add

* chore(docstrings): fixing deprecated `root` argument docstrings in LeRobotDataset class

* chore(draccus): updating draccus CLI help

* chore(revert): reverting changes in lerobot_dataset_viz.py

---------

Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/workflows/nightly.yml

@@ -188,7 +188,7 @@ jobs:
       - name: Verify GPU availability
         run: |
           nvidia-smi
-          python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
+          python -c "import torch; print(f'PyTorch version: {torch.__version__}'); print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
 
       - name: Run multi-GPU training tests
       # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup

AI_POLICY.md

@@ -0,0 +1,25 @@
+# AI Usage Policy
+
+The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
+
+- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
+- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
+- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
+
+Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
+
+## Remember the Human Maintainers
+
+Please remember that LeRobot is maintained by a dedicated team of humans.
+
+Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
+
+Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
+
+## AI is Welcome Here
+
+LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
+
+Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
+
+We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
 
 ## Ways to Contribute
 

docs/source/_toctree.yml

@@ -19,6 +19,8 @@
     title: Multi GPU training
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
+  - local: rename_map
+    title: Using Rename Map and Empty Cameras
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3

docs/source/rename_map.mdx

@@ -0,0 +1,145 @@
+# Understanding the Rename Map and Empty Cameras
+
+When you train or evaluate a robot policy, your **dataset** or **environment** hands you observations under one set of keys (e.g. `observation.images.front`, `observation.images.eagle`), while your **policy** was built to expect another (e.g. `observation.images.image`, `observation.images.image2`). The rename map is how you bridge that gap without changing the policy or the data source.
+
+This guide explains why it exists, how to use it in training and evaluation, and when to use **empty cameras** so you can fine-tune multi-camera policies on datasets that have fewer views.
+
+---
+
+## Why observation keys don’t always match
+
+Policies have a fixed set of **input feature names** (often coming from a pretrained config). For example:
+
+- **XVLA-base** expects three image keys: `observation.images.image`, `observation.images.image2`, `observation.images.image3`.
+- **pi0-fast-libero** might expect `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb`.
+
+Your dataset or sim might use completely different names: `observation.images.front`, `observation.images.eagle`, `observation.images.glove` (e.g. [svla_so100_sorting](https://huggingface.co/datasets/lerobot/svla_so100_sorting)). Or your eval env (e.g. LIBERO) might return `observation.images.image` and `observation.images.image2`.
+
+Rather than renaming columns in the dataset or editing the policy code, LeRobot lets you pass a **rename map**: a dictionary that says “when you see this key in the data, treat it as this key for the policy.” Renaming is applied in the preprocessing pipeline so the policy always receives the keys it expects.
+
+---
+
+## How the rename map works
+
+The rename map is a dictionary:
+
+- **Keys** = observation keys as produced by your **dataset** (training) or **environment** (evaluation).
+- **Values** = the observation keys your **policy** expects.
+
+Only keys listed in the map are renamed; everything else is left as-is. Under the hood, the [RenameObservationsProcessorStep](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/rename_processor.py) runs in the preprocessor and rewrites observation keys (and keeps normalization stats aligned) so the batch matches the policy’s `input_features`.
+
+You can use the same idea for **training** (dataset → policy) and **evaluation** (env → policy).
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/jadechoghari/images/resolve/main/rename-map.png"
+    alt="Rename map: mapping dataset or environment observation keys to policy input keys"
+    style="max-width: 100%; height: auto;"
+  />
+</p>
+
+---
+
+## Option 1: Use a rename map (recommended)
+
+You pass the mapping on the command line so dataset/env keys are renamed to what the policy expects. No need to change the policy repo or the data.
+
+### Training example: XVLA on a dataset with different camera names
+
+Suppose you fine-tune [lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base) on a dataset whose images are stored under `observation.images.front`, `observation.images.eagle`, and `observation.images.glove`. XVLA expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`. Map the dataset keys to the policy keys:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/xvla_training \
+  --job_name=xvla_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.repo_id="HF_USER/xvla-your-robot" \
+  --policy.dtype=bfloat16 \
+  --policy.action_mode=auto \
+  --steps=20000 \
+  --policy.device=cuda \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true \
+  --rename_map='{"observation.images.front": "observation.images.image", "observation.images.eagle": "observation.images.image2", "observation.images.glove": "observation.images.image3"}'
+```
+
+Order of entries in the map doesn’t matter; each dataset key is renamed to the corresponding policy key.
+
+### Evaluation example: Policy trained on different camera names than the env
+
+You trained (or downloaded) a policy that expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb` (e.g. [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)), but your evaluation environment (e.g. LIBERO) returns `observation.images.image` and `observation.images.image2`. Tell the eval script how to rename env keys to policy keys:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/pi0fast-libero \
+  --env.type=libero \
+  ... \
+  --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
+```
+
+So: **key = what the env gives, value = what the policy expects.** Same convention as in training.
+
+---
+
+## Option 2: Change the policy config (no rename map)
+
+If you prefer not to pass a rename map every time, you can **edit the policy’s `config.json`** so that its expected observation keys match your dataset or environment. For example, change the policy’s visual input keys to `observation.images.front`, `observation.images.eagle`, `observation.images.glove` to match your dataset, or to `observation.images.image` / `observation.images.image2` to match LIBERO.
+
+- **Training:** If the dataset’s camera keys match the (modified) policy config, you don’t need a rename map.
+- **Evaluation:** If the env’s keys match the (modified) policy config, you don’t need a rename map for eval either.
+
+The tradeoff: you’re changing the policy repo or your local checkpoint. That’s fine if you’re only ever using that one dataset or env; a rename map keeps the same policy usable across multiple data sources without touching the config.
+
+---
+
+## When you have fewer cameras than the policy expects: empty cameras
+
+Some policies (e.g. XVLA) are built for a fixed number of image inputs (e.g. three). Your dataset might only have **two** cameras. You still want to fine-tune without changing the model architecture.
+
+LeRobot supports this with **empty cameras**: the config declares extra “slots” that the policy expects, but the dataset (or env) does not provide. Those slots are filled with placeholder keys and typically zero or masked inputs so the policy can run with fewer real views.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/jadechoghari/images/resolve/main/empty_cam.png"
+    alt="Empty cameras: using placeholder slots when the dataset has fewer views than the policy expects"
+    style="max-width: 100%; height: auto;"
+  />
+</p>
+
+- In the policy config (e.g. [xvla-base config.json](https://huggingface.co/lerobot/xvla-base/blob/main/config.json)), `empty_cameras` is the number of these extra slots (default `0`).
+- For each slot, the config adds an observation key of the form:
+  `observation.images.empty_camera_0`, `observation.images.empty_camera_1`, …
+
+Example: XVLA-base has three visual inputs and `empty_cameras=0`. Your dataset has only two images. Set **`empty_cameras=1`**. Then:
+
+1. The config gains a third visual key: `observation.images.empty_camera_0`.
+2. You still use the rename map (or matching config keys) for the two real cameras.
+3. The third view is treated as “empty” (no corresponding dataset key); the policy ignores or masks it as needed.
+
+So you fine-tune on two observations only, and the third visual input is effectively unused. You do **not** need to add a fake third image to your dataset.
+
+---
+
+## Where the rename map is used in the codebase
+
+- **Training** ([`lerobot_train.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py)): `rename_map` is passed into `make_policy(..., rename_map=cfg.rename_map)` and into the preprocessor as `rename_observations_processor: {"rename_map": cfg.rename_map}`. Batches from the dataset are renamed before being fed to the policy.
+- **Evaluation** ([`lerobot_eval.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py)): Same idea—`rename_map` is passed to `make_policy` and to the preprocessor so env observations are renamed before the policy sees them.
+- **Processor** ([`rename_processor.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/rename_processor.py)): `RenameObservationsProcessorStep` does the actual key renaming and updates feature metadata so normalization stats stay consistent with the renamed keys.
+
+If you see a feature mismatch error (“Missing features” / “Extra features”), the error message suggests using `--rename_map` with a mapping from your data’s keys to the policy’s expected keys.
+
+---
+
+## Quick reference
+
+| Goal                                  | What to do                                                                                                 |
+| ------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
+| Dataset keys ≠ policy keys (training) | `--rename_map='{"dataset_key": "policy_key", ...}'`                                                        |
+| Env keys ≠ policy keys (eval)         | `--rename_map='{"env_key": "policy_key", ...}'`                                                            |
+| Fewer cameras than policy expects     | Set `empty_cameras` in the policy config (e.g. `1` when you have 2 real cameras and the policy expects 3). |
+| Avoid passing a rename map            | Edit the policy’s `config.json` so its observation keys match your dataset or env.                         |
+
+The rename map keeps your pipeline flexible: one policy, many data sources, no code changes—just a small dictionary on the command line or in your config.

examples/backward_compatibility/replay.py

@@ -57,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.4"
+version = "0.4.5"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -76,7 +76,7 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "torch==2.10.0",
     "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bump dependency
     "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
 

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/datasets/lerobot_dataset.py

@@ -664,11 +664,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for the README).
 
         Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
-                will be stored under root/repo_id.
-            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                set the HF_LEROBOT_HOME environment variable to point to a different location. Defaults to
-                '~/.cache/huggingface/lerobot'.
+            repo_id (str): This is the repo id that will be used to fetch the dataset.
+            root (Path | None, optional): Local directory where the dataset will be downloaded and
+                stored. If set, all dataset files will be stored directly under this path. If not set, the
+                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
+                HF_LEROBOT_HOME environment variable).
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
             image_transforms (Callable | None, optional): You can pass standard v2 image transforms from

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -55,10 +55,16 @@ class DiffusionConfig(PreTrainedConfig):
         normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
             a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
-            within the image size. If None, no cropping is done.
-        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
-            mode).
+        resize_shape: (H, W) shape to resize images to as a preprocessing step for the vision
+            backbone. If None, no resizing is done and the original image resolution is used.
+        crop_ratio: Ratio in (0, 1] used to derive the crop size from resize_shape
+            (crop_h = int(resize_shape[0] * crop_ratio), likewise for width).
+            Set to 1.0 to disable cropping. Only takes effect when resize_shape is not None.
+        crop_shape: (H, W) shape to crop images to. When resize_shape is set and crop_ratio < 1.0,
+            this is computed automatically. Can also be set directly for legacy configs that use
+            crop-only (without resize). If None and no derivation applies, no cropping is done.
+        crop_is_random: Whether the crop should be random at training time (it's always a center
+            crop in eval mode).
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
@@ -114,7 +120,9 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    crop_shape: tuple[int, int] | None = (84, 84)
+    resize_shape: tuple[int, int] | None = None
+    crop_ratio: float = 1.0
+    crop_shape: tuple[int, int] | None = None
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -175,6 +183,25 @@ class DiffusionConfig(PreTrainedConfig):
                 f"Got {self.noise_scheduler_type}."
             )
 
+        if self.resize_shape is not None and (
+            len(self.resize_shape) != 2 or any(d <= 0 for d in self.resize_shape)
+        ):
+            raise ValueError(f"`resize_shape` must be a pair of positive integers. Got {self.resize_shape}.")
+        if not (0 < self.crop_ratio <= 1.0):
+            raise ValueError(f"`crop_ratio` must be in (0, 1]. Got {self.crop_ratio}.")
+
+        if self.resize_shape is not None:
+            if self.crop_ratio < 1.0:
+                self.crop_shape = (
+                    int(self.resize_shape[0] * self.crop_ratio),
+                    int(self.resize_shape[1] * self.crop_ratio),
+                )
+            else:
+                # Explicitly disable cropping for resize+ratio path when crop_ratio == 1.0.
+                self.crop_shape = None
+        if self.crop_shape is not None and (self.crop_shape[0] <= 0 or self.crop_shape[1] <= 0):
+            raise ValueError(f"`crop_shape` must have positive dimensions. Got {self.crop_shape}.")
+
         # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
@@ -202,13 +229,12 @@ class DiffusionConfig(PreTrainedConfig):
         if len(self.image_features) == 0 and self.env_state_feature is None:
             raise ValueError("You must provide at least one image or the environment state among the inputs.")
 
-        if self.crop_shape is not None:
+        if self.resize_shape is None and self.crop_shape is not None:
             for key, image_ft in self.image_features.items():
                 if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
                     raise ValueError(
-                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for "
-                        f"`{key}`."
+                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
                     )
 
         # Check that all input images have the same shape.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -454,12 +454,18 @@ class DiffusionRgbEncoder(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if config.crop_shape is not None:
+        if config.resize_shape is not None:
+            self.resize = torchvision.transforms.Resize(config.resize_shape)
+        else:
+            self.resize = None
+
+        crop_shape = config.crop_shape
+        if crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
             if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
@@ -485,13 +491,16 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
-        # The dummy input should take the number of image channels from `config.image_features` and it should
-        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
-        # height and width from `config.image_features`.
+        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
 
         # Note: we have a check in the config class to make sure all images have the same shape.
         images_shape = next(iter(config.image_features.values())).shape
-        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
+        if config.crop_shape is not None:
+            dummy_shape_h_w = config.crop_shape
+        elif config.resize_shape is not None:
+            dummy_shape_h_w = config.resize_shape
+        else:
+            dummy_shape_h_w = images_shape[1:]
         dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
         feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
 
@@ -507,7 +516,10 @@ class DiffusionRgbEncoder(nn.Module):
         Returns:
             (B, D) image feature.
         """
-        # Preprocess: maybe crop (if it was set up in the __init__).
+        # Preprocess: resize if configured, then crop if configured.
+
+        if self.resize is not None:
+            x = self.resize(x)
         if self.do_crop:
             if self.training:  # noqa: SIM108
                 x = self.maybe_random_crop(x)

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -106,6 +106,9 @@ class SmolVLAConfig(PreTrainedConfig):
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 
+    compile_model: bool = False  # Whether to use torch.compile for model optimization
+    compile_mode: str = "max-autotune"  # Torch compile mode
+
     def __post_init__(self):
         super().__post_init__()
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -593,6 +593,12 @@ class VLAFlowMatching(nn.Module):
         self.prefix_length = self.config.prefix_length
         self.rtc_processor = rtc_processor
 
+        # Compile model if requested
+        if config.compile_model:
+            torch.set_float32_matmul_precision("high")
+            self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
+            self.forward = torch.compile(self.forward, mode=config.compile_mode)
+
     def _rtc_enabled(self):
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 

src/lerobot/scripts/lerobot_record.py

@@ -155,7 +155,7 @@ class DatasetRecordConfig:
     repo_id: str
     # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
     single_task: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second.
     fps: int = 30

src/lerobot/scripts/lerobot_replay.py

@@ -80,7 +80,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

src/lerobot/scripts/lerobot_train.py

@@ -380,10 +380,10 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         "dataloading_s": AverageMeter("data_s", ":.3f"),
     }
 
-    # Use effective batch size for proper epoch calculation in distributed training
+    # Keep global batch size for logging; MetricsTracker handles world size internally.
     effective_batch_size = cfg.batch_size * accelerator.num_processes
     train_tracker = MetricsTracker(
-        effective_batch_size,
+        cfg.batch_size,
         dataset.num_frames,
         dataset.num_episodes,
         train_metrics,

src/lerobot/utils/logging_utils.py

@@ -104,9 +104,10 @@ class MetricsTracker:
         self.metrics = metrics
 
         self.steps = initial_step
+        world_size = accelerator.num_processes if accelerator else 1
         # A sample is an (observation,action) pair, where observation and action
         # can be on multiple timestamps. In a batch, we have `batch_size` number of samples.
-        self.samples = self.steps * self._batch_size
+        self.samples = self.steps * self._batch_size * world_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
@@ -132,7 +133,8 @@ class MetricsTracker:
         Updates metrics that depend on 'step' for one step.
         """
         self.steps += 1
-        self.samples += self._batch_size * (self.accelerator.num_processes if self.accelerator else 1)
+        world_size = self.accelerator.num_processes if self.accelerator else 1
+        self.samples += self._batch_size * world_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
 

tests/artifacts/policies/pusht_diffusion_/actions.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:19eaaa85f66ba4aa6388dbb83819ffad6ea4363247208f871a8dc385689f6fc8
+oid sha256:54aecbc1af72a4cd5e9261492f5e7601890517516257aacdf2a0ffb3ce281f1b
 size 992

tests/artifacts/policies/pusht_diffusion_/grad_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:227296eaeeb54acdc3dae2eb8af3d4d08fb87e245337624447140b1e91cfd002
+oid sha256:88a9c3775a2aa1e90a08850521970070a4fcf0f6b82aab43cd8ccc5cf77e0013
 size 47424

tests/artifacts/policies/pusht_diffusion_/output_dict.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:271b00cb2f0cd5fd26b1d53463638e3d1a6e92692ec625fcffb420ca190869e5
+oid sha256:91a2635e05a75fe187a5081504c5f35ce3417378813fa2deaf9ca4e8200e1819
 size 68

tests/artifacts/policies/pusht_diffusion_/param_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:778fddbbaa64248cee35cb377c02cc2b6076f7ce5855146de677128900617ddf
+oid sha256:645bff922ac7bea63ad018ebf77c303c0e4cd2c1c0dc5ef3192865281bef3dc6
 size 47424

tests/utils/test_logging_utils.py

@@ -24,6 +24,11 @@ def mock_metrics():
     return {"loss": AverageMeter("loss", ":.3f"), "accuracy": AverageMeter("accuracy", ":.2f")}
 
 
+class MockAccelerator:
+    def __init__(self, num_processes: int):
+        self.num_processes = num_processes
+
+
 def test_average_meter_initialization():
     meter = AverageMeter("loss", ":.2f")
     assert meter.name == "loss"
@@ -82,6 +87,37 @@ def test_metrics_tracker_step(mock_metrics):
     assert tracker.epochs == tracker.samples / 1000
 
 
+def test_metrics_tracker_initialization_with_accelerator(mock_metrics):
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=mock_metrics,
+        initial_step=10,
+        accelerator=MockAccelerator(num_processes=2),
+    )
+    assert tracker.steps == 10
+    assert tracker.samples == 10 * 32 * 2
+    assert tracker.episodes == tracker.samples / (1000 / 50)
+    assert tracker.epochs == tracker.samples / 1000
+
+
+def test_metrics_tracker_step_with_accelerator(mock_metrics):
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=mock_metrics,
+        initial_step=5,
+        accelerator=MockAccelerator(num_processes=2),
+    )
+    tracker.step()
+    assert tracker.steps == 6
+    assert tracker.samples == (5 * 32 * 2) + (32 * 2)
+    assert tracker.episodes == tracker.samples / (1000 / 50)
+    assert tracker.epochs == tracker.samples / 1000
+
+
 def test_metrics_tracker_getattr(mock_metrics):
     tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics=mock_metrics)
     assert tracker.loss == mock_metrics["loss"]