license + peft-dep + init groot + flat import layering utils dataset

* fix(one shot load): adding metadata loading when reading from a dataset after writing

* refactor(one shot load): move metadata reload to ensure_readable() on LeRobotDatasetMetadata

Move the metadata reload from DatasetReader.load_and_activate() to a new
public ensure_readable() method on LeRobotDatasetMetadata, called from
LeRobotDataset._ensure_reader(). This places lifecycle management in the
right layer: metadata owns its readiness check, the dataset orchestrates
the write-to-read transition, and the reader stays clean.

Also adds a regression test using delta_timestamps to exercise the
meta.episodes access path in the create -> write -> finalize -> read flow.

Co-authored-by: Steven Palma <imstevenpmwork@users.noreply.github.com>

---------

Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Steven Palma <imstevenpmwork@users.noreply.github.com>

.github/workflows/fast_tests.yml

@@ -12,7 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# This workflow handles fast testing.
+# This workflow validates each optional-dependency tier in isolation.
+# Each tier installs a different extra and runs the full test suite.
+# Tests that require an extra not installed in the current tier are
+# skipped automatically via pytest.importorskip guards.
 name: Fast Tests
 
 on:
@@ -54,8 +57,9 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  # This job runs pytests with the default dependencies.
-  # It runs everytime we commit to a PR or push to main
+  # This job runs pytests in isolated dependency tiers.
+  # Each tier installs a different extra and runs the full suite;
+  # tests gated behind other extras skip automatically.
   fast-pytest-tests:
     name: Fast Pytest Tests
     runs-on: ubuntu-latest
@@ -89,8 +93,9 @@ jobs:
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
-      - name: Install lerobot with test extras
-        run: uv sync --locked --extra "test"
+      # ── Tier 1: Base ──────────────────────────────────────
+      - name: "Tier 1 — Install: base"
+        run: uv sync --locked --extra test
 
       - name: Login to Hugging Face
         if: env.HF_USER_TOKEN != ''
@@ -98,5 +103,26 @@ jobs:
           uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
           uv run hf auth whoami
 
-      - name: Run pytest
+      - name: "Tier 1 — Test: base"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 2: Dataset ──────────────────────────────────
+      - name: "Tier 2 — Install: dataset"
+        run: uv sync --locked --extra test --extra dataset
+
+      - name: "Tier 2 — Test: dataset"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 3: Hardware ─────────────────────────────────
+      - name: "Tier 3 — Install: hardware"
+        run: uv sync --locked --extra test --extra hardware
+
+      - name: "Tier 3 — Test: hardware"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 4: Viz ──────────────────────────────────────
+      - name: "Tier 4 — Install: viz"
+        run: uv sync --locked --extra test --extra viz
+
+      - name: "Tier 4 — Test: viz"
         run: uv run pytest tests -vv --maxfail=10

docs/source/adding_benchmarks.mdx

@@ -216,7 +216,7 @@ class MyBenchmarkEnvConfig(EnvConfig):
 
     def get_env_processors(self):
         """Override if your benchmark needs observation/action transforms."""
-        from lerobot.processor.pipeline import PolicyProcessorPipeline
+        from lerobot.processor import PolicyProcessorPipeline
         from lerobot.processor.env_processor import MyBenchmarkProcessorStep
         return (
             PolicyProcessorPipeline(steps=[MyBenchmarkProcessorStep()]),

docs/source/async.mdx

@@ -170,7 +170,7 @@ python -m lerobot.async_inference.robot_client \
 ```python
 import threading
 from lerobot.robots.so_follower import SO100FollowerConfig
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.robot_client import RobotClient
 from lerobot.async_inference.helpers import visualize_action_queue_size

docs/source/backwardcomp.mdx

@@ -41,7 +41,7 @@ The script:
 
 ```python
 # New usage pattern (after migration)
-from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies import make_policy, make_pre_post_processors
 
 # Load model and processors separately
 policy = make_policy(config, ds_meta=dataset.meta)

docs/source/bring_your_own_policies.mdx

@@ -47,9 +47,9 @@ Here is a template to get you started, customize the parameters and methods as n
 ```python
 # configuration_my_custom_policy.py
 from dataclasses import dataclass, field
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.optim.optimizers import AdamWConfig
-from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.optim import AdamWConfig
+from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
 
 @PreTrainedConfig.register_subclass("my_custom_policy")
 @dataclass
@@ -120,7 +120,7 @@ import torch
 import torch.nn as nn
 from typing import Any
 
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
 from .configuration_my_custom_policy import MyCustomPolicyConfig
 

docs/source/cameras.mdx

@@ -79,9 +79,8 @@ The following examples show how to use the camera API to configure and capture f
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.cameras.opencv.camera_opencv import OpenCVCamera
-from lerobot.cameras.configs import ColorMode, Cv2Rotation
+from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
+from lerobot.cameras import ColorMode, Cv2Rotation
 
 # Construct an `OpenCVCameraConfig` with your desired FPS, resolution, color mode, and rotation.
 config = OpenCVCameraConfig(
@@ -126,9 +125,8 @@ with OpenCVCamera(config) as camera:
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig
-from lerobot.cameras.realsense.camera_realsense import RealSenseCamera
-from lerobot.cameras.configs import ColorMode, Cv2Rotation
+from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
+from lerobot.cameras import ColorMode, Cv2Rotation
 
 # Create a `RealSenseCameraConfig` specifying your camera’s serial number and enabling depth.
 config = RealSenseCameraConfig(

docs/source/dataset_subtask.mdx

@@ -95,7 +95,7 @@ After completing your annotation:
 When you load a dataset with subtask annotations, the subtask information is automatically available:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 # Load a dataset with subtask annotations
 dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
@@ -133,11 +133,10 @@ if has_subtasks:
 The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
 
 ```python
-from lerobot.processor.tokenizer_processor import TokenizerProcessor
-from lerobot.processor.pipeline import ProcessorPipeline
+from lerobot.processor import TokenizerProcessorStep
 
-# Create a tokenizer processor
-tokenizer_processor = TokenizerProcessor(
+# Create a tokenizer processor step
+tokenizer_processor = TokenizerProcessorStep(
     tokenizer_name_or_path="google/paligemma-3b-pt-224",
     padding="max_length",
     max_length=64,
@@ -158,7 +157,7 @@ When subtasks are available in the batch, the tokenizer processor adds:
 
 ```python
 import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
 
@@ -182,7 +181,7 @@ for batch in dataloader:
 Try loading a dataset with subtask annotations:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 # Example dataset with subtask annotations
 dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")

docs/source/earthrover_mini_plus.mdx

@@ -66,10 +66,10 @@ The SDK gives you:
 
 Follow our [Installation Guide](./installation) to install LeRobot.
 
-In addition to the base installation, install the EarthRover Mini dependencies:
+In addition to the base installation, install the EarthRover Mini with hardware dependencies:
 
 ```bash
-pip install -e .
+pip install -e ".[hardware]"
 ```
 
 ## How It Works

docs/source/env_processor.mdx

@@ -173,8 +173,8 @@ observation = {
 The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:
 
 ```python
-from lerobot.envs.factory import make_env_pre_post_processors
-from lerobot.envs.configs import LiberoEnv, PushtEnv
+from lerobot.envs import make_env_pre_post_processors, PushtEnv
+from lerobot.envs.configs import LiberoEnv
 
 # For LIBERO: Returns LiberoProcessorStep in preprocessor
 libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
@@ -257,7 +257,7 @@ def eval_main(cfg: EvalPipelineConfig):
 The `LiberoProcessorStep` demonstrates a real-world environment processor:
 
 ```python
-from lerobot.processor.pipeline import ObservationProcessorStep
+from lerobot.processor import ObservationProcessorStep
 
 @dataclass
 @ProcessorStepRegistry.register(name="libero_processor")

docs/source/envhub.mdx

@@ -34,7 +34,7 @@ Finally, your environment must implement the standard `gym.vector.VectorEnv` int
 Loading an environment from the Hub is as simple as:
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load a hub environment (requires explicit consent to run remote code)
 env = make_env("lerobot/cartpole-env", trust_remote_code=True)
@@ -191,7 +191,7 @@ api.upload_folder(
 ### Basic Usage
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env(
@@ -314,7 +314,7 @@ env = make_env("trusted-org/verified-env@a1b2c3d4", trust_remote_code=True)
 Here's a complete example using the reference CartPole environment:
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 import numpy as np
 
 # Load the environment

docs/source/envhub_isaaclab_arena.mdx

@@ -58,10 +58,10 @@ pip install -e .
 cd ..
 
 
-# 5. Install LeRobot
+# 5. Install LeRobot (evaluation extra for env/policy evaluation)
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
-pip install -e .
+pip install -e ".[evaluation]"
 cd ..
 
 
@@ -262,7 +262,7 @@ def main(cfg: EvalPipelineConfig):
     """Run random action rollout for IsaacLab Arena environment."""
     logging.info(pformat(asdict(cfg)))
 
-    from lerobot.envs.factory import make_env
+    from lerobot.envs import make_env
 
     env_dict = make_env(
         cfg.env,

docs/source/envhub_leisaac.mdx

@@ -74,7 +74,7 @@ EnvHub exposes every LeIsaac-supported task in a uniform interface. The examples
 # envhub_random_action.py
 
 import torch
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
@@ -142,7 +142,7 @@ from lerobot.teleoperators import (  # noqa: F401
 )
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 
 @dataclass
@@ -282,7 +282,7 @@ Note: when working with `bi_so101_fold_cloth`, call `initialize()` immediately a
 
 ```python
 import torch
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)

docs/source/il_robots.mdx

@@ -58,8 +58,8 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
-from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 
 robot_config = SO101FollowerConfig(
     port="/dev/tty.usbmodem58760431541",
@@ -116,9 +116,9 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.teleoperators.koch_leader import KochLeaderConfig, KochLeader
-from lerobot.robots.koch_follower import KochFollowerConfig, KochFollower
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
+from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig
 
 camera_config = {
     "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
@@ -195,13 +195,12 @@ lerobot-record \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so_leader.so100_leader import SO100Leader
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 from lerobot.scripts.lerobot_record import record_loop
@@ -410,9 +409,8 @@ lerobot-replay \
 ```python
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so_follower.so100_follower import SO100Follower
+from lerobot.datasets import LeRobotDataset
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
@@ -532,15 +530,14 @@ lerobot-record  \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so_follower.so100_follower import SO100Follower
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.policies.act import ACTPolicy
+from lerobot.policies import make_pre_post_processors
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

docs/source/installation.mdx

@@ -116,6 +116,8 @@ brew install ffmpeg
 
 ## Step 3: Install LeRobot 🤗
 
+The base `lerobot` install is intentionally **lightweight** — it includes only core ML dependencies (PyTorch, torchvision, numpy, opencv, einops, draccus, huggingface-hub, gymnasium, safetensors). Heavier dependencies are gated behind optional extras so you only install what you need.
+
 ### From Source
 
 First, clone the repository and navigate into the directory:
@@ -131,12 +133,16 @@ Then, install the library in editable mode. This is useful if you plan to contri
 <hfoptions id="install_lerobot_src">
 <hfoption id="conda">
 ```bash
-pip install -e .
+pip install -e ".[core_scripts]"  # For robot workflows (recording, replaying, calibrate)
+pip install -e ".[training]"      # For training policies
+pip install -e ".[all]"           # Everything (all policies, envs, hardware, dev tools)
 ```
 </hfoption>
 <hfoption id="uv">
 ```bash
-uv pip install -e .
+uv pip install -e ".[core_scripts]"  # For robot workflows (recording, replaying, calibrate)
+uv pip install -e ".[training]"      # For training policies
+uv pip install -e ".[all]"           # Everything (all policies, envs, hardware, dev tools)
 ```
 </hfoption>
 </hfoptions>
@@ -162,26 +168,48 @@ uv pip install lerobot
 </hfoptions>
 <!-- prettier-ignore-end -->
 
-_This installs only the default dependencies._
+_This installs only the core ML dependencies. You will need to add extras for most workflows._
 
-**Extra Features:**
-To install additional functionality, use one of the following (If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.):
+**Feature Extras:**
+LeRobot provides **feature-scoped extras** that map to common workflows. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
+
+| Extra      | What it adds                                | Typical use case                    |
+| ---------- | ------------------------------------------- | ----------------------------------- |
+| `dataset`  | `datasets`, `av`, `torchcodec`, `jsonlines` | Loading & creating datasets         |
+| `training` | `dataset` + `accelerate`, `wandb`           | Training policies                   |
+| `hardware` | `pynput`, `pyserial`, `deepdiff`            | Connecting to real robots           |
+| `viz`      | `rerun-sdk`                                 | Visualization during recording/eval |
+
+**Composite Extras** combine feature extras for common CLI scripts:
+
+| Extra          | Includes                       | Typical use case                                        |
+| -------------- | ------------------------------ | ------------------------------------------------------- |
+| `core_scripts` | `dataset` + `hardware` + `viz` | `lerobot-record`, `lerobot-replay`, `lerobot-calibrate` |
+| `evaluation`   | `av`                           | `lerobot-eval` (add policy + env extras as needed)      |
+| `dataset_viz`  | `dataset` + `viz`              | `lerobot-dataset-viz`, `lerobot-imgtransform-viz`       |
 
 ```bash
-pip install 'lerobot[all]'          # All available features
-pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
-pip install 'lerobot[feetech]'      # Feetech motor support
+pip install 'lerobot[core_scripts]'          # Record, replay, calibrate
+pip install 'lerobot[training]'              # Train policies
+pip install 'lerobot[core_scripts,training]' # Record + train
+pip install 'lerobot[all]'                   # Everything
 ```
 
-_Replace `[...]` with your desired features._
+**Policy, environment, and hardware extras** are still available for specific dependencies:
 
-**Available Tags:**
-For a full list of optional dependencies, see:
-https://pypi.org/project/lerobot/
+```bash
+pip install 'lerobot[pi]'             # Pi0/Pi0.5/Pi0-FAST policy deps
+pip install 'lerobot[smolvla]'        # SmolVLA policy deps
+pip install 'lerobot[diffusion]'      # Diffusion policy deps (diffusers)
+pip install 'lerobot[aloha,pusht]'    # Simulation environments
+pip install 'lerobot[feetech]'        # Feetech motor support
+```
+
+_Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._
 
 ### Troubleshooting
 
-If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
+If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
 To install these for Linux run:
 
 ```bash
@@ -196,8 +224,8 @@ LeRobot provides optional extras for specific functionalities. Multiple extras c
 
 ### Simulations
 
-Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
-Example:
+Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht)).
+These automatically include the `dataset` extra.
 
 ```bash
 pip install -e ".[aloha]" # or "[pusht]" for example
@@ -213,7 +241,7 @@ pip install -e ".[feetech]" # or "[dynamixel]" for example
 
 ### Experiment Tracking
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
+Weights and Biases is included in the `training` extra. To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with:
 
 ```bash
 wandb login

docs/source/introduction_processors.mdx

@@ -19,10 +19,10 @@ This means that your favorite policy can be used like this:
 ```python
 import torch
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import make_pre_post_processors
 from lerobot.policies.your_policy import YourPolicy
-from lerobot.processor.pipeline import RobotProcessorPipeline, PolicyProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
 dataset = LeRobotDataset("hf_user/dataset", episodes=[0])
 sample = dataset[10]
 
@@ -260,7 +260,7 @@ Since processor pipelines can add new features (like velocity fields), change te
 These functions work together by starting with robot hardware specifications (`create_initial_features()`) then simulating the entire pipeline transformation (`aggregate_pipeline_dataset_features()`) to compute the final feature dictionary that gets passed to `LeRobotDataset.create()`, ensuring perfect alignment between what processors output and what datasets expect to store.
 
 ```python
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+from lerobot.datasets import aggregate_pipeline_dataset_features
 
 # Start with robot's raw features
 initial_features = create_initial_features(

docs/source/lerobot-dataset-v3.mdx

@@ -89,7 +89,7 @@ A core v3 principle is **decoupling storage from the user API**: data is stored
 
 ```python
 import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 repo_id = "yaak-ai/L2D-v3"
 
@@ -135,7 +135,7 @@ for batch in data_loader:
 Use `StreamingLeRobotDataset` to iterate directly from the Hub without local copies. This allows to stream large datasets without the need to downloading them onto disk or loading them onto memory, and is a key feature of the new dataset format.
 
 ```python
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
+from lerobot.datasets import StreamingLeRobotDataset
 
 repo_id = "yaak-ai/L2D-v3"
 dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
@@ -167,8 +167,8 @@ Currently, transforms are applied during **training time only**, not during reco
 Use the `image_transforms` parameter when loading a dataset for training:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
 
 # Option 1: Use default transform configuration (disabled by default)
 transforms_config = ImageTransformsConfig(
@@ -290,7 +290,7 @@ python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DAT
 When creating or recording datasets, you **must** call `dataset.finalize()` to properly close parquet writers. See the [PR #1903](https://github.com/huggingface/lerobot/pull/1903) for more details.
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 # Create dataset and record episodes
 dataset = LeRobotDataset.create(...)

docs/source/multi_gpu_training.mdx

@@ -4,10 +4,10 @@ This guide shows you how to train policies on multiple GPUs using [Hugging Face
 
 ## Installation
 
-First, ensure you have accelerate installed:
+`accelerate` is included in the `training` extra. Install it with:
 
 ```bash
-pip install accelerate
+pip install 'lerobot[training]'
 ```
 
 ## Training with Multiple GPUs

docs/source/phone_teleop.mdx

@@ -45,7 +45,8 @@ Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`.
 Teleoperation example:
 
 ```python
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 
 teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 teleop_device = Phone(teleop_config)

docs/source/pi0.mdx

@@ -110,8 +110,7 @@ lerobot-edit-dataset \
 Or equivalently in Python:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_dataset")
 recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])

docs/source/pi05.mdx

@@ -116,8 +116,7 @@ lerobot-edit-dataset \
 Or equivalently in Python:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_dataset")
 recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])

docs/source/policy_pi0_README.md

@@ -60,11 +60,10 @@ When `use_relative_actions=true`, the training script automatically:
 ### Recomputing stats for an existing dataset
 
 If you want to precompute relative action stats offline, use `recompute_stats` from
-`lerobot.datasets.dataset_tools`:
+`lerobot.datasets`:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_org/your_dataset")
 dataset = recompute_stats(

docs/source/rtc.mdx

@@ -39,9 +39,8 @@ The snippet below provides a simplified pseudo-example of how RTC operates with
 
 ```python
 from lerobot.policies.pi0 import PI0Policy, PI0Config
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.configs import RTCAttentionSchedule
+from lerobot.policies.rtc import RTCConfig, ActionQueue
 
 # Load Pi0 with RTC enabled
 policy_cfg = PI0Config()

docs/source/xvla.mdx

@@ -418,7 +418,7 @@ Create a custom preprocessing pipeline for your environment:
 
 ```python
 from lerobot.processor import PolicyProcessorPipeline
-from lerobot.policies.xvla.processor_xvla import (
+from lerobot.policies.xvla import (
     XVLAImageToFloatProcessorStep,
     XVLAImageNetNormalizeProcessorStep,
     XVLAAddDomainIdProcessorStep,

examples/backward_compatibility/replay.py

@@ -35,7 +35,7 @@ from pprint import pformat
 
 import draccus
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,

examples/dataset/load_lerobot_dataset.py

@@ -31,17 +31,11 @@ from pprint import pprint
 import torch
 from huggingface_hub import HfApi
 
-import lerobot
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
 
 
 def main():
-    # We ported a number of existing datasets ourselves, use this to see the list:
-    print("List of available datasets:")
-    pprint(lerobot.available_datasets)
-
-    # You can also browse through the datasets created/ported by the community on the hub using the hub api:
+    # Browse datasets created/ported by the community on the hub using the hub api:
     hub_api = HfApi()
     repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
     pprint(repo_ids)

examples/dataset/slurm_compute_rabc.py

@@ -231,7 +231,7 @@ class AggregateProgress(PipelineStep):
         import pyarrow as pa
         import pyarrow.parquet as pq
 
-        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.datasets import LeRobotDataset
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/dataset/use_dataset_image_transforms.py

@@ -26,8 +26,8 @@ import torch
 from torchvision.transforms import v2
 from torchvision.transforms.functional import to_pil_image
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
 
 
 def save_image(tensor, filename):

examples/dataset/use_dataset_tools.py

@@ -29,7 +29,8 @@ Usage:
 
 import numpy as np
 
-from lerobot.datasets.dataset_tools import (
+from lerobot.datasets import (
+    LeRobotDataset,
     add_features,
     delete_episodes,
     merge_datasets,
@@ -37,7 +38,6 @@ from lerobot.datasets.dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 
 def main():

examples/hil/hil_data_collection.py

@@ -112,17 +112,18 @@ from hil_utils import (
     teleop_smooth_move_to,
 )
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.configs import parser
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.datasets.feature_utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
-from lerobot.datasets.image_writer import safe_stop_image_writer
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.video_utils import VideoEncodingManager
-from lerobot.policies.factory import get_policy_class, make_policy, make_pre_post_processors
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.common.control_utils import is_headless, predict_action
+from lerobot.configs import PreTrainedConfig, parser
+from lerobot.datasets import (
+    LeRobotDataset,
+    VideoEncodingManager,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+    safe_stop_image_writer,
+)
+from lerobot.policies import PreTrainedPolicy, get_policy_class, make_policy, make_pre_post_processors
 from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
@@ -131,18 +132,18 @@ from lerobot.processor import (
     RelativeActionsProcessorStep,
     TransitionKey,
     create_transition,
+    rename_stats,
+    to_relative_actions,
 )
-from lerobot.processor.relative_action_processor import to_relative_actions
-from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import Robot, RobotConfig, make_robot_from_config
-from lerobot.robots.bi_openarm_follower.config_bi_openarm_follower import BiOpenArmFollowerConfig
-from lerobot.robots.so_follower.config_so_follower import SOFollowerRobotConfig  # noqa: F401
+from lerobot.robots.bi_openarm_follower import BiOpenArmFollowerConfig
+from lerobot.robots.so_follower import SOFollowerRobotConfig  # noqa: F401
 from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
-from lerobot.teleoperators.openarm_mini.config_openarm_mini import OpenArmMiniConfig  # noqa: F401
-from lerobot.teleoperators.so_leader.config_so_leader import SOLeaderTeleopConfig  # noqa: F401
+from lerobot.teleoperators.openarm_mini import OpenArmMiniConfig  # noqa: F401
+from lerobot.teleoperators.so_leader import SOLeaderTeleopConfig  # noqa: F401
+from lerobot.utils import get_safe_torch_device
 from lerobot.utils.constants import ACTION, OBS_STATE, OBS_STR
-from lerobot.utils.control_utils import is_headless, predict_action
-from lerobot.utils.device_utils import get_safe_torch_device
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging, log_say
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/hil/hil_utils.py

@@ -19,13 +19,12 @@ import time
 from dataclasses import dataclass, field
 from pathlib import Path
 
+from lerobot.common.control_utils import is_headless
 from lerobot.processor import (
     IdentityProcessorStep,
     RobotAction,
     RobotObservation,
     RobotProcessorPipeline,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -33,7 +32,6 @@ from lerobot.processor.converters import (
 )
 from lerobot.robots import Robot
 from lerobot.teleoperators import Teleoperator
-from lerobot.utils.control_utils import is_headless
 from lerobot.utils.robot_utils import precise_sleep
 
 logger = logging.getLogger(__name__)

examples/lekiwi/evaluate.py

@@ -14,15 +14,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/lekiwi/record.py

@@ -14,16 +14,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset
 from lerobot.processor import make_default_processors
-from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
-from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/lekiwi/replay.py

@@ -16,9 +16,8 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
-from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.datasets import LeRobotDataset
+from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/phone_to_so100/evaluate.py

@@ -14,19 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.configs import FeatureType, PolicyFeature
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -39,7 +36,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/phone_to_so100/record.py

@@ -14,13 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -35,11 +34,11 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/phone_to_so100/replay.py

@@ -16,10 +16,10 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )

examples/phone_to_so100/teleoperate.py

@@ -16,8 +16,8 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )
@@ -28,9 +28,9 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     GripperVelocityToJoint,
     InverseKinematicsEEToJoints,
 )
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/port_datasets/port_droid.py

@@ -22,8 +22,7 @@ from pathlib import Path
 import numpy as np
 import tensorflow_datasets as tfds
 
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
 
 DROID_SHARDS = 2048

examples/port_datasets/slurm_aggregate_shards.py

@@ -36,7 +36,7 @@ class AggregateDatasets(PipelineStep):
     def run(self, data=None, rank: int = 0, world_size: int = 1):
         import logging
 
-        from lerobot.datasets.aggregate import aggregate_datasets
+        from lerobot.datasets import aggregate_datasets
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/port_datasets/slurm_upload.py

@@ -26,8 +26,7 @@ from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
 from port_droid import DROID_SHARDS
 
-from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.utils import create_lerobot_dataset_card
+from lerobot.datasets import CODEBASE_VERSION, LeRobotDatasetMetadata, create_lerobot_dataset_card
 from lerobot.utils.utils import init_logging
 
 
@@ -155,7 +154,7 @@ class UploadDataset(PipelineStep):
         from datasets.utils.tqdm import disable_progress_bars
         from huggingface_hub import CommitOperationAdd, preupload_lfs_files
 
-        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+        from lerobot.datasets import LeRobotDatasetMetadata
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/rtc/eval_dataset.py

@@ -109,15 +109,10 @@ except ImportError:
     MATPLOTLIB_AVAILABLE = False
     plt = None
 
-from lerobot.configs import parser
-from lerobot.configs.default import DatasetConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.factory import resolve_delta_timestamps
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.configs import DatasetConfig, PreTrainedConfig, RTCAttentionSchedule, parser
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata, resolve_delta_timestamps
+from lerobot.policies import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc import RTCConfig
 from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import init_logging

examples/rtc/eval_with_real_robot.py

@@ -101,26 +101,21 @@ from threading import Event, Lock, Thread
 import torch
 from torch import Tensor
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig  # noqa: F401
-from lerobot.configs import parser
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
-from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.cameras.zmq import ZMQCameraConfig  # noqa: F401
+from lerobot.configs import PreTrainedConfig, RTCAttentionSchedule, parser
+from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.processor import (
     NormalizerProcessorStep,
     RelativeActionsProcessorStep,
     TransitionKey,
     create_transition,
-)
-from lerobot.processor.factory import (
     make_default_robot_action_processor,
     make_default_robot_observation_processor,
+    to_relative_actions,
 )
-from lerobot.processor.relative_action_processor import to_relative_actions
 from lerobot.rl.process import ProcessSignalHandler
 from lerobot.robots import (  # noqa: F401
     Robot,
@@ -133,6 +128,7 @@ from lerobot.robots import (  # noqa: F401
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import init_logging
 

examples/so100_to_so100_EE/evaluate.py

@@ -14,19 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.configs import FeatureType, PolicyFeature
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -39,7 +36,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/record.py

@@ -15,13 +15,12 @@
 # limitations under the License.
 
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -36,7 +35,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/replay.py

@@ -17,10 +17,10 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )

examples/so100_to_so100_EE/teleoperate.py

@@ -17,8 +17,8 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     robot_action_to_transition,
     transition_to_robot_action,

examples/training/train_policy.py

@@ -18,13 +18,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionConfig, DiffusionPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def main():

examples/training/train_with_streaming.py

@@ -19,14 +19,12 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDatasetMetadata, StreamingLeRobotDataset
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTConfig, ACTPolicy
 from lerobot.utils.constants import ACTION
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def main():

examples/tutorial/act/act_training_example.py

@@ -4,13 +4,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTConfig, ACTPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:

examples/tutorial/act/act_using_example.py

@@ -1,9 +1,9 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 

examples/tutorial/async-inf/robot_client.py

@@ -3,7 +3,7 @@ import threading
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.helpers import visualize_action_queue_size
 from lerobot.async_inference.robot_client import RobotClient
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.robots.so_follower import SO100FollowerConfig
 
 

examples/tutorial/diffusion/diffusion_training_example.py

@@ -4,13 +4,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionConfig, DiffusionPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:

examples/tutorial/diffusion/diffusion_using_example.py

@@ -1,9 +1,9 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 

examples/tutorial/pi0/using_pi0_example.py

@@ -1,11 +1,11 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.pi0.modeling_pi0 import PI0Policy
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.pi0 import PI0Policy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/tutorial/rl/hilserl_example.py

@@ -6,17 +6,17 @@ from queue import Empty, Full
 import torch
 import torch.optim as optim
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.policies import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
+from lerobot.teleoperators import TeleopEvents
 from lerobot.teleoperators.so_leader import SO100LeaderConfig
-from lerobot.teleoperators.utils import TeleopEvents
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 LOG_EVERY = 10
 SEND_EVERY = 10

examples/tutorial/rl/reward_classifier_example.py

@@ -1,8 +1,7 @@
 import torch
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import make_policy, make_pre_post_processors
-from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import RewardClassifierConfig, make_policy, make_pre_post_processors
 
 
 def main():

examples/tutorial/smolvla/using_smolvla_example.py

@@ -1,11 +1,11 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.smolvla import SmolVLAPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

pyproject.toml

@@ -58,45 +58,74 @@ classifiers = [
 keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artificial intelligence"]
 
 dependencies = [
-
-    # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
-    "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
-    "accelerate>=1.10.0,<2.0.0",
-
-    # Core dependencies
-    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
-    "setuptools>=71.0.0,<81.0.0",
-    "cmake>=3.29.0.1,<4.2.0",
-    "packaging>=24.2,<26.0",
-
+    # Core ML
     "torch>=2.7,<2.11.0",
-    "torchcodec>=0.3.0,<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')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
     "torchvision>=0.22.0,<0.26.0",
-
-    "einops>=0.8.0,<0.9.0",
+    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "opencv-python-headless>=4.9.0,<4.14.0",
-    "av>=15.0.0,<16.0.0",
-    "jsonlines>=4.0.0,<5.0.0",
-    "pynput>=1.7.8,<1.9.0",
-    "pyserial>=3.5,<4.0",
+    "Pillow>=10.0.0,<13.0.0",
+    "einops>=0.8.0,<0.9.0",
 
-    "wandb>=0.24.0,<0.25.0",
+    # Config & Hub
     "draccus==0.10.0", # TODO: Relax version constraint
-    "gymnasium>=1.1.1,<2.0.0",
-    "rerun-sdk>=0.24.0,<0.27.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
+    "requests>=2.32.0,<3.0.0",
 
-    # Support dependencies
-    "deepdiff>=7.0.1,<9.0.0",
-    "imageio[ffmpeg]>=2.34.0,<3.0.0",
+    # Environments
+    # NOTE: gymnasium is used in lerobot.envs (lerobot-train, lerobot-eval), policies/factory,
+    # and robots/unitree. Moving it to an optional extra would require import guards across many
+    # tightly-coupled modules. Candidate for a future refactor to decouple envs from the core.
+    "gymnasium>=1.1.1,<2.0.0",
+
+    # Serialization & checkpointing
+    "safetensors>=0.4.3,<1.0.0",
+
+    # Lightweight utilities
+    "packaging>=24.2,<26.0",
     "termcolor>=2.4.0,<4.0.0",
+    "tqdm>=4.66.0,<5.0.0",
+
+    # Build tools (required by opencv-python-headless on some platforms)
+    "cmake>=3.29.0.1,<4.2.0",
+    "setuptools>=71.0.0,<81.0.0",
 ]
 
 # Optional dependencies
 [project.optional-dependencies]
 
+# ── Feature-scoped extras ──────────────────────────────────
+dataset = [
+    "datasets>=4.0.0,<5.0.0",
+    "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
+    "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
+    "lerobot[av-dep]",
+    "torchcodec>=0.3.0,<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')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
+    "jsonlines>=4.0.0,<5.0.0",
+]
+training = [
+    "lerobot[dataset]",
+    "accelerate>=1.10.0,<2.0.0",
+    "wandb>=0.24.0,<0.25.0",
+]
+hardware = [
+    "pynput>=1.7.8,<1.9.0",
+    "pyserial>=3.5,<4.0",
+    "deepdiff>=7.0.1,<9.0.0",
+]
+viz = [
+    "rerun-sdk>=0.24.0,<0.27.0",
+]
+# ── User-facing composite extras (map to CLI scripts) ─────
+# lerobot-record, lerobot-replay, lerobot-calibrate, lerobot-teleoperate, etc.
+core_scripts = ["lerobot[dataset]", "lerobot[hardware]", "lerobot[viz]"]
+# lerobot-eval -- base evaluation framework. You also need the policy's extra (e.g., lerobot[pi])
+# and the environment's extra (e.g., lerobot[pusht]) if evaluating in simulation.
+evaluation = ["lerobot[av-dep]"]
+# lerobot-dataset-viz, lerobot-imgtransform-viz
+dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
+
 # Common
+av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.9.17"]
 transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
@@ -104,6 +133,7 @@ grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
 scipy-dep = ["scipy>=1.14.0,<2.0.0"]
+diffusers-dep = ["diffusers>=0.27.2,<0.36.0"]
 qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
 matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
@@ -136,28 +166,28 @@ intelrealsense = [
 phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
+diffusion = ["lerobot[diffusers-dep]"]
 wallx = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "lerobot[peft-dep]",
     "lerobot[scipy-dep]",
     "torchdiffeq>=0.2.4,<0.3.0",
     "lerobot[qwen-vl-utils-dep]",
 ]
 pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
-smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
-multi_task_dit = ["lerobot[transformers-dep]"]
+smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0"]
+multi_task_dit = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]"]
 groot = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "lerobot[peft-dep]",
+    "lerobot[diffusers-dep]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
-    "safetensors>=0.4.3,<1.0.0",
-    "Pillow>=10.0.0,<13.0.0",
     "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
@@ -166,31 +196,42 @@ async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1"]
 test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
 # NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
-pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
-metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
+aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
+libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
 
 # All
 all = [
+    # Feature-scoped extras
+    "lerobot[dataset]",
+    "lerobot[training]",
+    "lerobot[hardware]",
+    "lerobot[viz]",
     # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
     # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
     # helps pip's resolver converge by constraining scipy early, before it encounters
     # the loose scipy requirements from transitive deps like dm-control and metaworld.
     "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
+    "lerobot[feetech]",
+    "lerobot[damiao]",
+    "lerobot[robstride]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
     "lerobot[lekiwi]",
+    "lerobot[openarms]",
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
+    "lerobot[diffusion]",
+    "lerobot[multi_task_dit]",
     "lerobot[wallx]",
     "lerobot[pi]",
     "lerobot[smolvla]",
@@ -267,7 +308,9 @@ ignore = [
 ]
 
 [tool.ruff.lint.per-file-ignores]
-"__init__.py" = ["F401", "F403"]
+"__init__.py" = ["F401", "F403", "E402"]
+# E402: conditional-import guards (TYPE_CHECKING / is_package_available) must precede the imports they protect
+"src/lerobot/scripts/convert_dataset_v21_to_v30.py" = ["E402"]
 "src/lerobot/policies/wall_x/**" = ["N801", "N812", "SIM102", "SIM108", "SIM210", "SIM211", "B006", "B007", "SIM118"] # Supprese these as they are coming from original Qwen2_5_vl code TODO(pepijn): refactor original
 
 [tool.ruff.lint.isort]

src/lerobot/__init__.py

@@ -13,188 +13,39 @@
 # 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.
+
 """
-This file contains lists of available environments, dataset and policies to reflect the current state of LeRobot library.
-We do not want to import all the dependencies, but instead we keep it lightweight to ensure fast access to these variables.
+LeRobot -- PyTorch library for real-world robotics.
 
-Example:
-    ```python
-        import lerobot
-        print(lerobot.available_envs)
-        print(lerobot.available_tasks_per_env)
-        print(lerobot.available_datasets)
-        print(lerobot.available_datasets_per_env)
-        print(lerobot.available_real_world_datasets)
-        print(lerobot.available_policies)
-        print(lerobot.available_policies_per_env)
-        print(lerobot.available_robots)
-        print(lerobot.available_cameras)
-        print(lerobot.available_motors)
-    ```
+Provides datasets, pretrained policies, and tools for training, evaluation,
+data collection, and robot control. Integrates with Hugging Face Hub for
+model and dataset sharing.
 
-When implementing a new dataset loadable with LeRobotDataset follow these steps:
-- Update `available_datasets_per_env` in `lerobot/__init__.py`
+The base install is intentionally lightweight. Feature-specific dependencies
+are gated behind optional extras::
 
-When implementing a new environment (e.g. `gym_aloha`), follow these steps:
-- Update `available_tasks_per_env` and `available_datasets_per_env` in `lerobot/__init__.py`
-
-When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps:
-- Update `available_policies` and `available_policies_per_env`, in `lerobot/__init__.py`
-- Set the required `name` class attribute.
-- Update variables in `tests/test_available.py` by importing your new Policy class
+    pip install 'lerobot[dataset]'       # dataset loading & creation
+    pip install 'lerobot[training]'      # training loop + wandb
+    pip install 'lerobot[hardware]'      # real robot control
+    pip install 'lerobot[core_scripts]'  # dataset + hardware + viz (record, replay, calibrate, etc.)
+    pip install 'lerobot[all]'           # everything
 """
 
-import itertools
+from lerobot.__version__ import __version__
 
-from lerobot.__version__ import __version__  # noqa: F401
-
-# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
-# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
-# a yaml file AND a environment name. The difference should be more obvious.
-available_tasks_per_env = {
-    "aloha": [
-        "AlohaInsertion-v0",
-        "AlohaTransferCube-v0",
+# Maps optional extras to the CLI entry-points they unlock.
+available_extras: dict[str, list[str]] = {
+    "dataset": ["lerobot-dataset-viz", "lerobot-imgtransform-viz", "lerobot-edit-dataset"],
+    "training": ["lerobot-train"],
+    "hardware": [
+        "lerobot-calibrate",
+        "lerobot-find-port",
+        "lerobot-find-cameras",
+        "lerobot-find-joint-limits",
+        "lerobot-setup-motors",
     ],
-    "pusht": ["PushT-v0"],
-}
-available_envs = list(available_tasks_per_env.keys())
-
-available_datasets_per_env = {
-    "aloha": [
-        "lerobot/aloha_sim_insertion_human",
-        "lerobot/aloha_sim_insertion_scripted",
-        "lerobot/aloha_sim_transfer_cube_human",
-        "lerobot/aloha_sim_transfer_cube_scripted",
-        "lerobot/aloha_sim_insertion_human_image",
-        "lerobot/aloha_sim_insertion_scripted_image",
-        "lerobot/aloha_sim_transfer_cube_human_image",
-        "lerobot/aloha_sim_transfer_cube_scripted_image",
-    ],
-    # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
-    # coupled with tests.
-    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
+    "core_scripts": ["lerobot-record", "lerobot-replay", "lerobot-teleoperate"],
+    "evaluation": ["lerobot-eval"],
 }
 
-available_real_world_datasets = [
-    "lerobot/aloha_mobile_cabinet",
-    "lerobot/aloha_mobile_chair",
-    "lerobot/aloha_mobile_elevator",
-    "lerobot/aloha_mobile_shrimp",
-    "lerobot/aloha_mobile_wash_pan",
-    "lerobot/aloha_mobile_wipe_wine",
-    "lerobot/aloha_static_battery",
-    "lerobot/aloha_static_candy",
-    "lerobot/aloha_static_coffee",
-    "lerobot/aloha_static_coffee_new",
-    "lerobot/aloha_static_cups_open",
-    "lerobot/aloha_static_fork_pick_up",
-    "lerobot/aloha_static_pingpong_test",
-    "lerobot/aloha_static_pro_pencil",
-    "lerobot/aloha_static_screw_driver",
-    "lerobot/aloha_static_tape",
-    "lerobot/aloha_static_thread_velcro",
-    "lerobot/aloha_static_towel",
-    "lerobot/aloha_static_vinh_cup",
-    "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",
-    "lerobot/nyu_rot_dataset",
-    "lerobot/utokyo_saytap",
-    "lerobot/imperialcollege_sawyer_wrist_cam",
-    "lerobot/utokyo_xarm_bimanual",
-    "lerobot/tokyo_u_lsmo",
-    "lerobot/utokyo_pr2_opening_fridge",
-    "lerobot/cmu_franka_exploration_dataset",
-    "lerobot/cmu_stretch",
-    "lerobot/asu_table_top",
-    "lerobot/utokyo_pr2_tabletop_manipulation",
-    "lerobot/utokyo_xarm_pick_and_place",
-    "lerobot/ucsd_kitchen_dataset",
-    "lerobot/austin_buds_dataset",
-    "lerobot/dlr_sara_grid_clamp",
-    "lerobot/conq_hose_manipulation",
-    "lerobot/columbia_cairlab_pusht_real",
-    "lerobot/dlr_sara_pour",
-    "lerobot/dlr_edan_shared_control",
-    "lerobot/ucsd_pick_and_place_dataset",
-    "lerobot/berkeley_cable_routing",
-    "lerobot/nyu_franka_play_dataset",
-    "lerobot/austin_sirius_dataset",
-    "lerobot/cmu_play_fusion",
-    "lerobot/berkeley_gnm_sac_son",
-    "lerobot/nyu_door_opening_surprising_effectiveness",
-    "lerobot/berkeley_fanuc_manipulation",
-    "lerobot/jaco_play",
-    "lerobot/viola",
-    "lerobot/kaist_nonprehensile",
-    "lerobot/berkeley_mvp",
-    "lerobot/uiuc_d3field",
-    "lerobot/berkeley_gnm_recon",
-    "lerobot/austin_sailor_dataset",
-    "lerobot/utaustin_mutex",
-    "lerobot/roboturk",
-    "lerobot/stanford_hydra_dataset",
-    "lerobot/berkeley_autolab_ur5",
-    "lerobot/stanford_robocook",
-    "lerobot/toto",
-    "lerobot/fmb",
-    "lerobot/droid_100",
-    "lerobot/berkeley_rpt",
-    "lerobot/stanford_kuka_multimodal_dataset",
-    "lerobot/iamlab_cmu_pickup_insert",
-    "lerobot/taco_play",
-    "lerobot/berkeley_gnm_cory_hall",
-    "lerobot/usc_cloth_sim",
-]
-
-available_datasets = sorted(
-    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
-)
-
-# lists all available policies from `lerobot/policies`
-available_policies = ["act", "diffusion", "tdmpc", "vqbet"]
-
-# lists all available robots from `lerobot/robots`
-available_robots = [
-    "koch",
-    "koch_bimanual",
-    "aloha",
-    "so100",
-    "so101",
-]
-
-# lists all available cameras from `lerobot/cameras`
-available_cameras = [
-    "opencv",
-    "intelrealsense",
-]
-
-# lists all available motors from `lerobot/motors`
-available_motors = [
-    "dynamixel",
-    "feetech",
-]
-
-# keys and values refer to yaml files
-available_policies_per_env = {
-    "aloha": ["act"],
-    "pusht": ["diffusion", "vqbet"],
-    "koch_real": ["act_koch_real"],
-    "aloha_real": ["act_aloha_real"],
-}
-
-env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]
-env_dataset_pairs = [
-    (env, dataset) for env, datasets in available_datasets_per_env.items() for dataset in datasets
-]
-env_dataset_policy_triplets = [
-    (env, dataset, policy)
-    for env, datasets in available_datasets_per_env.items()
-    for dataset in datasets
-    for policy in available_policies_per_env[env]
-]
+__all__ = ["__version__", "available_extras"]

src/lerobot/async_inference/__init__.py

@@ -0,0 +1,30 @@
+# 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.
+
+"""
+Async inference server/client.
+
+Requires: ``pip install 'lerobot[async]'``
+
+Available modules (import directly)::
+
+    from lerobot.async_inference.policy_server import ...
+    from lerobot.async_inference.robot_client import ...
+"""
+
+from lerobot.utils.import_utils import require_package
+
+require_package("grpcio", extra="async", import_name="grpc")
+
+__all__: list[str] = []

src/lerobot/async_inference/helpers.py

@@ -22,8 +22,7 @@ from typing import Any
 
 import torch
 
-from lerobot.configs.types import PolicyFeature
-from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.configs import PolicyFeature
 
 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
 from lerobot.policies import (  # noqa: F401
@@ -36,6 +35,7 @@ from lerobot.policies import (  # noqa: F401
 )
 from lerobot.robots.robot import Robot
 from lerobot.utils.constants import OBS_IMAGES, OBS_STATE, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.utils.utils import init_logging
 
 Action = torch.Tensor

src/lerobot/async_inference/policy_server.py

@@ -38,7 +38,7 @@ import draccus
 import grpc
 import torch
 
-from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
     services_pb2,  # type: ignore

src/lerobot/async_inference/robot_client.py

@@ -47,8 +47,8 @@ import draccus
 import grpc
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,

src/lerobot/cameras/__init__.py

@@ -15,3 +15,9 @@
 from .camera import Camera
 from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
 from .utils import make_cameras_from_configs
+
+# NOTE: Camera submodule configs and implementations (OpenCVCameraConfig, RealSenseCamera, etc.)
+# are intentionally NOT re-exported here to avoid pulling backend-specific dependencies.
+# Import from submodules: ``from lerobot.cameras.opencv import OpenCVCameraConfig``
+
+__all__ = ["Camera", "CameraConfig", "ColorMode", "Cv2Backends", "Cv2Rotation", "make_cameras_from_configs"]

src/lerobot/cameras/reachy2_camera/__init__.py

@@ -14,3 +14,5 @@
 
 from .configuration_reachy2_camera import Reachy2CameraConfig
 from .reachy2_camera import Reachy2Camera
+
+__all__ = ["Reachy2Camera", "Reachy2CameraConfig"]

src/lerobot/cameras/realsense/__init__.py

@@ -14,3 +14,5 @@
 
 from .camera_realsense import RealSenseCamera
 from .configuration_realsense import RealSenseCameraConfig
+
+__all__ = ["RealSenseCamera", "RealSenseCameraConfig"]

src/lerobot/cameras/zmq/image_server.py

@@ -31,8 +31,8 @@ import cv2
 import numpy as np
 import zmq
 
-from lerobot.cameras.configs import ColorMode
-from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
+from ..configs import ColorMode
+from ..opencv import OpenCVCamera, OpenCVCameraConfig
 
 logger = logging.getLogger(__name__)
 

src/lerobot/common/__init__.py

@@ -0,0 +1,30 @@
+# 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.
+
+"""
+Cross-cutting modules that bridge multiple lerobot packages.
+
+Unlike ``lerobot.utils`` (which must remain dependency-free), modules here
+are allowed to import from ``lerobot.policies``, ``lerobot.processor``,
+``lerobot.configs``, etc.  They are deliberately NOT re-exported from the
+top-level ``lerobot`` package.
+
+Available modules (import directly)::
+
+    from lerobot.common.control_utils import predict_action, ...
+    from lerobot.common.train_utils import save_checkpoint, ...
+    from lerobot.common.wandb_utils import WandBLogger, ...
+"""
+
+__all__: list[str] = []

src/lerobot/common/control_utils.py

@@ -12,26 +12,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import annotations
+
 ########################################################################################
 # Utilities
 ########################################################################################
-
-
 import logging
 import traceback
 from contextlib import nullcontext
 from copy import copy
 from functools import cache
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import numpy as np
 import torch
-from deepdiff import DeepDiff
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import DEFAULT_FEATURES
-from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.utils import prepare_observation_for_inference
+from lerobot.policies import PreTrainedPolicy, prepare_observation_for_inference
+
+if TYPE_CHECKING:
+    from lerobot.datasets import LeRobotDataset
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.robots import Robot
 from lerobot.types import PolicyAction
@@ -218,6 +217,13 @@ def sanity_check_dataset_robot_compatibility(
     Raises:
         ValueError: If any of the checked metadata fields do not match.
     """
+    from lerobot.utils.import_utils import require_package
+
+    require_package("deepdiff", extra="hardware")
+    from deepdiff import DeepDiff
+
+    from lerobot.utils.constants import DEFAULT_FEATURES
+
     fields = [
         ("robot_type", dataset.meta.robot_type, robot.robot_type),
         ("fps", dataset.fps, fps),

src/lerobot/common/train_utils.py

@@ -19,10 +19,13 @@ from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
 
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets.io_utils import load_json, write_json
-from lerobot.optim.optimizers import load_optimizer_state, save_optimizer_state
-from lerobot.optim.schedulers import load_scheduler_state, save_scheduler_state
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.optim import (
+    load_optimizer_state,
+    load_scheduler_state,
+    save_optimizer_state,
+    save_scheduler_state,
+)
+from lerobot.policies import PreTrainedPolicy
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.utils.constants import (
     CHECKPOINTS_DIR,
@@ -31,6 +34,7 @@ from lerobot.utils.constants import (
     TRAINING_STATE_DIR,
     TRAINING_STEP,
 )
+from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.random_utils import load_rng_state, save_rng_state
 
 

src/lerobot/configs/__init__.py

@@ -0,0 +1,47 @@
+# 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.
+
+"""
+Public API for lerobot configuration types and base config classes.
+
+NOTE: TrainPipelineConfig, EvalPipelineConfig, and TrainRLServerPipelineConfig
+are intentionally NOT re-exported here to avoid circular dependencies
+(they import lerobot.envs and lerobot.policies at module level).
+Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
+"""
+
+from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from .policies import PreTrainedConfig
+from .types import (
+    FeatureType,
+    NormalizationMode,
+    PipelineFeatureType,
+    PolicyFeature,
+    RTCAttentionSchedule,
+)
+
+__all__ = [
+    # Types
+    "FeatureType",
+    "NormalizationMode",
+    "PipelineFeatureType",
+    "PolicyFeature",
+    "RTCAttentionSchedule",
+    # Config classes
+    "DatasetConfig",
+    "EvalConfig",
+    "PeftConfig",
+    "PreTrainedConfig",
+    "WandBConfig",
+]

src/lerobot/configs/default.py

@@ -16,8 +16,8 @@
 
 from dataclasses import dataclass, field
 
-from lerobot.datasets.transforms import ImageTransformsConfig
-from lerobot.datasets.video_utils import get_safe_default_codec
+from lerobot.transforms import ImageTransformsConfig
+from lerobot.utils.import_utils import get_safe_default_codec
 
 
 @dataclass

src/lerobot/configs/eval.py

@@ -19,8 +19,9 @@ from pathlib import Path
 
 from lerobot import envs, policies  # noqa: F401
 from lerobot.configs import parser
-from lerobot.configs.default import EvalConfig
-from lerobot.configs.policies import PreTrainedConfig
+
+from .default import EvalConfig
+from .policies import PreTrainedConfig
 
 logger = getLogger(__name__)
 

src/lerobot/configs/policies.py

@@ -26,13 +26,13 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.optim.optimizers import OptimizerConfig
-from lerobot.optim.schedulers import LRSchedulerConfig
+from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.device_utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 from lerobot.utils.hub import HubMixin
 
+from .types import FeatureType, PolicyFeature
+
 T = TypeVar("T", bound="PreTrainedConfig")
 logger = getLogger(__name__)
 

src/lerobot/configs/train.py

@@ -24,12 +24,12 @@ from huggingface_hub.errors import HfHubHTTPError
 
 from lerobot import envs
 from lerobot.configs import parser
-from lerobot.configs.default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.optim import OptimizerConfig
-from lerobot.optim.schedulers import LRSchedulerConfig
+from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.hub import HubMixin
 
+from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from .policies import PreTrainedConfig
+
 TRAIN_CONFIG_NAME = "train_config.json"
 
 

src/lerobot/data_processing/__init__.py

@@ -11,3 +11,13 @@
 # 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.
+
+"""
+Data processing utilities (annotation tools, dataset transformations).
+
+Available sub-modules (import directly)::
+
+    from lerobot.data_processing.sarm_annotations import ...
+"""
+
+__all__: list[str] = []

src/lerobot/data_processing/sarm_annotations/__init__.py

@@ -11,3 +11,13 @@
 # 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.
+
+"""
+SARM subtask annotation tools.
+
+Available modules (import directly)::
+
+    from lerobot.data_processing.sarm_annotations.subtask_annotation import ...
+"""
+
+__all__: list[str] = []

src/lerobot/data_processing/sarm_annotations/subtask_annotation.py

@@ -76,7 +76,7 @@ import torch
 from pydantic import BaseModel, Field
 from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 
 # Pydantic Models for SARM Subtask Annotation
@@ -746,8 +746,7 @@ def save_annotations_to_dataset(
     dataset_path: Path, annotations: dict[int, SubtaskAnnotation], fps: int, prefix: str = "sparse"
 ):
     """Save annotations to LeRobot dataset parquet format."""
-    from lerobot.datasets.io_utils import load_episodes
-    from lerobot.datasets.utils import DEFAULT_EPISODES_PATH
+    from lerobot.datasets import DEFAULT_EPISODES_PATH, load_episodes
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:
@@ -841,7 +840,7 @@ def generate_auto_sparse_annotations(
 
 def load_annotations_from_dataset(dataset_path: Path, prefix: str = "sparse") -> dict[int, SubtaskAnnotation]:
     """Load annotations from LeRobot dataset parquet files."""
-    from lerobot.datasets.io_utils import load_episodes
+    from lerobot.datasets import load_episodes
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:

src/lerobot/datasets/__init__.py

@@ -15,19 +15,68 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.multi_dataset import MultiLeRobotDataset
-from lerobot.datasets.sampler import EpisodeAwareSampler
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig
+from lerobot.utils.import_utils import require_package
+
+require_package("datasets", extra="dataset")
+require_package("av", extra="dataset")
+
+from .aggregate import aggregate_datasets
+from .compute_stats import DEFAULT_QUANTILES, aggregate_stats, get_feature_stats
+from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from .dataset_tools import (
+    add_features,
+    convert_image_to_video_dataset,
+    delete_episodes,
+    merge_datasets,
+    modify_features,
+    modify_tasks,
+    recompute_stats,
+    remove_feature,
+    split_dataset,
+)
+from .factory import make_dataset, resolve_delta_timestamps
+from .image_writer import safe_stop_image_writer
+from .io_utils import load_episodes, write_stats
+from .lerobot_dataset import LeRobotDataset
+from .multi_dataset import MultiLeRobotDataset
+from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from .sampler import EpisodeAwareSampler
+from .streaming_dataset import StreamingLeRobotDataset
+from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
+from .video_utils import VideoEncodingManager
+
+# NOTE: Low-level I/O functions (cast_stats_to_numpy, get_parquet_file_size_in_mb, etc.)
+# and legacy migration constants are intentionally NOT re-exported here.
+# Import directly: ``from lerobot.datasets.io_utils import ...``
 
 __all__ = [
+    "CODEBASE_VERSION",
+    "DEFAULT_EPISODES_PATH",
+    "DEFAULT_QUANTILES",
     "EpisodeAwareSampler",
-    "ImageTransforms",
-    "ImageTransformsConfig",
     "LeRobotDataset",
     "LeRobotDatasetMetadata",
     "MultiLeRobotDataset",
     "StreamingLeRobotDataset",
+    "VideoEncodingManager",
+    "add_features",
+    "aggregate_datasets",
+    "aggregate_pipeline_dataset_features",
+    "aggregate_stats",
+    "convert_image_to_video_dataset",
+    "create_initial_features",
+    "create_lerobot_dataset_card",
+    "delete_episodes",
+    "get_feature_stats",
+    "load_episodes",
+    "make_dataset",
+    "merge_datasets",
+    "modify_features",
+    "modify_tasks",
+    "recompute_stats",
+    "remove_feature",
+    "resolve_delta_timestamps",
+    "safe_stop_image_writer",
+    "split_dataset",
+    "write_stats",
 ]

src/lerobot/datasets/aggregate.py

@@ -23,10 +23,10 @@ import datasets
 import pandas as pd
 import tqdm
 
-from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import get_hf_features_from_features
-from lerobot.datasets.io_utils import (
+from .compute_stats import aggregate_stats
+from .dataset_metadata import LeRobotDatasetMetadata
+from .feature_utils import get_hf_features_from_features
+from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
     to_parquet_with_hf_images,
@@ -34,7 +34,7 @@ from lerobot.datasets.io_utils import (
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.utils import (
+from .utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
@@ -43,7 +43,7 @@ from lerobot.datasets.utils import (
     DEFAULT_VIDEO_PATH,
     update_chunk_file_indices,
 )
-from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
+from .video_utils import concatenate_video_files, get_video_duration_in_s
 
 
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):

src/lerobot/datasets/compute_stats.py

@@ -19,9 +19,11 @@ import logging
 
 import numpy as np
 
-from lerobot.datasets.io_utils import load_image_as_numpy
+from lerobot.processor import RelativeActionsProcessorStep
 from lerobot.utils.constants import ACTION, OBS_STATE
 
+from .io_utils import load_image_as_numpy
+
 DEFAULT_QUANTILES = [0.01, 0.10, 0.50, 0.90, 0.99]
 
 
@@ -696,8 +698,6 @@ def compute_relative_action_stats(
         ValueError: If the dataset has fewer frames than ``chunk_size``.
         RuntimeError: If no valid (single-episode) chunks are found.
     """
-    from lerobot.processor.relative_action_processor import RelativeActionsProcessorStep
-
     if exclude_joints is None:
         exclude_joints = []
 

src/lerobot/datasets/dataset_metadata.py

@@ -23,9 +23,13 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download
 
-from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.feature_utils import _validate_feature_names, create_empty_dataset_info
-from lerobot.datasets.io_utils import (
+from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
+from lerobot.utils.feature_utils import _validate_feature_names
+from lerobot.utils.utils import flatten_dict
+
+from .compute_stats import aggregate_stats
+from .feature_utils import create_empty_dataset_info
+from .io_utils import (
     get_file_size_in_mb,
     load_episodes,
     load_info,
@@ -37,19 +41,16 @@ from lerobot.datasets.io_utils import (
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.utils import (
+from .utils import (
     DEFAULT_EPISODES_PATH,
-    DEFAULT_FEATURES,
     INFO_PATH,
     check_version_compatibility,
-    flatten_dict,
     get_safe_version,
     has_legacy_hub_download_metadata,
     is_valid_version,
     update_chunk_file_indices,
 )
-from lerobot.datasets.video_utils import get_video_info
-from lerobot.utils.constants import HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
+from .video_utils import get_video_info
 
 CODEBASE_VERSION = "v3.0"
 
@@ -180,6 +181,16 @@ class LeRobotDatasetMetadata:
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
+    def ensure_readable(self) -> None:
+        """Guarantee metadata is fully loaded for read operations.
+
+        Idempotent — when metadata is already in memory this is a single
+        ``is None`` check.  Call this before transitioning from write to
+        read mode on the same instance.
+        """
+        if self.episodes is None:
+            self._load_metadata()
+
     def _pull_from_repo(
         self,
         allow_patterns: list[str] | str | None = None,

src/lerobot/datasets/dataset_reader.py

@@ -21,17 +21,17 @@ from pathlib import Path
 import datasets
 import torch
 
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import (
+from .dataset_metadata import LeRobotDatasetMetadata
+from .feature_utils import (
     check_delta_timestamps,
     get_delta_indices,
     get_hf_features_from_features,
 )
-from lerobot.datasets.io_utils import (
+from .io_utils import (
     hf_transform_to_torch,
     load_nested_dataset,
 )
-from lerobot.datasets.video_utils import decode_video_frames
+from .video_utils import decode_video_frames
 
 
 class DatasetReader:

src/lerobot/datasets/dataset_tools.py

@@ -36,22 +36,25 @@ import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm
 
-from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.datasets.compute_stats import (
+from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
+from lerobot.utils.utils import flatten_dict
+
+from .aggregate import aggregate_datasets
+from .compute_stats import (
     aggregate_stats,
     compute_episode_stats,
     compute_relative_action_stats,
 )
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.io_utils import (
+from .dataset_metadata import LeRobotDatasetMetadata
+from .io_utils import (
     get_parquet_file_size_in_mb,
     load_episodes,
     write_info,
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import (
+from .lerobot_dataset import LeRobotDataset
+from .utils import (
     DATA_DIR,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -59,8 +62,7 @@ from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
     update_chunk_file_indices,
 )
-from lerobot.datasets.video_utils import encode_video_frames, get_video_info
-from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
+from .video_utils import encode_video_frames, get_video_info
 
 
 def _load_episode_with_stats(src_dataset: LeRobotDataset, episode_idx: int) -> dict:
@@ -829,8 +831,6 @@ def _copy_and_reindex_episodes_metadata(
         data_metadata: Dict mapping new episode index to its data file metadata
         video_metadata: Optional dict mapping new episode index to its video metadata
     """
-    from lerobot.datasets.utils import flatten_dict
-
     if src_dataset.meta.episodes is None:
         src_dataset.meta.episodes = load_episodes(src_dataset.meta.root)
 
@@ -922,8 +922,8 @@ def _write_parquet(df: pd.DataFrame, path: Path, meta: LeRobotDatasetMetadata) -
 
     This ensures images are properly embedded and the file can be loaded correctly by HF datasets.
     """
-    from lerobot.datasets.feature_utils import get_hf_features_from_features
-    from lerobot.datasets.io_utils import embed_images
+    from .feature_utils import get_hf_features_from_features
+    from .io_utils import embed_images
 
     hf_features = get_hf_features_from_features(meta.features)
     ep_dataset = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=hf_features, split="train")
@@ -1367,7 +1367,7 @@ def _copy_data_without_images(
         episode_indices: Episodes to include
         img_keys: Image keys to remove
     """
-    from lerobot.datasets.utils import DATA_DIR
+    from .utils import DATA_DIR
 
     data_dir = src_dataset.root / DATA_DIR
     parquet_files = sorted(data_dir.glob("*/*.parquet"))

src/lerobot/datasets/dataset_writer.py

@@ -31,26 +31,26 @@ import PIL.Image
 import pyarrow.parquet as pq
 import torch
 
-from lerobot.datasets.compute_stats import compute_episode_stats
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import (
+from .compute_stats import compute_episode_stats
+from .dataset_metadata import LeRobotDatasetMetadata
+from .feature_utils import (
     get_hf_features_from_features,
     validate_episode_buffer,
     validate_frame,
 )
-from lerobot.datasets.image_writer import AsyncImageWriter, write_image
-from lerobot.datasets.io_utils import (
+from .image_writer import AsyncImageWriter, write_image
+from .io_utils import (
     embed_images,
     get_file_size_in_mb,
     load_episodes,
     write_info,
 )
-from lerobot.datasets.utils import (
+from .utils import (
     DEFAULT_EPISODES_PATH,
     DEFAULT_IMAGE_PATH,
     update_chunk_file_indices,
 )
-from lerobot.datasets.video_utils import (
+from .video_utils import (
     StreamingVideoEncoder,
     concatenate_video_files,
     encode_video_frames,

src/lerobot/datasets/factory.py

@@ -18,19 +18,15 @@ from pprint import pformat
 
 import torch
 
-from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.multi_dataset import MultiLeRobotDataset
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-from lerobot.datasets.transforms import ImageTransforms
-from lerobot.utils.constants import ACTION, OBS_PREFIX, REWARD
+from lerobot.transforms import ImageTransforms
+from lerobot.utils.constants import ACTION, IMAGENET_STATS, OBS_PREFIX, REWARD
 
-IMAGENET_STATS = {
-    "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
-    "std": [[[0.229]], [[0.224]], [[0.225]]],  # (c,1,1)
-}
+from .dataset_metadata import LeRobotDatasetMetadata
+from .lerobot_dataset import LeRobotDataset
+from .multi_dataset import MultiLeRobotDataset
+from .streaming_dataset import StreamingLeRobotDataset
 
 
 def resolve_delta_timestamps(

src/lerobot/datasets/feature_utils.py

@@ -14,23 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from pprint import pformat
-from typing import Any
 
 import datasets
 import numpy as np
 from PIL import Image as PILImage
 
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.utils import (
+from lerobot.utils.constants import DEFAULT_FEATURES
+from lerobot.utils.utils import is_valid_numpy_dtype_string
+
+from .utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
-    DEFAULT_FEATURES,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
 )
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STR
-from lerobot.utils.utils import is_valid_numpy_dtype_string
 
 
 def get_hf_features_from_features(features: dict) -> datasets.Features:
@@ -71,199 +69,6 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
     return datasets.Features(hf_features)
 
 
-def _validate_feature_names(features: dict[str, dict]) -> None:
-    """Validate that feature names do not contain invalid characters.
-
-    Args:
-        features (dict): The LeRobot features dictionary.
-
-    Raises:
-        ValueError: If any feature name contains '/'.
-    """
-    invalid_features = {name: ft for name, ft in features.items() if "/" in name}
-    if invalid_features:
-        raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
-
-
-def hw_to_dataset_features(
-    hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
-) -> dict[str, dict]:
-    """Convert hardware-specific features to a LeRobot dataset feature dictionary.
-
-    This function takes a dictionary describing hardware outputs (like joint states
-    or camera image shapes) and formats it into the standard LeRobot feature
-    specification.
-
-    Args:
-        hw_features (dict): Dictionary mapping feature names to their type (float for
-            joints) or shape (tuple for images).
-        prefix (str): The prefix to add to the feature keys (e.g., "observation"
-            or "action").
-        use_video (bool): If True, image features are marked as "video", otherwise "image".
-
-    Returns:
-        dict: A LeRobot features dictionary.
-    """
-    features = {}
-    joint_fts = {
-        key: ftype
-        for key, ftype in hw_features.items()
-        if ftype is float or (isinstance(ftype, PolicyFeature) and ftype.type != FeatureType.VISUAL)
-    }
-    cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
-
-    if joint_fts and prefix == ACTION:
-        features[prefix] = {
-            "dtype": "float32",
-            "shape": (len(joint_fts),),
-            "names": list(joint_fts),
-        }
-
-    if joint_fts and prefix == OBS_STR:
-        features[f"{prefix}.state"] = {
-            "dtype": "float32",
-            "shape": (len(joint_fts),),
-            "names": list(joint_fts),
-        }
-
-    for key, shape in cam_fts.items():
-        features[f"{prefix}.images.{key}"] = {
-            "dtype": "video" if use_video else "image",
-            "shape": shape,
-            "names": ["height", "width", "channels"],
-        }
-
-    _validate_feature_names(features)
-    return features
-
-
-def build_dataset_frame(
-    ds_features: dict[str, dict], values: dict[str, Any], prefix: str
-) -> dict[str, np.ndarray]:
-    """Construct a single data frame from raw values based on dataset features.
-
-    A "frame" is a dictionary containing all the data for a single timestep,
-    formatted as numpy arrays according to the feature specification.
-
-    Args:
-        ds_features (dict): The LeRobot dataset features dictionary.
-        values (dict): A dictionary of raw values from the hardware/environment.
-        prefix (str): The prefix to filter features by (e.g., "observation"
-            or "action").
-
-    Returns:
-        dict: A dictionary representing a single frame of data.
-    """
-    frame = {}
-    for key, ft in ds_features.items():
-        if key in DEFAULT_FEATURES or not key.startswith(prefix):
-            continue
-        elif ft["dtype"] == "float32" and len(ft["shape"]) == 1:
-            frame[key] = np.array([values[name] for name in ft["names"]], dtype=np.float32)
-        elif ft["dtype"] in ["image", "video"]:
-            frame[key] = values[key.removeprefix(f"{prefix}.images.")]
-
-    return frame
-
-
-def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
-    """Convert dataset features to policy features.
-
-    This function transforms the dataset's feature specification into a format
-    that a policy can use, classifying features by type (e.g., visual, state,
-    action) and ensuring correct shapes (e.g., channel-first for images).
-
-    Args:
-        features (dict): The LeRobot dataset features dictionary.
-
-    Returns:
-        dict: A dictionary mapping feature keys to `PolicyFeature` objects.
-
-    Raises:
-        ValueError: If an image feature does not have a 3D shape.
-    """
-    # TODO(aliberts): Implement "type" in dataset features and simplify this
-    policy_features = {}
-    for key, ft in features.items():
-        shape = ft["shape"]
-        if ft["dtype"] in ["image", "video"]:
-            type = FeatureType.VISUAL
-            if len(shape) != 3:
-                raise ValueError(f"Number of dimensions of {key} != 3 (shape={shape})")
-
-            names = ft["names"]
-            # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
-            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
-                shape = (shape[2], shape[0], shape[1])
-        elif key == OBS_ENV_STATE:
-            type = FeatureType.ENV
-        elif key.startswith(OBS_STR):
-            type = FeatureType.STATE
-        elif key.startswith(ACTION):
-            type = FeatureType.ACTION
-        else:
-            continue
-
-        policy_features[key] = PolicyFeature(
-            type=type,
-            shape=shape,
-        )
-
-    return policy_features
-
-
-def combine_feature_dicts(*dicts: dict) -> dict:
-    """Merge LeRobot grouped feature dicts.
-
-    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
-    - For others (e.g. `observation.images.*`), the last one wins (if they are identical).
-
-    Args:
-        *dicts: A variable number of LeRobot feature dictionaries to merge.
-
-    Returns:
-        dict: A single merged feature dictionary.
-
-    Raises:
-        ValueError: If there's a dtype mismatch for a feature being merged.
-    """
-    out: dict = {}
-    for d in dicts:
-        for key, value in d.items():
-            if not isinstance(value, dict):
-                out[key] = value
-                continue
-
-            dtype = value.get("dtype")
-            shape = value.get("shape")
-            is_vector = (
-                dtype not in ("image", "video", "string")
-                and isinstance(shape, tuple)
-                and len(shape) == 1
-                and "names" in value
-            )
-
-            if is_vector:
-                # Initialize or retrieve the accumulating dict for this feature key
-                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
-                # Ensure consistent data types across merged entries
-                if "dtype" in target and dtype != target["dtype"]:
-                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
-
-                # Merge feature names: append only new ones to preserve order without duplicates
-                seen = set(target["names"])
-                for n in value["names"]:
-                    if n not in seen:
-                        target["names"].append(n)
-                        seen.add(n)
-                # Recompute the shape to reflect the updated number of features
-                target["shape"] = (len(target["names"]),)
-            else:
-                # For images/videos and non-1D entries: override with the latest definition
-                out[key] = value
-    return out
-
-
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,

src/lerobot/datasets/io_utils.py

@@ -13,7 +13,6 @@
 # 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 json
 from pathlib import Path
 from typing import Any
 
@@ -29,7 +28,10 @@ from datasets.table import embed_table_storage
 from PIL import Image as PILImage
 from torchvision import transforms
 
-from lerobot.datasets.utils import (
+from lerobot.utils.io_utils import load_json, write_json
+from lerobot.utils.utils import SuppressProgressBars, flatten_dict, unflatten_dict
+
+from .utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_EPISODES_PATH,
     DEFAULT_SUBTASKS_PATH,
@@ -37,11 +39,8 @@ from lerobot.datasets.utils import (
     EPISODES_DIR,
     INFO_PATH,
     STATS_PATH,
-    flatten_dict,
     serialize_dict,
-    unflatten_dict,
 )
-from lerobot.utils.utils import SuppressProgressBars
 
 
 def get_parquet_file_size_in_mb(parquet_path: str | Path) -> float:
@@ -116,33 +115,6 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
     return dataset
 
 
-def load_json(fpath: Path) -> Any:
-    """Load data from a JSON file.
-
-    Args:
-        fpath (Path): Path to the JSON file.
-
-    Returns:
-        Any: The data loaded from the JSON file.
-    """
-    with open(fpath) as f:
-        return json.load(f)
-
-
-def write_json(data: dict, fpath: Path) -> None:
-    """Write data to a JSON file.
-
-    Creates parent directories if they don't exist.
-
-    Args:
-        data (dict): The dictionary to write.
-        fpath (Path): The path to the output JSON file.
-    """
-    fpath.parent.mkdir(exist_ok=True, parents=True)
-    with open(fpath, "w") as f:
-        json.dump(data, f, indent=4, ensure_ascii=False)
-
-
 def write_info(info: dict, local_dir: Path) -> None:
     write_json(info, local_dir / INFO_PATH)
 

src/lerobot/datasets/lerobot_dataset.py

@@ -24,20 +24,21 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.dataset_reader import DatasetReader
-from lerobot.datasets.dataset_writer import DatasetWriter
-from lerobot.datasets.utils import (
+from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE
+
+from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from .dataset_reader import DatasetReader
+from .dataset_writer import DatasetWriter
+from .utils import (
     create_lerobot_dataset_card,
     get_safe_version,
     is_valid_version,
 )
-from lerobot.datasets.video_utils import (
+from .video_utils import (
     StreamingVideoEncoder,
     get_safe_default_codec,
     resolve_vcodec,
 )
-from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE
 
 logger = logging.getLogger(__name__)
 
@@ -278,6 +279,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def _ensure_reader(self) -> DatasetReader:
         """Lazily create the reader on first access."""
         if self.reader is None:
+            self.meta.ensure_readable()
             self.reader = DatasetReader(
                 meta=self.meta,
                 root=self.root,

src/lerobot/datasets/multi_dataset.py

@@ -21,12 +21,13 @@ import datasets
 import torch
 import torch.utils
 
-from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.feature_utils import get_hf_features_from_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.video_utils import VideoFrame
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
+from .compute_stats import aggregate_stats
+from .feature_utils import get_hf_features_from_features
+from .lerobot_dataset import LeRobotDataset
+from .video_utils import VideoFrame
+
 logger = logging.getLogger(__name__)
 
 

src/lerobot/datasets/pipeline_features.py

@@ -16,11 +16,11 @@ import re
 from collections.abc import Sequence
 from typing import Any
 
-from lerobot.configs.types import PipelineFeatureType
-from lerobot.datasets.feature_utils import hw_to_dataset_features
+from lerobot.configs import PipelineFeatureType
 from lerobot.processor import DataProcessorPipeline
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 
 def create_initial_features(

src/lerobot/datasets/streaming_dataset.py

@@ -22,20 +22,21 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import get_delta_indices
-from lerobot.datasets.io_utils import item_to_torch
-from lerobot.datasets.utils import (
+from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
+
+from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from .feature_utils import get_delta_indices
+from .io_utils import item_to_torch
+from .utils import (
     check_version_compatibility,
     find_float_index,
     is_float_in_list,
     safe_shard,
 )
-from lerobot.datasets.video_utils import (
+from .video_utils import (
     VideoDecoderCache,
     decode_video_frames_torchcodec,
 )
-from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 
 
 class LookBackError(Exception):

src/lerobot/datasets/utils.py

@@ -17,9 +17,7 @@ import contextlib
 import importlib.resources
 import json
 import logging
-from collections.abc import Iterator
 from pathlib import Path
-from typing import Any
 
 import datasets
 import numpy as np
@@ -28,6 +26,8 @@ import torch
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
 
+from lerobot.utils.utils import flatten_dict, unflatten_dict
+
 V30_MESSAGE = """
 The dataset you requested ({repo_id}) is in {version} format.
 
@@ -93,14 +93,6 @@ LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
 LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
 LEGACY_TASKS_PATH = "meta/tasks.jsonl"
 
-DEFAULT_FEATURES = {
-    "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
-    "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
-    "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
-    "index": {"dtype": "int64", "shape": (1,), "names": None},
-    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
-}
-
 
 def has_legacy_hub_download_metadata(root: Path) -> bool:
     """Return ``True`` when *root* looks like a legacy Hub ``local_dir`` mirror.
@@ -123,59 +115,6 @@ def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int) -
     return chunk_idx, file_idx
 
 
-def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
-    """Flatten a nested dictionary by joining keys with a separator.
-
-    Example:
-        >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}
-        >>> print(flatten_dict(dct))
-        {'a/b': 1, 'a/c/d': 2, 'e': 3}
-
-    Args:
-        d (dict): The dictionary to flatten.
-        parent_key (str): The base key to prepend to the keys in this level.
-        sep (str): The separator to use between keys.
-
-    Returns:
-        dict: A flattened dictionary.
-    """
-    items = []
-    for k, v in d.items():
-        new_key = f"{parent_key}{sep}{k}" if parent_key else k
-        if isinstance(v, dict):
-            items.extend(flatten_dict(v, new_key, sep=sep).items())
-        else:
-            items.append((new_key, v))
-    return dict(items)
-
-
-def unflatten_dict(d: dict, sep: str = "/") -> dict:
-    """Unflatten a dictionary with delimited keys into a nested dictionary.
-
-    Example:
-        >>> flat_dct = {"a/b": 1, "a/c/d": 2, "e": 3}
-        >>> print(unflatten_dict(flat_dct))
-        {'a': {'b': 1, 'c': {'d': 2}}, 'e': 3}
-
-    Args:
-        d (dict): A dictionary with flattened keys.
-        sep (str): The separator used in the keys.
-
-    Returns:
-        dict: A nested dictionary.
-    """
-    outdict = {}
-    for key, value in d.items():
-        parts = key.split(sep)
-        d = outdict
-        for part in parts[:-1]:
-            if part not in d:
-                d[part] = {}
-            d = d[part]
-        d[parts[-1]] = value
-    return outdict
-
-
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     """Serialize a dictionary containing tensors or numpy arrays to be JSON-compatible.
 
@@ -332,27 +271,6 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
     raise ForwardCompatibilityError(repo_id, min(upper_versions))
 
 
-def cycle(iterable: Any) -> Iterator[Any]:
-    """Create a dataloader-safe cyclical iterator.
-
-    This is an equivalent of `itertools.cycle` but is safe for use with
-    PyTorch DataLoaders with multiple workers.
-    See https://github.com/pytorch/pytorch/issues/23900 for details.
-
-    Args:
-        iterable: The iterable to cycle over.
-
-    Yields:
-        Items from the iterable, restarting from the beginning when exhausted.
-    """
-    iterator = iter(iterable)
-    while True:
-        try:
-            yield next(iterator)
-        except StopIteration:
-            iterator = iter(iterable)
-
-
 def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) -> None:
     """Create a branch on an existing Hugging Face repo.
 

src/lerobot/datasets/video_utils.py

@@ -37,6 +37,8 @@ import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
+from lerobot.utils.import_utils import get_safe_default_codec
+
 logger = logging.getLogger(__name__)
 
 # List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
@@ -116,16 +118,6 @@ def resolve_vcodec(vcodec: str) -> str:
     return "libsvtav1"
 
 
-def get_safe_default_codec():
-    if importlib.util.find_spec("torchcodec"):
-        return "torchcodec"
-    else:
-        logger.warning(
-            "'torchcodec' is not available in your platform, falling back to 'pyav' as a default decoder"
-        )
-        return "pyav"
-
-
 def decode_video_frames(
     video_path: Path | str,
     timestamps: list[float],
@@ -271,7 +263,10 @@ class VideoDecoderCache:
         if importlib.util.find_spec("torchcodec"):
             from torchcodec.decoders import VideoDecoder
         else:
-            raise ImportError("torchcodec is required but not available.")
+            raise ImportError(
+                "'torchcodec' is required but not installed. "
+                "Install it with: pip install 'lerobot[dataset]' (or uv pip install 'lerobot[dataset]')"
+            )
 
         video_path = str(video_path)
 
@@ -606,7 +601,7 @@ class _CameraEncoderThread(threading.Thread):
         self.encoder_threads = encoder_threads
 
     def run(self) -> None:
-        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
+        from .compute_stats import RunningQuantileStats, auto_downsample_height_width
 
         container = None
         output_stream = None

src/lerobot/envs/__init__.py

@@ -12,4 +12,27 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configs import AlohaEnv, EnvConfig, HubEnvConfig, PushtEnv  # noqa: F401
+# NOTE: gymnasium is currently a core dependency but is a candidate for moving to an
+# optional extra in the future. When that transition happens, uncomment the guard below
+# and update the extra name to the one that will contain gymnasium.
+# from lerobot.utils.import_utils import require_package
+# require_package("gymnasium", extra="<update_extra>", import_name="gymnasium")
+
+from .configs import AlohaEnv, EnvConfig, HILSerlRobotEnvConfig, HubEnvConfig, PushtEnv
+from .factory import make_env, make_env_config, make_env_pre_post_processors
+from .utils import check_env_attributes_and_types, close_envs, env_to_policy_features, preprocess_observation
+
+__all__ = [
+    "AlohaEnv",
+    "EnvConfig",
+    "HILSerlRobotEnvConfig",
+    "HubEnvConfig",
+    "PushtEnv",
+    "check_env_attributes_and_types",
+    "close_envs",
+    "env_to_policy_features",
+    "make_env",
+    "make_env_config",
+    "make_env_pre_post_processors",
+    "preprocess_observation",
+]

src/lerobot/envs/configs.py

@@ -23,7 +23,8 @@ import draccus
 import gymnasium as gym
 from gymnasium.envs.registration import registry as gym_registry
 
-from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.configs import FeatureType, PolicyFeature
+from lerobot.processor import IsaaclabArenaProcessorStep, LiberoProcessorStep, PolicyProcessorPipeline
 from lerobot.robots import RobotConfig
 from lerobot.teleoperators.config import TeleoperatorConfig
 from lerobot.utils.constants import (
@@ -124,8 +125,6 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
 
     def get_env_processors(self):
         """Return (preprocessor, postprocessor) for this env. Default: identity."""
-        from lerobot.processor.pipeline import PolicyProcessorPipeline
-
         return PolicyProcessorPipeline(steps=[]), PolicyProcessorPipeline(steps=[])
 
 
@@ -418,7 +417,7 @@ class LiberoEnv(EnvConfig):
         return kwargs
 
     def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from lerobot.envs.libero import create_libero_envs
+        from .libero import create_libero_envs
 
         if self.task is None:
             raise ValueError("LiberoEnv requires a task to be specified")
@@ -436,9 +435,6 @@ class LiberoEnv(EnvConfig):
         )
 
     def get_env_processors(self):
-        from lerobot.processor.env_processor import LiberoProcessorStep
-        from lerobot.processor.pipeline import PolicyProcessorPipeline
-
         return (
             PolicyProcessorPipeline(steps=[LiberoProcessorStep()]),
             PolicyProcessorPipeline(steps=[]),
@@ -487,7 +483,7 @@ class MetaworldEnv(EnvConfig):
         }
 
     def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from lerobot.envs.metaworld import create_metaworld_envs
+        from .metaworld import create_metaworld_envs
 
         if self.task is None:
             raise ValueError("MetaWorld requires a task to be specified")
@@ -568,9 +564,6 @@ class IsaaclabArenaEnv(HubEnvConfig):
         return {}
 
     def get_env_processors(self):
-        from lerobot.processor.env_processor import IsaaclabArenaProcessorStep
-        from lerobot.processor.pipeline import PolicyProcessorPipeline
-
         state_keys = tuple(k.strip() for k in (self.state_keys or "").split(",") if k.strip())
         camera_keys = tuple(k.strip() for k in (self.camera_keys or "").split(",") if k.strip())
         if not state_keys and not camera_keys:

src/lerobot/envs/factory.py

@@ -19,8 +19,8 @@ from typing import Any
 
 import gymnasium as gym
 
-from lerobot.envs.configs import EnvConfig, HubEnvConfig
-from lerobot.envs.utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
+from .configs import EnvConfig, HubEnvConfig
+from .utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:

src/lerobot/envs/libero.py

@@ -29,9 +29,10 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
 
-from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation
 
+from .utils import _LazyAsyncVectorEnv
+
 
 def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
     """Normalize camera_name into a non-empty list of strings."""

src/lerobot/envs/metaworld.py

@@ -25,9 +25,10 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
 
-from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation
 
+from .utils import _LazyAsyncVectorEnv
+
 # ---- Load configuration data from the external JSON file ----
 CONFIG_PATH = Path(__file__).parent / "metaworld_config.json"
 try: