5aadedfa235f5ece8c4f8f8b0a3cf01066bc8d29de7539532687"},
+    {file = "pyarrow-17.0.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a155acc7f154b9ffcc85497509bcd0d43efb80d6f733b0dc3bb14e281f131c8b"},
+    {file = "pyarrow-17.0.0-cp39-cp39-manylinux_2_28_aarch64.whl", hash = "sha256:dec8d129254d0188a49f8a1fc99e0560dc1b85f60af729f47de4046015f9b0a5"},
+    {file = "pyarrow-17.0.0-cp39-cp39-manylinux_2_28_x86_64.whl", hash = "sha256:a48ddf5c3c6a6c505904545c25a4ae13646ae1f8ba703c4df4a1bfe4f4006bda"},
+    {file = "pyarrow-17.0.0-cp39-cp39-win_amd64.whl", hash = "sha256:42bf93249a083aca230ba7e2786c5f673507fa97bbd9725a1e2754715151a204"},
+    {file = "pyarrow-17.0.0.tar.gz", hash = "sha256:4beca9521ed2c0921c1023e68d097d0299b62c362639ea315572a58f3f50fd28"},
 ]
 
 [package.dependencies]
 numpy = ">=1.16.6"
 
+[package.extras]
+test = ["cffi", "hypothesis", "pandas", "pytest", "pytz"]
+
 [[package]]
 name = "pyarrow-hotfix"
 version = "0.6"
@@ -3221,23 +3212,6 @@ pytest = ">=4.6"
 [package.extras]
 testing = ["fields", "hunter", "process-tests", "pytest-xdist", "virtualenv"]
 
-[[package]]
-name = "pytest-mock"
-version = "3.14.0"
-description = "Thin-wrapper around the mock package for easier use with pytest"
-optional = true
-python-versions = ">=3.8"
-files = [
-    {file = "pytest-mock-3.14.0.tar.gz", hash = "sha256:2719255a1efeceadbc056d6bf3df3d1c5015530fb40cf347c0f9afac88410bd0"},
-    {file = "pytest_mock-3.14.0-py3-none-any.whl", hash = "sha256:0b72c38033392a5f4621342fe11e9219ac11ec9d375f8e2a0c164539e0d70f6f"},
-]
-
-[package.dependencies]
-pytest = ">=6.2.5"
-
-[package.extras]
-dev = ["pre-commit", "pytest-asyncio", "tox"]
-
 [[package]]
 name = "python-dateutil"
 version = "2.9.0.post0"
@@ -3302,7 +3276,6 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
-    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -3449,16 +3422,16 @@ use-chardet-on-py3 = ["chardet (>=3.0.2,<6)"]
 
 [[package]]
 name = "rerun-sdk"
-version = "0.16.1"
+version = "0.17.0"
 description = "The Rerun Logging SDK"
 optional = false
 python-versions = "<3.13,>=3.8"
 files = [
-    {file = "rerun_sdk-0.16.1-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:170c6976634008611753e10dfef8cdc395ce8180e634c169e7c61cef2f89a277"},
-    {file = "rerun_sdk-0.16.1-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:c9a76eab7eb5559276737dad655200e9350df0837158dbc5a896970ab4201454"},
-    {file = "rerun_sdk-0.16.1-cp38-abi3-manylinux_2_31_aarch64.whl", hash = "sha256:4d6436752d57e8b8038489a0e7e37f0c760b088e96db5fb81667d3a376d63fea"},
-    {file = "rerun_sdk-0.16.1-cp38-abi3-manylinux_2_31_x86_64.whl", hash = "sha256:37b7b47948471873e84f224b16f417a94a91c7cbd6c72c68281eeff1ba414b8f"},
-    {file = "rerun_sdk-0.16.1-cp38-abi3-win_amd64.whl", hash = "sha256:be88799c8afdf68eafa99e64e2e4f0a484e187e017a180219abbe6bb988acd4e"},
+    {file = "rerun_sdk-0.17.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:abd34f746eada83b8bb0bc50007183151981d7ccf18306f3d42165819a3f6fcb"},
+    {file = "rerun_sdk-0.17.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:8b0a8a6feab3f8e679801d158216a71d88a81480021587719330f50d083c4d26"},
+    {file = "rerun_sdk-0.17.0-cp38-abi3-manylinux_2_31_aarch64.whl", hash = "sha256:ad55807abafb01e527846742e087819aac8e103f1ec15aadc563a4038bb44e1d"},
+    {file = "rerun_sdk-0.17.0-cp38-abi3-manylinux_2_31_x86_64.whl", hash = "sha256:9d41f1f475270b1e0d50ddb8cb62e0d828988f0c371ac8457af25c8be5aa1dc0"},
+    {file = "rerun_sdk-0.17.0-cp38-abi3-win_amd64.whl", hash = "sha256:34e5595a326cbdddfebdf00b08e877358c564fce74cc8c6d617fc89ef3a6aa70"},
 ]
 
 [package.dependencies]
@@ -3469,6 +3442,7 @@ pyarrow = ">=14.0.2"
 typing-extensions = ">=4.5"
 
 [package.extras]
+notebook = ["rerun-notebook (==0.17.0)"]
 tests = ["pytest (==7.1.2)"]
 
 [[package]]
@@ -3595,27 +3569,32 @@ torch = ["safetensors[numpy]", "torch (>=1.10)"]
 
 [[package]]
 name = "scikit-image"
-version = "0.23.2"
+version = "0.24.0"
 description = "Image processing in Python"
 optional = true
-python-versions = ">=3.10"
+python-versions = ">=3.9"
 files = [
-    {file = "scikit_image-0.23.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:f9a8db6c52f8d0e1474ea8320d7b8db442b4d6baa29dd0acbd02f8a49572f18a"},
-    {file = "scikit_image-0.23.2-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:524b51a7440e46ed2ebbde7bc288bf2dde1dee2caafdd9513b2aca38a48223b7"},
-    {file = "scikit_image-0.23.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8b335c229170d787b3fb8c60d220f72049ccf862d5191a3cfda6ac84b995ac4e"},
-    {file = "scikit_image-0.23.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:08b10781efbd6b084f3c847ff4049b657241ea866b9e331b14bf791dcb3e6661"},
-    {file = "scikit_image-0.23.2-cp310-cp310-win_amd64.whl", hash = "sha256:a207352e9a1956dda1424bbe872c7795345187138118e8be6a421aef3b988c2a"},
-    {file = "scikit_image-0.23.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ee83fdb1843ee938eabdfeb9498623282935ea30aa20dffc5d5d16698efb4b2a"},
-    {file = "scikit_image-0.23.2-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:a158f50d3df4867bbd1c698520ede8bc493e430ad83f54ac1f0d8f57b328779b"},
-    {file = "scikit_image-0.23.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:55de3326be124334b89314e9e04c8971ad98d6681e11a243f71bfb85ef9554b0"},
-    {file = "scikit_image-0.23.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fce619a6d84fe40c1208fa579b646e93ce13ef0afc3652a23e9782b2c183291a"},
-    {file = "scikit_image-0.23.2-cp311-cp311-win_amd64.whl", hash = "sha256:ee65669aa586e110346f567ed5c92d1bd63799a19e951cb83da3f54b0caf7c52"},
-    {file = "scikit_image-0.23.2-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:15bfb4e8d7bd90a967e6a3c3ab6be678063fc45e950b730684a8db46a02ff892"},
-    {file = "scikit_image-0.23.2-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:5736e66d01b11cd90988ec24ab929c80a03af28f690189c951886891ebf63154"},
-    {file = "scikit_image-0.23.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3597ac5d8f51dafbcb7433ef1fdefdefb535f50745b2002ae0a5d651df4f063b"},
-    {file = "scikit_image-0.23.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1978be2abe3c3c3189a99a411d48bbb1306f7c2debb3aefbf426e23947f26623"},
-    {file = "scikit_image-0.23.2-cp312-cp312-win_amd64.whl", hash = "sha256:ae32bf0cb02b672ed74d28880ca6f88928ae8dd794d67e04fa3ff4836feb9bd6"},
-    {file = "scikit_image-0.23.2.tar.gz", hash = "sha256:c9da4b2c3117e3e30364a3d14496ee5c72b09eb1a4ab1292b302416faa360590"},
+    {file = "scikit_image-0.24.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:cb3bc0264b6ab30b43c4179ee6156bc18b4861e78bb329dd8d16537b7bbf827a"},
+    {file = "scikit_image-0.24.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:9c7a52e20cdd760738da38564ba1fed7942b623c0317489af1a598a8dedf088b"},
+    {file = "scikit_image-0.24.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93f46e6ce42e5409f4d09ce1b0c7f80dd7e4373bcec635b6348b63e3c886eac8"},
+    {file = "scikit_image-0.24.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:39ee0af13435c57351a3397eb379e72164ff85161923eec0c38849fecf1b4764"},
+    {file = "scikit_image-0.24.0-cp310-cp310-win_amd64.whl", hash = "sha256:7ac7913b028b8aa780ffae85922894a69e33d1c0bf270ea1774f382fe8bf95e7"},
+    {file = "scikit_image-0.24.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:272909e02a59cea3ed4aa03739bb88df2625daa809f633f40b5053cf09241831"},
+    {file = "scikit_image-0.24.0-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:190ebde80b4470fe8838764b9b15f232a964f1a20391663e31008d76f0c696f7"},
+    {file = "scikit_image-0.24.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:59c98cc695005faf2b79904e4663796c977af22586ddf1b12d6af2fa22842dc2"},
+    {file = "scikit_image-0.24.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa27b3a0dbad807b966b8db2d78da734cb812ca4787f7fbb143764800ce2fa9c"},
+    {file = "scikit_image-0.24.0-cp311-cp311-win_amd64.whl", hash = "sha256:dacf591ac0c272a111181afad4b788a27fe70d213cfddd631d151cbc34f8ca2c"},
+    {file = "scikit_image-0.24.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:6fccceb54c9574590abcddc8caf6cefa57c13b5b8b4260ab3ff88ad8f3c252b3"},
+    {file = "scikit_image-0.24.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:ccc01e4760d655aab7601c1ba7aa4ddd8b46f494ac46ec9c268df6f33ccddf4c"},
+    {file = "scikit_image-0.24.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:18836a18d3a7b6aca5376a2d805f0045826bc6c9fc85331659c33b4813e0b563"},
+    {file = "scikit_image-0.24.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8579bda9c3f78cb3b3ed8b9425213c53a25fa7e994b7ac01f2440b395babf660"},
+    {file = "scikit_image-0.24.0-cp312-cp312-win_amd64.whl", hash = "sha256:82ab903afa60b2da1da2e6f0c8c65e7c8868c60a869464c41971da929b3e82bc"},
+    {file = "scikit_image-0.24.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:ef04360eda372ee5cd60aebe9be91258639c86ae2ea24093fb9182118008d009"},
+    {file = "scikit_image-0.24.0-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:e9aadb442360a7e76f0c5c9d105f79a83d6df0e01e431bd1d5757e2c5871a1f3"},
+    {file = "scikit_image-0.24.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5e37de6f4c1abcf794e13c258dc9b7d385d5be868441de11c180363824192ff7"},
+    {file = "scikit_image-0.24.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4688c18bd7ec33c08d7bf0fd19549be246d90d5f2c1d795a89986629af0a1e83"},
+    {file = "scikit_image-0.24.0-cp39-cp39-win_amd64.whl", hash = "sha256:56dab751d20b25d5d3985e95c9b4e975f55573554bd76b0aedf5875217c93e69"},
+    {file = "scikit_image-0.24.0.tar.gz", hash = "sha256:5d16efe95da8edbeb363e0c4157b99becbd650a60b77f6e3af5768b66cf007ab"},
 ]
 
 [package.dependencies]
@@ -3680,13 +3659,13 @@ test = ["Cython", "array-api-strict", "asv", "gmpy2", "hypothesis (>=6.30)", "me
 
 [[package]]
 name = "sentry-sdk"
-version = "2.7.1"
+version = "2.10.0"
 description = "Python client for Sentry (https://sentry.io)"
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "sentry_sdk-2.7.1-py2.py3-none-any.whl", hash = "sha256:ef1b3d54eb715825657cd4bb3cb42bb4dc85087bac14c56b0fd8c21abd968c9a"},
-    {file = "sentry_sdk-2.7.1.tar.gz", hash = "sha256:25006c7e68b75aaa5e6b9c6a420ece22e8d7daec4b7a906ffd3a8607b67c037b"},
+    {file = "sentry_sdk-2.10.0-py2.py3-none-any.whl", hash = "sha256:87b3d413c87d8e7f816cc9334bff255a83d8b577db2b22042651c30c19c09190"},
+    {file = "sentry_sdk-2.10.0.tar.gz", hash = "sha256:545fcc6e36c335faa6d6cda84669b6e17025f31efbf3b2211ec14efe008b75d1"},
 ]
 
 [package.dependencies]
@@ -3830,67 +3809,63 @@ test = ["pytest"]
 
 [[package]]
 name = "setuptools"
-version = "70.2.0"
+version = "71.0.1"
 description = "Easily download, build, install, upgrade, and uninstall Python packages"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "setuptools-70.2.0-py3-none-any.whl", hash = "sha256:b8b8060bb426838fbe942479c90296ce976249451118ef566a5a0b7d8b78fb05"},
-    {file = "setuptools-70.2.0.tar.gz", hash = "sha256:bd63e505105011b25c3c11f753f7e3b8465ea739efddaccef8f0efac2137bac1"},
+    {file = "setuptools-71.0.1-py3-none-any.whl", hash = "sha256:1eb8ef012efae7f6acbc53ec0abde4bc6746c43087fd215ee09e1df48998711f"},
+    {file = "setuptools-71.0.1.tar.gz", hash = "sha256:c51d7fd29843aa18dad362d4b4ecd917022131425438251f4e3d766c964dd1ad"},
 ]
 
 [package.extras]
-doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier"]
-test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "importlib-metadata", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "jaraco.test", "mypy (==1.10.0)", "packaging (>=23.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-home (>=0.5)", "pytest-mypy", "pytest-perf", "pytest-ruff (>=0.3.2)", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel"]
+core = ["importlib-metadata (>=6)", "importlib-resources (>=5.10.2)", "jaraco.text (>=3.7)", "more-itertools (>=8.8)", "ordered-set (>=3.1.1)", "packaging (>=24)", "platformdirs (>=2.6.2)", "tomli (>=2.0.1)", "wheel (>=0.43.0)"]
+doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (<7.4)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier"]
+test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "importlib-metadata", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "jaraco.test", "mypy (==1.10.0)", "packaging (>=23.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-home (>=0.5)", "pytest-mypy", "pytest-perf", "pytest-ruff (<0.4)", "pytest-ruff (>=0.2.1)", "pytest-ruff (>=0.3.2)", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel"]
 
 [[package]]
 name = "shapely"
-version = "2.0.4"
+version = "2.0.5"
 description = "Manipulation and analysis of geometric objects"
 optional = true
 python-versions = ">=3.7"
 files = [
-    {file = "shapely-2.0.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:011b77153906030b795791f2fdfa2d68f1a8d7e40bce78b029782ade3afe4f2f"},
-    {file = "shapely-2.0.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9831816a5d34d5170aa9ed32a64982c3d6f4332e7ecfe62dc97767e163cb0b17"},
-    {file = "shapely-2.0.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5c4849916f71dc44e19ed370421518c0d86cf73b26e8656192fcfcda08218fbd"},
-    {file = "shapely-2.0.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:841f93a0e31e4c64d62ea570d81c35de0f6cea224568b2430d832967536308e6"},
-    {file = "shapely-2.0.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d2b4431f522b277c79c34b65da128029a9955e4481462cbf7ebec23aab61fc58"},
-    {file = "shapely-2.0.4-cp310-cp310-win32.whl", hash = "sha256:92a41d936f7d6743f343be265ace93b7c57f5b231e21b9605716f5a47c2879e7"},
-    {file = "shapely-2.0.4-cp310-cp310-win_amd64.whl", hash = "sha256:30982f79f21bb0ff7d7d4a4e531e3fcaa39b778584c2ce81a147f95be1cd58c9"},
-    {file = "shapely-2.0.4-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:de0205cb21ad5ddaef607cda9a3191eadd1e7a62a756ea3a356369675230ac35"},
-    {file = "shapely-2.0.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:7d56ce3e2a6a556b59a288771cf9d091470116867e578bebced8bfc4147fbfd7"},
-    {file = "shapely-2.0.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:58b0ecc505bbe49a99551eea3f2e8a9b3b24b3edd2a4de1ac0dc17bc75c9ec07"},
-    {file = "shapely-2.0.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:790a168a808bd00ee42786b8ba883307c0e3684ebb292e0e20009588c426da47"},
-    {file = "shapely-2.0.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4310b5494271e18580d61022c0857eb85d30510d88606fa3b8314790df7f367d"},
-    {file = "shapely-2.0.4-cp311-cp311-win32.whl", hash = "sha256:63f3a80daf4f867bd80f5c97fbe03314348ac1b3b70fb1c0ad255a69e3749879"},
-    {file = "shapely-2.0.4-cp311-cp311-win_amd64.whl", hash = "sha256:c52ed79f683f721b69a10fb9e3d940a468203f5054927215586c5d49a072de8d"},
-    {file = "shapely-2.0.4-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:5bbd974193e2cc274312da16b189b38f5f128410f3377721cadb76b1e8ca5328"},
-    {file = "shapely-2.0.4-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:41388321a73ba1a84edd90d86ecc8bfed55e6a1e51882eafb019f45895ec0f65"},
-    {file = "shapely-2.0.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0776c92d584f72f1e584d2e43cfc5542c2f3dd19d53f70df0900fda643f4bae6"},
-    {file = "shapely-2.0.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c75c98380b1ede1cae9a252c6dc247e6279403fae38c77060a5e6186c95073ac"},
-    {file = "shapely-2.0.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c3e700abf4a37b7b8b90532fa6ed5c38a9bfc777098bc9fbae5ec8e618ac8f30"},
-    {file = "shapely-2.0.4-cp312-cp312-win32.whl", hash = "sha256:4f2ab0faf8188b9f99e6a273b24b97662194160cc8ca17cf9d1fb6f18d7fb93f"},
-    {file = "shapely-2.0.4-cp312-cp312-win_amd64.whl", hash = "sha256:03152442d311a5e85ac73b39680dd64a9892fa42bb08fd83b3bab4fe6999bfa0"},
-    {file = "shapely-2.0.4-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:994c244e004bc3cfbea96257b883c90a86e8cbd76e069718eb4c6b222a56f78b"},
-    {file = "shapely-2.0.4-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:05ffd6491e9e8958b742b0e2e7c346635033d0a5f1a0ea083547fcc854e5d5cf"},
-    {file = "shapely-2.0.4-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2fbdc1140a7d08faa748256438291394967aa54b40009f54e8d9825e75ef6113"},
-    {file = "shapely-2.0.4-cp37-cp37m-win32.whl", hash = "sha256:5af4cd0d8cf2912bd95f33586600cac9c4b7c5053a036422b97cfe4728d2eb53"},
-    {file = "shapely-2.0.4-cp37-cp37m-win_amd64.whl", hash = "sha256:464157509ce4efa5ff285c646a38b49f8c5ef8d4b340f722685b09bb033c5ccf"},
-    {file = "shapely-2.0.4-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:489c19152ec1f0e5c5e525356bcbf7e532f311bff630c9b6bc2db6f04da6a8b9"},
-    {file = "shapely-2.0.4-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:b79bbd648664aa6f44ef018474ff958b6b296fed5c2d42db60078de3cffbc8aa"},
-    {file = "shapely-2.0.4-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:674d7baf0015a6037d5758496d550fc1946f34bfc89c1bf247cabdc415d7747e"},
-    {file = "shapely-2.0.4-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6cd4ccecc5ea5abd06deeaab52fcdba372f649728050c6143cc405ee0c166679"},
-    {file = "shapely-2.0.4-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fb5cdcbbe3080181498931b52a91a21a781a35dcb859da741c0345c6402bf00c"},
-    {file = "shapely-2.0.4-cp38-cp38-win32.whl", hash = "sha256:55a38dcd1cee2f298d8c2ebc60fc7d39f3b4535684a1e9e2f39a80ae88b0cea7"},
-    {file = "shapely-2.0.4-cp38-cp38-win_amd64.whl", hash = "sha256:ec555c9d0db12d7fd777ba3f8b75044c73e576c720a851667432fabb7057da6c"},
-    {file = "shapely-2.0.4-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:3f9103abd1678cb1b5f7e8e1af565a652e036844166c91ec031eeb25c5ca8af0"},
-    {file = "shapely-2.0.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:263bcf0c24d7a57c80991e64ab57cba7a3906e31d2e21b455f493d4aab534aaa"},
-    {file = "shapely-2.0.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:ddf4a9bfaac643e62702ed662afc36f6abed2a88a21270e891038f9a19bc08fc"},
-    {file = "shapely-2.0.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:485246fcdb93336105c29a5cfbff8a226949db37b7473c89caa26c9bae52a242"},
-    {file = "shapely-2.0.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8de4578e838a9409b5b134a18ee820730e507b2d21700c14b71a2b0757396acc"},
-    {file = "shapely-2.0.4-cp39-cp39-win32.whl", hash = "sha256:9dab4c98acfb5fb85f5a20548b5c0abe9b163ad3525ee28822ffecb5c40e724c"},
-    {file = "shapely-2.0.4-cp39-cp39-win_amd64.whl", hash = "sha256:31c19a668b5a1eadab82ff070b5a260478ac6ddad3a5b62295095174a8d26398"},
-    {file = "shapely-2.0.4.tar.gz", hash = "sha256:5dc736127fac70009b8d309a0eeb74f3e08979e530cf7017f2f507ef62e6cfb8"},
+    {file = "shapely-2.0.5-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:89d34787c44f77a7d37d55ae821f3a784fa33592b9d217a45053a93ade899375"},
+    {file = "shapely-2.0.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:798090b426142df2c5258779c1d8d5734ec6942f778dab6c6c30cfe7f3bf64ff"},
+    {file = "shapely-2.0.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:45211276900c4790d6bfc6105cbf1030742da67594ea4161a9ce6812a6721e68"},
+    {file = "shapely-2.0.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2e119444bc27ca33e786772b81760f2028d930ac55dafe9bc50ef538b794a8e1"},
+    {file = "shapely-2.0.5-cp310-cp310-win32.whl", hash = "sha256:9a4492a2b2ccbeaebf181e7310d2dfff4fdd505aef59d6cb0f217607cb042fb3"},
+    {file = "shapely-2.0.5-cp310-cp310-win_amd64.whl", hash = "sha256:1e5cb5ee72f1bc7ace737c9ecd30dc174a5295fae412972d3879bac2e82c8fae"},
+    {file = "shapely-2.0.5-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:5bbfb048a74cf273db9091ff3155d373020852805a37dfc846ab71dde4be93ec"},
+    {file = "shapely-2.0.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:93be600cbe2fbaa86c8eb70656369f2f7104cd231f0d6585c7d0aa555d6878b8"},
+    {file = "shapely-2.0.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0f8e71bb9a46814019f6644c4e2560a09d44b80100e46e371578f35eaaa9da1c"},
+    {file = "shapely-2.0.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d5251c28a29012e92de01d2e84f11637eb1d48184ee8f22e2df6c8c578d26760"},
+    {file = "shapely-2.0.5-cp311-cp311-win32.whl", hash = "sha256:35110e80070d664781ec7955c7de557456b25727a0257b354830abb759bf8311"},
+    {file = "shapely-2.0.5-cp311-cp311-win_amd64.whl", hash = "sha256:6c6b78c0007a34ce7144f98b7418800e0a6a5d9a762f2244b00ea560525290c9"},
+    {file = "shapely-2.0.5-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:03bd7b5fa5deb44795cc0a503999d10ae9d8a22df54ae8d4a4cd2e8a93466195"},
+    {file = "shapely-2.0.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:2ff9521991ed9e201c2e923da014e766c1aa04771bc93e6fe97c27dcf0d40ace"},
+    {file = "shapely-2.0.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1b65365cfbf657604e50d15161ffcc68de5cdb22a601bbf7823540ab4918a98d"},
+    {file = "shapely-2.0.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:21f64e647a025b61b19585d2247137b3a38a35314ea68c66aaf507a1c03ef6fe"},
+    {file = "shapely-2.0.5-cp312-cp312-win32.whl", hash = "sha256:3ac7dc1350700c139c956b03d9c3df49a5b34aaf91d024d1510a09717ea39199"},
+    {file = "shapely-2.0.5-cp312-cp312-win_amd64.whl", hash = "sha256:30e8737983c9d954cd17feb49eb169f02f1da49e24e5171122cf2c2b62d65c95"},
+    {file = "shapely-2.0.5-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:ff7731fea5face9ec08a861ed351734a79475631b7540ceb0b66fb9732a5f529"},
+    {file = "shapely-2.0.5-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ff9e520af0c5a578e174bca3c18713cd47a6c6a15b6cf1f50ac17dc8bb8db6a2"},
+    {file = "shapely-2.0.5-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:49b299b91557b04acb75e9732645428470825061f871a2edc36b9417d66c1fc5"},
+    {file = "shapely-2.0.5-cp37-cp37m-win32.whl", hash = "sha256:b5870633f8e684bf6d1ae4df527ddcb6f3895f7b12bced5c13266ac04f47d231"},
+    {file = "shapely-2.0.5-cp37-cp37m-win_amd64.whl", hash = "sha256:401cb794c5067598f50518e5a997e270cd7642c4992645479b915c503866abed"},
+    {file = "shapely-2.0.5-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:e91ee179af539100eb520281ba5394919067c6b51824e6ab132ad4b3b3e76dd0"},
+    {file = "shapely-2.0.5-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:8af6f7260f809c0862741ad08b1b89cb60c130ae30efab62320bbf4ee9cc71fa"},
+    {file = "shapely-2.0.5-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f5456dd522800306ba3faef77c5ba847ec30a0bd73ab087a25e0acdd4db2514f"},
+    {file = "shapely-2.0.5-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b714a840402cde66fd7b663bb08cacb7211fa4412ea2a209688f671e0d0631fd"},
+    {file = "shapely-2.0.5-cp38-cp38-win32.whl", hash = "sha256:7e8cf5c252fac1ea51b3162be2ec3faddedc82c256a1160fc0e8ddbec81b06d2"},
+    {file = "shapely-2.0.5-cp38-cp38-win_amd64.whl", hash = "sha256:4461509afdb15051e73ab178fae79974387f39c47ab635a7330d7fee02c68a3f"},
+    {file = "shapely-2.0.5-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7545a39c55cad1562be302d74c74586f79e07b592df8ada56b79a209731c0219"},
+    {file = "shapely-2.0.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:4c83a36f12ec8dee2066946d98d4d841ab6512a6ed7eb742e026a64854019b5f"},
+    {file = "shapely-2.0.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:89e640c2cd37378480caf2eeda9a51be64201f01f786d127e78eaeff091ec897"},
+    {file = "shapely-2.0.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:06efe39beafde3a18a21dde169d32f315c57da962826a6d7d22630025200c5e6"},
+    {file = "shapely-2.0.5-cp39-cp39-win32.whl", hash = "sha256:8203a8b2d44dcb366becbc8c3d553670320e4acf0616c39e218c9561dd738d92"},
+    {file = "shapely-2.0.5-cp39-cp39-win_amd64.whl", hash = "sha256:7fed9dbfbcfec2682d9a047b9699db8dcc890dfca857ecba872c42185fc9e64e"},
+    {file = "shapely-2.0.5.tar.gz", hash = "sha256:bff2366bc786bfa6cb353d6b47d0443c570c32776612e527ee47b6df63fcfe32"},
 ]
 
 [package.dependencies]
@@ -3935,17 +3910,20 @@ files = [
 
 [[package]]
 name = "sympy"
-version = "1.12.1"
+version = "1.13.0"
 description = "Computer algebra system (CAS) in Python"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "sympy-1.12.1-py3-none-any.whl", hash = "sha256:9b2cbc7f1a640289430e13d2a56f02f867a1da0190f2f99d8968c2f74da0e515"},
-    {file = "sympy-1.12.1.tar.gz", hash = "sha256:2877b03f998cd8c08f07cd0de5b767119cd3ef40d09f41c30d722f6686b0fb88"},
+    {file = "sympy-1.13.0-py3-none-any.whl", hash = "sha256:6b0b32a4673fb91bd3cac3b55406c8e01d53ae22780be467301cc452f6680c92"},
+    {file = "sympy-1.13.0.tar.gz", hash = "sha256:3b6af8f4d008b9a1a6a4268b335b984b23835f26d1d60b0526ebc71d48a25f57"},
 ]
 
 [package.dependencies]
-mpmath = ">=1.1.0,<1.4.0"
+mpmath = ">=1.1.0,<1.4"
+
+[package.extras]
+dev = ["hypothesis (>=6.70.0)", "pytest (>=7.1.0)"]
 
 [[package]]
 name = "tbb"
@@ -4499,7 +4477,7 @@ dev = ["debugpy", "pre-commit"]
 dora = ["gym-dora"]
 koch = ["dynamixel-sdk", "pynput"]
 pusht = ["gym-pusht"]
-test = ["pytest", "pytest-cov", "pytest-mock"]
+test = ["pytest", "pytest-cov"]
 umi = ["imagecodecs"]
 video-benchmark = ["pandas", "scikit-image"]
 xarm = ["gym-xarm"]
@@ -4507,4 +4485,4 @@ xarm = ["gym-xarm"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<3.13"
-content-hash = "2c59d869c6b1f2132070387f3d371b5b004765ae853501bbd522eb400738f2d0"
+content-hash = "dfe9c6a54e0382156e62e7bd2c7aab1be6372da76d30c61b06d27232276638cb"

pyproject.toml

@@ -38,12 +38,12 @@ einops = ">=0.8.0"
 pymunk = ">=6.6.0"
 zarr = ">=2.17.0"
 numba = ">=0.59.0"
-torch = "^2.2.1"
+torch = ">=2.2.1"
 opencv-python = ">=4.9.0"
-diffusers = "^0.27.2"
+diffusers = ">=0.27.2"
 torchvision = ">=0.17.1"
 h5py = ">=3.10.0"
-huggingface-hub = {extras = ["hf-transfer"], version = "^0.23.0"}
+huggingface-hub = {extras = ["hf-transfer"], version = ">=0.23.0"}
 gymnasium = ">=0.29.1"
 cmake = ">=3.29.0.1"
 gym-dora = { git = "https://github.com/dora-rs/dora-lerobot.git", subdirectory = "gym_dora", optional = true }
@@ -54,17 +54,16 @@ pre-commit = {version = ">=3.7.0", optional = true}
 debugpy = {version = ">=1.8.1", optional = true}
 pytest = {version = ">=8.1.0", optional = true}
 pytest-cov = {version = ">=5.0.0", optional = true}
-datasets = "^2.19.0"
+datasets = ">=2.19.0"
 imagecodecs = { version = ">=2024.1.1", optional = true }
 pyav = ">=12.0.5"
 moviepy = ">=1.0.3"
 rerun-sdk = ">=0.15.1"
 deepdiff = ">=7.0.1"
-scikit-image = {version = "^0.23.2", optional = true}
-pandas = {version = "^2.2.2", optional = true}
-pytest-mock = {version = "^3.14.0", optional = true}
-dynamixel-sdk = {version = "^3.7.31", optional = true}
-pynput = {version = "^1.7.7", optional = true}
+scikit-image = {version = ">=0.23.2", optional = true}
+pandas = {version = ">=2.2.2", optional = true}
+dynamixel-sdk = {version = ">=3.7.31", optional = true}
+pynput = {version = ">=1.7.7", optional = true}
 
 
 
@@ -74,7 +73,7 @@ pusht = ["gym-pusht"]
 xarm = ["gym-xarm"]
 aloha = ["gym-aloha"]
 dev = ["pre-commit", "debugpy"]
-test = ["pytest", "pytest-cov", "pytest-mock"]
+test = ["pytest", "pytest-cov"]
 umi = ["imagecodecs"]
 video_benchmark = ["scikit-image", "pandas"]
 koch = ["dynamixel-sdk", "pynput"]
@@ -110,7 +109,6 @@ exclude = [
 
 [tool.ruff.lint]
 select = ["E4", "E7", "E9", "F", "I", "N", "B", "C4", "SIM"]
-ignore-init-module-imports = true
 
 
 [build-system]

tests/data/lerobot/aloha_mobile_cabinet/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9f9347c8d9ac90ee44e6dd86f65043438168df6bbe4bab2d2b875e55ef7376ef
-size 1488
+oid sha256:7841afb9ef99c0601448c43a20c25eb029440c73816319c67c5d7e1c5cde2445
+size 136

tests/data/lerobot/aloha_mobile_cabinet/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_cabinet/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:02fc4ea25766269f65752a60b0594c43d799b0ae528cd773bf024b064b5aa329
+oid sha256:03508d82db846a804aef1a28aec3cb9572e3105b55a02b6ddbb09b2522d57b84
 size 4344

tests/data/lerobot/aloha_mobile_cabinet/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:55d7b1a06fe3e3051482752740074348bdb5fc98fb2e305b06d6203994117b27
+oid sha256:7009b3d2f14d6af497eeb32a52332e79cb9c07db24a6c2bbfbeffbaa8151dd69
 size 592448

tests/data/lerobot/aloha_mobile_cabinet/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:98329e4b40e9be0d63f7d36da9d86c44bbe7eeeb1b10d3ba973c923f3be70867
+oid sha256:34ece24fb6b302db0b68987858509f31713fb299faa9a9d34b8fd68f10bc3100
 size 247

tests/data/lerobot/aloha_mobile_cabinet/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:54e42cdfd016a0ced2ab1fe2966a8c15a2384e0dbe1a2fe87433a2d1b8209ac0
-size 5220057
+oid sha256:a70cc17019407cf6bee44fa2c78b4f29e48eb1696aa1a4ff4c048ba256574523
+size 6356921

tests/data/lerobot/aloha_mobile_cabinet/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:af1ded2a244cb47a96255b75f584a643edf6967e13bb5464b330ffdd9d7ad859
-size 5284692
+oid sha256:2b35992036e6dcee7d4df6d1675d55d1dd2d658b2d65442737e709895699a2f0
+size 5084448

tests/data/lerobot/aloha_mobile_cabinet/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:13d1bebabd79984fd6715971be758ef9a354495adea5e8d33f4e7904365e112b
-size 5258380
+oid sha256:3aa92e6b6bd0e39f6de530ea6a270671db7350cdc101c9d9030c775539c708c1
+size 5441406

tests/data/lerobot/aloha_mobile_chair/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f33bc6810f0b91817a42610364cb49ed1b99660f058f0f9407e6f5920d0aee02
-size 1008
+oid sha256:4ee862b1a6dc1d11df77c36c47ea00db88ad35a48e4d71c2940ad26b55fe2167
+size 136

tests/data/lerobot/aloha_mobile_chair/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_chair/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b58d6c89e936a781a307805ebecf0dd473fbc02d52a7094da62e54bffb9454a
+oid sha256:095c30bfe3c5da168c85aceef905e74e2142866332282965aa6812f6e6e48448
 size 4344

tests/data/lerobot/aloha_mobile_chair/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a08be578285cbe2d35b78f150d464ff3e10604a9865398c976983e0d711774f9
+oid sha256:98859f2d87e1a0abb9a930a82af623504b3efb26f70fe576f05bab7f19024427
 size 788528

tests/data/lerobot/aloha_mobile_chair/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34e36233477c8aa0b0840314ddace072062d4f486d06546bbd6550832c370065
+oid sha256:38cf4116a65cb92a5c43f9b9da7a7b81cfa9168b17605c8c456f7d3a3a23b77a
 size 247

tests/data/lerobot/aloha_mobile_chair/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:66e7349a4a82ca6042a7189608d01eb1cfa38d100d039b5445ae1a9e65d824ab
-size 14470946
+oid sha256:596dda720d378a44b6b61a6a72b44bec3e55e85198bca37f9dace6fe84af7ff0
+size 16062396

tests/data/lerobot/aloha_mobile_chair/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a2146f0c10c9f2611e57e617983aa4f91ad681b4fc50d91b992b97abd684f926
-size 11662185
+oid sha256:c614bbaf93d65354a82001b357682a0bd36f9603685f6c735c5e377b763d0bdb
+size 10317415

tests/data/lerobot/aloha_mobile_chair/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5affbaf1c48895ba3c626e0d8cf1309e5f4ec6bbaa135313096f52a22de66c05
-size 11410342
+oid sha256:868788028a38334b6b566cb17ffcc2ace2ec2b2b68ff2a58b6d29eb3c3e2ec1f
+size 9516445

tests/data/lerobot/aloha_mobile_elevator/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6c2b195ca91b88fd16422128d386d2cabd808a1862c6d127e6bf2e83e1fe819a
-size 448
+oid sha256:f365a02b052a2697b1558f4ab9b813f0d4ba46a5bc6ae3da30bbc4b135426aa6
+size 136

tests/data/lerobot/aloha_mobile_elevator/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_elevator/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b360b6b956d2adcb20589947c553348ef1eb6b70743c989dcbe95243d8592ce5
+oid sha256:5c96f47b569b7af82e05200213d733626664150aa7c5ae3298fd04a2138a2023
 size 4344

tests/data/lerobot/aloha_mobile_elevator/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3f5c3926b4d4da9271abefcdf6a8952bb1f13258a9c39fe0fd223f548dc89dcb
+oid sha256:75f53d221827f17cc2ded3908452e24331b39b79dc3a26f2b9d89a6e6894baab
 size 887728

tests/data/lerobot/aloha_mobile_elevator/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4993b05fb026619eec5eb70db8cadaa041ba4ab92d38b4a387167ace03b1018b
+oid sha256:d394d451929b805f2d94f9fc5b12d15c31cfc494df76d7d642b63378b8ba0131
 size 247

tests/data/lerobot/aloha_mobile_elevator/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bd25d17ef5b7500386761b5e32920879bbdcafe0e17a8a8845628525d861e644
-size 10231081
+oid sha256:73ddb898f83589b4bcabe978e46e75f20be215492f115bf6ebc98f1d01e1eff8
+size 9696507

tests/data/lerobot/aloha_mobile_elevator/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b557acbfeb0681c0a38e47263d945f6cd3a03461298d8b17209c81e3fd0aae8
-size 9701371
+oid sha256:d3d993977bee96882732d4a9c9d082c356fc9fcd8199c027b016207d60494c2f
+size 8957007

tests/data/lerobot/aloha_mobile_elevator/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da8f3b4f9f965da63819652b2c042d4cf7e07d14631113ea072087d56370310e
-size 10473741
+oid sha256:c9321627184c14af4a6ba64d02e86f7253bc1f563a3adef17036d68480d2bb3e
+size 9938178

tests/data/lerobot/aloha_mobile_shrimp/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a053506017d8a78cfd307b2912eeafa1ac1485a280cf90913985fcc40120b5ec
-size 416
+oid sha256:88346956fdf58f17dba7b08cc858364ed8278a7baa20febd9c68ae959d2c9c82
+size 136

tests/data/lerobot/aloha_mobile_shrimp/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_shrimp/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6d172d1bca02face22ceb4c21ea2b054cf3463025485dce64711b6f36b31f8a
+oid sha256:de80d5afc044be903a89ee08f30cfef5fb4c1e928d8ba8f4d81ea9d0bb4fb011
 size 4344

tests/data/lerobot/aloha_mobile_shrimp/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7e5ce817a2c188041f57f8d4c465dab3b9c3e4e1aeb7a9fb270230d1b36df530
+oid sha256:79c2a3da1024fa140d23e8438b2756d27cf5db65ac70d7ac4215260b55ca55f8
 size 1477064

tests/data/lerobot/aloha_mobile_shrimp/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4eb2dc373e4ea7d474742590f9073d66a773f6ab94b9e73a8673df19f93fae6d
+oid sha256:69435f30146a309c8d7d0eb01216555bf0547095db1fc9c20218d481d6fe62c8
 size 247

tests/data/lerobot/aloha_mobile_shrimp/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d2c55b146fabe78b18c8a28a7746ab56e1ee7a6918e9e3dad9bd196f97975895
-size 26158915
+oid sha256:3fc89b720dfb7511d5dd9eba31494cf720e6a89519067b7b5a4d65f0a539c811
+size 35137505

tests/data/lerobot/aloha_mobile_shrimp/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:71e1958d77f56843acf1ec48da4f04311a5836c87a0e77dbe26aa47c27c6347e
-size 18786848
+oid sha256:26b8d97a096aa8a1d686d86fc93bde1dcdd50a9dc273f49f3b6a700fe6610e88
+size 20387806

tests/data/lerobot/aloha_mobile_shrimp/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:20780718399b5759ff9a3a79824986310524793066198e3b9a307222f11a93df
-size 17769988
+oid sha256:72000be2803259f40da6d093279d17ed194ead3ebc508bf2d77cb463bcb67c4d
+size 17594265

tests/data/lerobot/aloha_mobile_wash_pan/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:279916f7689ae46af90e92a46eba9486a71fc762e3e2679ab5441eb37126827b
-size 928
+oid sha256:fb6de86fee6ff3cc5d61d591fe480a50feb289c05770e3f4b76e24138b571c65
+size 136

tests/data/lerobot/aloha_mobile_wash_pan/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_wash_pan/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7a7731051b521694b52b5631470720a7f05331915f4ac4e7f8cd83f9ff459bce
+oid sha256:d79027c2513c01a7e360f3177e62ab955e5d3f704f1e7127a6e1e852158ec42c
 size 4344

tests/data/lerobot/aloha_mobile_wash_pan/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:99608258e8c9fe5191f1a12edc29b47d307790104149dffb6d3046ddad6aeb1b
+oid sha256:0a2c1f98c816728136291fcb7530cd0ebcf4ea47b0f6750836da56b8324d64c1
 size 435600

tests/data/lerobot/aloha_mobile_wash_pan/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ae6735b7b394914824e974a7461019373a10f9e2d84ddf834bec8ea268d9ec1e
+oid sha256:921505133c62906bd53034a613a827996994875d84c8b26d69d188df9a7ffeba
 size 247

tests/data/lerobot/aloha_mobile_wash_pan/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:833e288c5fdacbbe10a5d048cb6f49fe1a396d91b2117b827e130ec11069256a
-size 8397615
+oid sha256:7e298db7d820e2ff9f0b9c250e800e8ada3521fdeae3c4127452dd62700e9ac8
+size 10980189

tests/data/lerobot/aloha_mobile_wash_pan/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2cb870acb4855fef70f19c5f632d94e4c25eef59eeea92f4b1167a44b1b36b33
-size 5912007
+oid sha256:29b46c2e823d62b1329b98a3d7efffd24fc6c904e9cea115e2f0adb1bb45db44
+size 7229025

tests/data/lerobot/aloha_mobile_wash_pan/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8be36298141b455ea51d17a78e4bbc6619639302139fe2db605bdfa3ff5e91bd
-size 4794018
+oid sha256:f34ddbd109b212260c758d54a0930f75a38666a178a0d26eeefa846cfeac86c0
+size 5944469

tests/data/lerobot/aloha_mobile_wipe_wine/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:348d0ee38a71929b2017d540de870b9dff6d79efdd0cbc5352fa9697e350134a
-size 928
+oid sha256:1386f9030607facefe56f429c93e50df0e22017914ce3f21ab67edc87b936d9d
+size 136

tests/data/lerobot/aloha_mobile_wipe_wine/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_mobile_wipe_wine/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c5c2996f58d5277fa19cf56ec143334fbee940d1de37530452496a6f0aa11f88
+oid sha256:7ffb173891cebb47a4d24d051f5fdd2ec44493d0a1a48d11f4d1410049aadd5b
 size 4344

tests/data/lerobot/aloha_mobile_wipe_wine/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da3a8efea9ba60d1fdd209d45a3387df22a09f7c156904ecb03f10456736fb74
+oid sha256:ae1760af2d3bf13c6e868643f203e76e1faf81a237715f72f2b81c3199e95e96
 size 514056

tests/data/lerobot/aloha_mobile_wipe_wine/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b7111ff1ef5c4d6a2990f5f39f42398f061da8c4e81adf46b9d9150ec2feeaf
+oid sha256:505a42c408d56c8a7d3e2367280b41e27667b58334f32e84c937c44c38217bd6
 size 247

tests/data/lerobot/aloha_mobile_wipe_wine/videos/observation.images.cam_high_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2ac8c2755d940534042595ecad33ebea358974ec67bc041c8675e53b7d2272ff
-size 9182551
+oid sha256:1489dac711fb99b192f064f9dbe56ed0e9e80fedc34da469e85acc7d5b4d75bf
+size 12316772

tests/data/lerobot/aloha_mobile_wipe_wine/videos/observation.images.cam_left_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b35aaa37e66dd5563d93e6059d5b645e112e020e03bd398f7098a5289970953a
-size 6378566
+oid sha256:20edc20184b5e4eb45194016fe7a0a5673665e3105286e0c6563767b5ff461f3
+size 6365474

tests/data/lerobot/aloha_mobile_wipe_wine/videos/observation.images.cam_right_wrist_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6138247ba7160a3de6c50111e6fcc5ae075044086d8527ae5d435b1f8a7c7a93
-size 6439183
+oid sha256:4ccdc96d9fe560a841e45e9fa636b69ef35b518271982339516517a4ae47d04f
+size 7449799

tests/data/lerobot/aloha_sim_insertion_human/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e11c127b1ab12f0761bce6651fa5a4882093924df230294f2f34309bc74b0707
-size 672
+oid sha256:9ee4f3c571ce6822e157e60133bee02245febee93eba5d35458d3c83345f7b87
+size 136

tests/data/lerobot/aloha_sim_insertion_human/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_sim_insertion_human/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5efab3606f50ee7ac8bb0c88cfeefd86bfd060dbb75d063e01d09456da020026
+oid sha256:b05f933aa67d559e44f062c8428b2f85ee7b49d3bf0e0302b9b83fb7d48ed0a3
 size 2904

tests/data/lerobot/aloha_sim_insertion_human/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:724a26cc4a3fc54ea5deb835816afa4a1c9712958ba402cd3067c22f4556a532
+oid sha256:8698f98e3fe36e321ba99a9b60facaab4abffb26916042b021adc1b41e8fb877
 size 100040

tests/data/lerobot/aloha_sim_insertion_human/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:71d6ba89bee5a4ee2761220452999e415bc838a44bebf1b5a2e4ba8622369798
+oid sha256:c0b18566cbf59e399ea40f1630df12ffbbb9f73bbc733d1d4eba62d675b1fda5
 size 247

tests/data/lerobot/aloha_sim_insertion_human/videos/observation.images.top_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12cd101db746125d40cf2e27c79340a3786c2906feca11a34e380c5d88280d36
-size 1329662
+oid sha256:5a57aade7d8510ef1cc8778f90cfa86749c95fa0c5a5e80cb166b2edd0f7189a
+size 1788513

tests/data/lerobot/aloha_sim_insertion_human_image/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:083db9efc5c9e3396c5e1159d020c2a3786f1f1a4b069719d327ed7fbc65c34d
-size 33
+oid sha256:e7ab5c2bd7d176d4d7902a600240318c2828b7d75f4a888d0887327e4eff089d
+size 65

tests/data/lerobot/aloha_sim_insertion_scripted/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4500f31e62f0928a837fa71783acacda0db516c7b00d0586a41ea5fd8fc5e772
-size 928
+oid sha256:4e910eac6a1c94f4c194b05e908dcc973dd4227b18eb80c374d7a1150f166c34
+size 136

tests/data/lerobot/aloha_sim_insertion_scripted/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
-size 33
+oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
+size 188

tests/data/lerobot/aloha_sim_insertion_scripted/meta_data/stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ae67bab70f3b313427fdcb447ed0a1a3d09581ff7ae8cc64ddd2243ef9ccb6c0
+oid sha256:a85e57264325cc0927450e30a85dd0eacb0a70ebdb00c4e2ac043a57f9c200e2
 size 2904

tests/data/lerobot/aloha_sim_insertion_scripted/train/data-00000-of-00001.arrow

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:277340fe2c2ca9f40a2cf27caf66dbb47089d690917a076e341d3be586b874d1
+oid sha256:171a9efc9c45601688821936ec9a1dcf91f16b1bbab4e8246f18b4d4cc6ac6ee
 size 80432

tests/data/lerobot/aloha_sim_insertion_scripted/train/state.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e48156ce4f71ac15d78732312fbc7e199f0ecdaac3604231e6be2e3e5b31a0ad
+oid sha256:5fd5fe80657788d044cdc8a1baf1456c7695cc951049347a469165002a83c6c7
 size 247

tests/data/lerobot/aloha_sim_insertion_scripted/videos/observation.images.top_episode_000000.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:292e6815ae7431d07ee75a5a770fdc8fd6fe8479eb104c33774ef0049f0dd768
-size 963206
+oid sha256:cb4810728c3d642326bf5fa2cd1632a60e68880faace4ec7368c6ee7992dabfb
+size 1297818

tests/data/lerobot/aloha_sim_insertion_scripted_image/meta_data/info.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:083db9efc5c9e3396c5e1159d020c2a3786f1f1a4b069719d327ed7fbc65c34d
-size 33
+oid sha256:e7ab5c2bd7d176d4d7902a600240318c2828b7d75f4a888d0887327e4eff089d
+size 65

tests/data/lerobot/aloha_sim_transfer_cube_human/meta_data/episode_data_index.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4500f31e62f0928a837fa71783acacda0db516c7b00d0586a41ea5fd8fc5e772
-size 928
+oid sha256:4e910eac6a1c94f4c194b05e908dcc973dd4227b18eb80c374d7a1150f166c34
+size 136