diff --git a/src/lerobot/async_inference/__init__.py b/src/lerobot/async_inference/__init__.py
index 0da31b6a8..f1edccb87 100644
--- a/src/lerobot/async_inference/__init__.py
+++ b/src/lerobot/async_inference/__init__.py
@@ -1,6 +1,16 @@
-# Async inference server/client.
-# Requires: lerobot[async]
+"""
+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] = []
diff --git a/src/lerobot/common/__init__.py b/src/lerobot/common/__init__.py
index 2f9bf10d9..dfd97496d 100644
--- a/src/lerobot/common/__init__.py
+++ b/src/lerobot/common/__init__.py
@@ -5,4 +5,12 @@ 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] = []
diff --git a/src/lerobot/rl/wandb_utils.py b/src/lerobot/common/wandb_utils.py
similarity index 100%
rename from src/lerobot/rl/wandb_utils.py
rename to src/lerobot/common/wandb_utils.py
diff --git a/src/lerobot/data_processing/__init__.py b/src/lerobot/data_processing/__init__.py
index 2f76d5676..cd55d46fc 100644
--- a/src/lerobot/data_processing/__init__.py
+++ b/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] = []
diff --git a/src/lerobot/data_processing/sarm_annotations/__init__.py b/src/lerobot/data_processing/sarm_annotations/__init__.py
index 2f76d5676..cd4c38f33 100644
--- a/src/lerobot/data_processing/sarm_annotations/__init__.py
+++ b/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] = []
diff --git a/src/lerobot/envs/__init__.py b/src/lerobot/envs/__init__.py
index adb40490c..730207d6c 100644
--- a/src/lerobot/envs/__init__.py
+++ b/src/lerobot/envs/__init__.py
@@ -12,6 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# NOTE: gymnasium is currently a core dependency but is a candidate for moving to an
+# optional extra in the future. This guard is here to ensure a clear error message
+# if/when that transition happens.
+from lerobot.utils.import_utils import require_package
+
+require_package("gymnasium", extra="evaluation", 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
diff --git a/src/lerobot/policies/__init__.py b/src/lerobot/policies/__init__.py
index 896cc8617..e138a84d9 100644
--- a/src/lerobot/policies/__init__.py
+++ b/src/lerobot/policies/__init__.py
@@ -22,11 +22,10 @@ from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .rtc import ActionInterpolator as ActionInterpolator
+from .sac.configuration_sac import SACConfig as SACConfig
+from .sac.reward_model.configuration_classifier import RewardClassifierConfig as RewardClassifierConfig
 from .sarm.configuration_sarm import SARMConfig as SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
-from .smolvla.processor_smolvla import (
-    SmolVLANewLineProcessor as SmolVLANewLineProcessor,  # noqa: F401 - registers with ProcessorStepRegistry
-)
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
@@ -46,6 +45,8 @@ __all__ = [
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
+    "RewardClassifierConfig",
+    "SACConfig",
     "SARMConfig",
     "SmolVLAConfig",
     "TDMPCConfig",
diff --git a/src/lerobot/policies/sarm/__init__.py b/src/lerobot/policies/sarm/__init__.py
index 3d52651d9..85b24c59e 100644
--- a/src/lerobot/policies/sarm/__init__.py
+++ b/src/lerobot/policies/sarm/__init__.py
@@ -1,3 +1,4 @@
 from .configuration_sarm import SARMConfig
+from .modeling_sarm import SARMRewardModel
 
-__all__ = ["SARMConfig"]
+__all__ = ["SARMConfig", "SARMRewardModel"]
diff --git a/src/lerobot/policies/smolvla/__init__.py b/src/lerobot/policies/smolvla/__init__.py
index d610dcf46..516faaaf9 100644
--- a/src/lerobot/policies/smolvla/__init__.py
+++ b/src/lerobot/policies/smolvla/__init__.py
@@ -1,7 +1,5 @@
 from .configuration_smolvla import SmolVLAConfig
-from .processor_smolvla import (
-    SmolVLANewLineProcessor,
-    make_smolvla_pre_post_processors,
-)
+from .modeling_smolvla import SmolVLAPolicy
+from .processor_smolvla import make_smolvla_pre_post_processors
 
-__all__ = ["SmolVLAConfig", "SmolVLANewLineProcessor", "make_smolvla_pre_post_processors"]
+__all__ = ["SmolVLAConfig", "SmolVLAPolicy", "make_smolvla_pre_post_processors"]
diff --git a/src/lerobot/policies/smolvla/processor_smolvla.py b/src/lerobot/policies/smolvla/processor_smolvla.py
index 0aa2beb19..8d6c8aca4 100644
--- a/src/lerobot/policies/smolvla/processor_smolvla.py
+++ b/src/lerobot/policies/smolvla/processor_smolvla.py
@@ -18,15 +18,13 @@ from typing import Any
 
 import torch
 
-from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
-    ComplementaryDataProcessorStep,
     DeviceProcessorStep,
+    NewLineTaskProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    ProcessorStepRegistry,
     RenameObservationsProcessorStep,
     TokenizerProcessorStep,
     UnnormalizerProcessorStep,
@@ -71,7 +69,7 @@ def make_smolvla_pre_post_processors(
     input_steps = [
         RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
         AddBatchDimensionProcessorStep(),
-        SmolVLANewLineProcessor(),
+        NewLineTaskProcessorStep(),
         TokenizerProcessorStep(
             tokenizer_name=config.vlm_model_name,
             padding=config.pad_language_to,
@@ -103,41 +101,3 @@ def make_smolvla_pre_post_processors(
             to_output=transition_to_policy_action,
         ),
     )
-
-
-@ProcessorStepRegistry.register(name="smolvla_new_line_processor")
-class SmolVLANewLineProcessor(ComplementaryDataProcessorStep):
-    """
-    A processor step that ensures the 'task' description ends with a newline character.
-
-    This step is necessary for certain tokenizers (e.g., PaliGemma) that expect a
-    newline at the end of the prompt. It handles both single string tasks and lists
-    of string tasks.
-    """
-
-    def complementary_data(self, complementary_data):
-        if "task" not in complementary_data:
-            return complementary_data
-
-        task = complementary_data["task"]
-        if task is None:
-            return complementary_data
-
-        new_complementary_data = dict(complementary_data)
-
-        # Handle both string and list of strings
-        if isinstance(task, str):
-            # Single string: add newline if not present
-            if not task.endswith("\n"):
-                new_complementary_data["task"] = f"{task}\n"
-        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
-            # List of strings: add newline to each if not present
-            new_complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
-        # If task is neither string nor list of strings, leave unchanged
-
-        return new_complementary_data
-
-    def transform_features(
-        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        return features
diff --git a/src/lerobot/policies/smolvla/smolvlm_with_expert.py b/src/lerobot/policies/smolvla/smolvlm_with_expert.py
index b1d93d72d..ea806f185 100644
--- a/src/lerobot/policies/smolvla/smolvlm_with_expert.py
+++ b/src/lerobot/policies/smolvla/smolvlm_with_expert.py
@@ -13,21 +13,27 @@
 # limitations under the License.
 
 import copy
+from typing import TYPE_CHECKING
 
 import torch
 from torch import nn
 
-from lerobot.utils.import_utils import require_package
+from lerobot.utils.import_utils import _transformers_available, require_package
 
-require_package("transformers", extra="smolvla")
-
-from transformers import (
-    AutoConfig,
-    AutoModel,
-    AutoModelForImageTextToText,
-    AutoProcessor,
-    SmolVLMForConditionalGeneration,
-)
+if TYPE_CHECKING or _transformers_available:
+    from transformers import (
+        AutoConfig,
+        AutoModel,
+        AutoModelForImageTextToText,
+        AutoProcessor,
+        SmolVLMForConditionalGeneration,
+    )
+else:
+    AutoConfig = None
+    AutoModel = None
+    AutoModelForImageTextToText = None
+    AutoProcessor = None
+    SmolVLMForConditionalGeneration = None
 
 
 def apply_rope(x, positions, max_wavelength=10_000):
@@ -78,6 +84,7 @@ class SmolVLMWithExpertModel(nn.Module):
         device: str = "auto",
     ):
         super().__init__()
+        require_package("transformers", extra="smolvla")
         if load_vlm_weights:
             print(f"Loading  {model_id} weights ...")
             self.vlm = AutoModelForImageTextToText.from_pretrained(
diff --git a/src/lerobot/policies/wall_x/__init__.py b/src/lerobot/policies/wall_x/__init__.py
index 7c10b8b2a..16fd2c8ab 100644
--- a/src/lerobot/policies/wall_x/__init__.py
+++ b/src/lerobot/policies/wall_x/__init__.py
@@ -15,6 +15,7 @@
 # limitations under the License.
 
 from .configuration_wall_x import WallXConfig
+from .modeling_wall_x import WallXPolicy
 from .processor_wall_x import make_wall_x_pre_post_processors
 
-__all__ = ["WallXConfig", "make_wall_x_pre_post_processors"]
+__all__ = ["WallXConfig", "WallXPolicy", "make_wall_x_pre_post_processors"]
diff --git a/src/lerobot/policies/wall_x/modeling_wall_x.py b/src/lerobot/policies/wall_x/modeling_wall_x.py
index a95b42e9d..b90b7e564 100644
--- a/src/lerobot/policies/wall_x/modeling_wall_x.py
+++ b/src/lerobot/policies/wall_x/modeling_wall_x.py
@@ -34,40 +34,28 @@ lerobot-train \
 ```
 """
 
+import logging
 import math
 from collections import deque
 from os import PathLike
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-
-from lerobot.utils.import_utils import require_package
-
-require_package("transformers", extra="wallx")
-require_package("peft", extra="wallx")
-require_package("torchdiffeq", extra="wallx")
-require_package("qwen-vl-utils", extra="wallx", import_name="qwen_vl_utils")
-
-from peft import LoraConfig, get_peft_model
 from PIL import Image
-from qwen_vl_utils.vision_process import smart_resize
 from torch import Tensor
 from torch.distributions import Beta
 from torch.nn import CrossEntropyLoss
-from torchdiffeq import odeint
-from transformers import AutoProcessor, BatchFeature
-from transformers.cache_utils import (
-    StaticCache,
-)
-from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
-    Qwen2_5_VLForConditionalGeneration,
-)
-from transformers.utils import is_torchdynamo_compiling, logging
 
 from lerobot.utils.constants import ACTION, OBS_STATE
+from lerobot.utils.import_utils import (
+    _peft_available,
+    _transformers_available,
+    is_package_available,
+    require_package,
+)
 
 from ..pretrained import PreTrainedPolicy
 from ..utils import populate_queues
@@ -82,12 +70,45 @@ from .constant import (
     RESOLUTION,
     TOKENIZER_MAX_LENGTH,
 )
-from .qwen_model.configuration_qwen2_5_vl import Qwen2_5_VLConfig
-from .qwen_model.qwen2_5_vl_moe import (
-    Qwen2_5_VisionTransformerPretrainedModel,
-    Qwen2_5_VLACausalLMOutputWithPast,
-    Qwen2_5_VLMoEModel,
+
+_torchdiffeq_available = is_package_available("torchdiffeq")
+_qwen_vl_utils_available = is_package_available("qwen-vl-utils", import_name="qwen_vl_utils")
+_wallx_deps_available = (
+    _transformers_available and _peft_available and _torchdiffeq_available and _qwen_vl_utils_available
 )
+
+if TYPE_CHECKING or _wallx_deps_available:
+    from peft import LoraConfig, get_peft_model
+    from qwen_vl_utils.vision_process import smart_resize
+    from torchdiffeq import odeint
+    from transformers import AutoProcessor, BatchFeature
+    from transformers.cache_utils import StaticCache
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+        Qwen2_5_VLForConditionalGeneration,
+    )
+    from transformers.utils import is_torchdynamo_compiling
+
+    from .qwen_model.configuration_qwen2_5_vl import Qwen2_5_VLConfig
+    from .qwen_model.qwen2_5_vl_moe import (
+        Qwen2_5_VisionTransformerPretrainedModel,
+        Qwen2_5_VLACausalLMOutputWithPast,
+        Qwen2_5_VLMoEModel,
+    )
+else:
+    LoraConfig = None
+    get_peft_model = None
+    smart_resize = None
+    odeint = None
+    AutoProcessor = None
+    BatchFeature = None
+    StaticCache = None
+    Qwen2_5_VLForConditionalGeneration = None
+    is_torchdynamo_compiling = None
+    Qwen2_5_VLConfig = None
+    Qwen2_5_VisionTransformerPretrainedModel = None
+    Qwen2_5_VLACausalLMOutputWithPast = None
+    Qwen2_5_VLMoEModel = None
+
 from .utils import (
     get_wallx_normal_text,
     preprocesser_call,
@@ -95,7 +116,7 @@ from .utils import (
     replace_action_token,
 )
 
-logger = logging.get_logger(__name__)
+logger = logging.getLogger(__name__)
 
 
 class SinusoidalPosEmb(nn.Module):
@@ -262,7 +283,13 @@ class ActionHead(nn.Module):
         return self.propri_proj(proprioception)
 
 
-class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
+# Conditional base: when transformers is unavailable the class still parses
+# (inheriting from nn.Module) but cannot be instantiated—require_package in
+# WallXPolicy.__init__ gives the user a clear error before that happens.
+_Qwen2_5_VLForAction_Base = Qwen2_5_VLForConditionalGeneration if _wallx_deps_available else nn.Module
+
+
+class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
     """
     Qwen2.5 Vision-Language Mixture of Experts model for action processing.
 
@@ -1717,6 +1744,10 @@ class WallXPolicy(PreTrainedPolicy):
     name = "wall_x"
 
     def __init__(self, config: WallXConfig, **kwargs):
+        require_package("transformers", extra="wallx")
+        require_package("peft", extra="wallx")
+        require_package("torchdiffeq", extra="wallx")
+        require_package("qwen-vl-utils", extra="wallx", import_name="qwen_vl_utils")
         super().__init__(config)
         config.validate_features()
         self.config = config
diff --git a/src/lerobot/policies/wall_x/utils.py b/src/lerobot/policies/wall_x/utils.py
index 39a20afce..d38a2d509 100644
--- a/src/lerobot/policies/wall_x/utils.py
+++ b/src/lerobot/policies/wall_x/utils.py
@@ -25,10 +25,16 @@ import random
 import re
 from collections import OrderedDict
 from dataclasses import dataclass, field
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import torch
-from transformers import BatchFeature
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import BatchFeature
+else:
+    BatchFeature = None
 
 from lerobot.utils.constants import OBS_IMAGES
 
diff --git a/src/lerobot/policies/xvla/__init__.py b/src/lerobot/policies/xvla/__init__.py
index 7a8cce8bb..58609e91c 100644
--- a/src/lerobot/policies/xvla/__init__.py
+++ b/src/lerobot/policies/xvla/__init__.py
@@ -1,4 +1,5 @@
 from .configuration_xvla import XVLAConfig
+from .modeling_xvla import XVLAPolicy
 from .processor_xvla import (
     XVLAAddDomainIdProcessorStep,
     XVLAImageNetNormalizeProcessorStep,
@@ -7,6 +8,7 @@ from .processor_xvla import (
 
 __all__ = [
     "XVLAConfig",
+    "XVLAPolicy",
     "XVLAAddDomainIdProcessorStep",
     "XVLAImageNetNormalizeProcessorStep",
     "XVLAImageToFloatProcessorStep",
diff --git a/src/lerobot/policies/xvla/modeling_xvla.py b/src/lerobot/policies/xvla/modeling_xvla.py
index 0f8bdd036..04e923fdd 100644
--- a/src/lerobot/policies/xvla/modeling_xvla.py
+++ b/src/lerobot/policies/xvla/modeling_xvla.py
@@ -23,10 +23,7 @@ import logging
 import os
 from collections import deque
 from pathlib import Path
-
-from lerobot.utils.import_utils import require_package
-
-require_package("transformers", extra="xvla")
+from typing import TYPE_CHECKING
 
 import torch
 import torch.nn.functional as F  # noqa: N812
@@ -34,15 +31,22 @@ from torch import Tensor, nn
 
 from lerobot.configs import PreTrainedConfig
 from lerobot.utils.constants import ACTION, OBS_LANGUAGE_TOKENS, OBS_STATE
+from lerobot.utils.import_utils import _transformers_available, require_package
 
 from ..pretrained import PreTrainedPolicy, T
 from ..utils import populate_queues
 from .action_hub import build_action_space
-from .configuration_florence2 import Florence2Config
 from .configuration_xvla import XVLAConfig
-from .modeling_florence2 import Florence2ForConditionalGeneration
 from .soft_transformer import SoftPromptedTransformer
 
+# Florence2 config and modeling depend on transformers
+if TYPE_CHECKING or _transformers_available:
+    from .configuration_florence2 import Florence2Config
+    from .modeling_florence2 import Florence2ForConditionalGeneration
+else:
+    Florence2Config = None
+    Florence2ForConditionalGeneration = None
+
 
 class XVLAModel(nn.Module):
     """
@@ -278,6 +282,7 @@ class XVLAPolicy(PreTrainedPolicy):
     name = "xvla"
 
     def __init__(self, config: XVLAConfig, **kwargs):
+        require_package("transformers", extra="xvla")
         super().__init__(config)
         config.validate_features()
         florence_config = config.get_florence_config()
diff --git a/src/lerobot/processor/__init__.py b/src/lerobot/processor/__init__.py
index 213bd93f2..20cdf310c 100644
--- a/src/lerobot/processor/__init__.py
+++ b/src/lerobot/processor/__init__.py
@@ -56,6 +56,7 @@ from .hil_processor import (
     RewardClassifierProcessorStep,
     TimeLimitProcessorStep,
 )
+from .newline_task_processor import NewLineTaskProcessorStep
 from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep, hotswap_stats
 from .observation_processor import VanillaObservationProcessorStep
 from .pipeline import (
@@ -119,6 +120,7 @@ __all__ = [
     "RelativeActionsProcessorStep",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
+    "NewLineTaskProcessorStep",
     "NormalizerProcessorStep",
     "Numpy2TorchActionProcessorStep",
     "ObservationProcessorStep",
diff --git a/src/lerobot/processor/newline_task_processor.py b/src/lerobot/processor/newline_task_processor.py
new file mode 100644
index 000000000..ea61bdd71
--- /dev/null
+++ b/src/lerobot/processor/newline_task_processor.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python
+
+# Copyright 2025 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.
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+
+from .pipeline import ComplementaryDataProcessorStep, ProcessorStepRegistry
+
+
+# NOTE: The registry name "smolvla_new_line_processor" is kept for backward compatibility
+# with serialized processor configs that reference this name.
+@ProcessorStepRegistry.register(name="smolvla_new_line_processor")
+class NewLineTaskProcessorStep(ComplementaryDataProcessorStep):
+    """
+    A processor step that ensures the 'task' description ends with a newline character.
+
+    This step is necessary for certain tokenizers (e.g., PaliGemma) that expect a
+    newline at the end of the prompt. It handles both single string tasks and lists
+    of string tasks.
+    """
+
+    def complementary_data(self, complementary_data):
+        if "task" not in complementary_data:
+            return complementary_data
+
+        task = complementary_data["task"]
+        if task is None:
+            return complementary_data
+
+        new_complementary_data = dict(complementary_data)
+
+        # Handle both string and list of strings
+        if isinstance(task, str):
+            # Single string: add newline if not present
+            if not task.endswith("\n"):
+                new_complementary_data["task"] = f"{task}\n"
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            # List of strings: add newline to each if not present
+            new_complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return new_complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
diff --git a/src/lerobot/processor/policy_robot_bridge.py b/src/lerobot/processor/policy_robot_bridge.py
index 1c5327bb5..25d622dc2 100644
--- a/src/lerobot/processor/policy_robot_bridge.py
+++ b/src/lerobot/processor/policy_robot_bridge.py
@@ -20,9 +20,11 @@ from typing import Any
 import torch
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.processor import ActionProcessorStep, PolicyAction, ProcessorStepRegistry, RobotAction
+from lerobot.types import PolicyAction, RobotAction
 from lerobot.utils.constants import ACTION
 
+from .pipeline import ActionProcessorStep, ProcessorStepRegistry
+
 
 @dataclass
 @ProcessorStepRegistry.register("robot_action_to_policy_action_processor")
diff --git a/src/lerobot/rl/__init__.py b/src/lerobot/rl/__init__.py
index 41037b98d..f6c5df0db 100644
--- a/src/lerobot/rl/__init__.py
+++ b/src/lerobot/rl/__init__.py
@@ -1,6 +1,20 @@
-# Reinforcement learning modules.
-# Requires: lerobot[hilserl]
+"""
+Reinforcement learning modules.
+
+Requires: ``pip install 'lerobot[hilserl]'``
+
+Available modules (import directly)::
+
+    from lerobot.rl.actor import ...
+    from lerobot.rl.learner import ...
+    from lerobot.rl.learner_service import ...
+    from lerobot.rl.buffer import ...
+    from lerobot.rl.eval_policy import ...
+    from lerobot.rl.gym_manipulator import ...
+"""
 
 from lerobot.utils.import_utils import require_package
 
 require_package("grpcio", extra="hilserl", import_name="grpc")
+
+__all__: list[str] = []
diff --git a/src/lerobot/rl/learner.py b/src/lerobot/rl/learner.py
index 8f54f8c1b..073d9a65f 100644
--- a/src/lerobot/rl/learner.py
+++ b/src/lerobot/rl/learner.py
@@ -66,6 +66,7 @@ from lerobot.common.train_utils import (
     save_checkpoint,
     update_last_checkpoint,
 )
+from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets import LeRobotDataset, make_dataset
@@ -99,7 +100,6 @@ from lerobot.utils.utils import (
 from .buffer import ReplayBuffer, concatenate_batch_transitions
 from .learner_service import MAX_WORKERS, SHUTDOWN_TIMEOUT, LearnerService
 from .process import ProcessSignalHandler
-from .wandb_utils import WandBLogger
 
 
 @parser.wrap()
@@ -151,7 +151,7 @@ def train(cfg: TrainRLServerPipelineConfig, job_name: str | None = None):
 
     # Setup WandB logging if enabled
     if cfg.wandb.enable and cfg.wandb.project:
-        from .wandb_utils import WandBLogger
+        from lerobot.common.wandb_utils import WandBLogger
 
         wandb_logger = WandBLogger(cfg)
     else:
diff --git a/src/lerobot/scripts/lerobot_find_cameras.py b/src/lerobot/scripts/lerobot_find_cameras.py
index ab999f7ae..72f4096da 100644
--- a/src/lerobot/scripts/lerobot_find_cameras.py
+++ b/src/lerobot/scripts/lerobot_find_cameras.py
@@ -38,10 +38,8 @@ import numpy as np
 from PIL import Image
 
 from lerobot.cameras import ColorMode
-from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.cameras.opencv.camera_opencv import OpenCVCamera
-from lerobot.cameras.realsense import RealSenseCameraConfig
-from lerobot.cameras.realsense.camera_realsense import RealSenseCamera
+from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
+from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
 
 logger = logging.getLogger(__name__)
 
diff --git a/src/lerobot/scripts/lerobot_record.py b/src/lerobot/scripts/lerobot_record.py
index f6787e149..8f1d27602 100644
--- a/src/lerobot/scripts/lerobot_record.py
+++ b/src/lerobot/scripts/lerobot_record.py
@@ -142,7 +142,7 @@ from lerobot.teleoperators import (  # noqa: F401
     so_leader,
     unitree_g1,
 )
-from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop
+from lerobot.teleoperators.keyboard import KeyboardTeleop
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
diff --git a/src/lerobot/scripts/lerobot_train.py b/src/lerobot/scripts/lerobot_train.py
index 0e27b45bd..c943767e3 100644
--- a/src/lerobot/scripts/lerobot_train.py
+++ b/src/lerobot/scripts/lerobot_train.py
@@ -35,13 +35,13 @@ from lerobot.common.train_utils import (
     save_checkpoint,
     update_last_checkpoint,
 )
+from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.datasets import EpisodeAwareSampler, make_dataset
 from lerobot.envs import close_envs, make_env, make_env_pre_post_processors
 from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
-from lerobot.rl.wandb_utils import WandBLogger
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
 from lerobot.utils.random_utils import set_seed
diff --git a/src/lerobot/transport/__init__.py b/src/lerobot/transport/__init__.py
index 7ea343691..355ad4cb9 100644
--- a/src/lerobot/transport/__init__.py
+++ b/src/lerobot/transport/__init__.py
@@ -1,2 +1,11 @@
-# gRPC transport layer for async inference.
-# Requires: lerobot[grpcio-dep]
+"""
+gRPC transport layer for async inference.
+
+Requires: ``pip install 'lerobot[grpcio-dep]'``
+
+Available modules (import directly)::
+
+    from lerobot.transport.utils import ...
+"""
+
+__all__: list[str] = []
diff --git a/tests/processor/test_smolvla_processor.py b/tests/processor/test_smolvla_processor.py
index 227b1dc35..2aa7d4bdf 100644
--- a/tests/processor/test_smolvla_processor.py
+++ b/tests/processor/test_smolvla_processor.py
@@ -22,14 +22,12 @@ import torch
 
 from lerobot.configs.types import FeatureType, NormalizationMode, PipelineFeatureType, PolicyFeature
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
-from lerobot.policies.smolvla.processor_smolvla import (
-    SmolVLANewLineProcessor,
-    make_smolvla_pre_post_processors,
-)
+from lerobot.policies.smolvla.processor_smolvla import make_smolvla_pre_post_processors
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
     EnvTransition,
+    NewLineTaskProcessorStep,
     NormalizerProcessorStep,
     ProcessorStep,
     RenameObservationsProcessorStep,
@@ -108,7 +106,7 @@ def test_make_smolvla_processor_basic():
     assert len(preprocessor.steps) == 6
     assert isinstance(preprocessor.steps[0], RenameObservationsProcessorStep)
     assert isinstance(preprocessor.steps[1], AddBatchDimensionProcessorStep)
-    assert isinstance(preprocessor.steps[2], SmolVLANewLineProcessor)
+    assert isinstance(preprocessor.steps[2], NewLineTaskProcessorStep)
     # Step 3 would be TokenizerProcessorStep but it's mocked
     assert isinstance(preprocessor.steps[4], DeviceProcessorStep)
     assert isinstance(preprocessor.steps[5], NormalizerProcessorStep)
@@ -120,8 +118,8 @@ def test_make_smolvla_processor_basic():
 
 
 def test_smolvla_newline_processor_single_task():
-    """Test SmolVLANewLineProcessor with single task string."""
-    processor = SmolVLANewLineProcessor()
+    """Test NewLineTaskProcessorStep with single task string."""
+    processor = NewLineTaskProcessorStep()
 
     # Test with task that doesn't have newline
     transition = create_transition(complementary_data={"task": "test task"})
@@ -135,8 +133,8 @@ def test_smolvla_newline_processor_single_task():
 
 
 def test_smolvla_newline_processor_list_of_tasks():
-    """Test SmolVLANewLineProcessor with list of task strings."""
-    processor = SmolVLANewLineProcessor()
+    """Test NewLineTaskProcessorStep with list of task strings."""
+    processor = NewLineTaskProcessorStep()
 
     # Test with list of tasks
     tasks = ["task1", "task2\n", "task3"]
@@ -147,8 +145,8 @@ def test_smolvla_newline_processor_list_of_tasks():
 
 
 def test_smolvla_newline_processor_empty_transition():
-    """Test SmolVLANewLineProcessor with empty transition."""
-    processor = SmolVLANewLineProcessor()
+    """Test NewLineTaskProcessorStep with empty transition."""
+    processor = NewLineTaskProcessorStep()
 
     # Test with no complementary_data
     transition = create_transition()
@@ -361,8 +359,8 @@ def test_smolvla_processor_without_stats():
 
 
 def test_smolvla_newline_processor_state_dict():
-    """Test SmolVLANewLineProcessor state dict methods."""
-    processor = SmolVLANewLineProcessor()
+    """Test NewLineTaskProcessorStep state dict methods."""
+    processor = NewLineTaskProcessorStep()
 
     # Test state_dict (should be empty)
     state = processor.state_dict()
@@ -380,8 +378,8 @@ def test_smolvla_newline_processor_state_dict():
 
 
 def test_smolvla_newline_processor_transform_features():
-    """Test SmolVLANewLineProcessor transform_features method."""
-    processor = SmolVLANewLineProcessor()
+    """Test NewLineTaskProcessorStep transform_features method."""
+    processor = NewLineTaskProcessorStep()
 
     # Test transform_features
     features = {