pre-commit

src/lerobot/configs/policies.py

@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy.
+        output_shapes: A dictionary defining the shapes of the output data for the policy.
+        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
+            normalization mode to apply.
+        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
+            the original scale.
     """
 
     n_obs_steps: int = 1

src/lerobot/envs/configs.py

@@ -205,7 +205,6 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 

src/lerobot/policies/act/configuration_act.py

@@ -28,7 +28,7 @@ class ACTConfig(PreTrainedConfig):
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and 'output_shapes`.
 
     Notes on the inputs and outputs:
         - Either:
@@ -48,12 +48,21 @@ class ACTConfig(PreTrainedConfig):
             This should be no greater than the chunk size. For example, if the chunk size size 100, you may
             set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
             environment, and throws the other 50 out.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and `output_shapes`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -48,12 +48,21 @@ class DiffusionConfig(PreTrainedConfig):
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.
@@ -64,7 +73,7 @@ class DiffusionConfig(PreTrainedConfig):
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
-        use_separate_rgb_encoder_per_camera: Whether to use a separate RGB encoder for each camera view.
+        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
         down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
             You may provide a variable number of dimensions, therefore also controlling the degree of
             downsampling.

src/lerobot/policies/tdmpc/configuration_tdmpc.py

@@ -30,7 +30,7 @@ class TDMPCConfig(PreTrainedConfig):
     camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features`, `output_features`, and perhaps `max_random_shift_ratio`.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
 
     Args:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
@@ -40,12 +40,24 @@ class TDMPCConfig(PreTrainedConfig):
             is an alternative to using action repeats. If this is set to more than 1, then we require
             `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
             approach of using multiple steps from the plan is not in the original implementation.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
+            match the original implementation.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
+            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
+            normalization mode here.
         image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
         state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
         latent_dim: Observation's latent embedding dimension.

src/lerobot/policies/vqbet/configuration_vqbet.py

@@ -32,7 +32,7 @@ class VQBeTConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and `output_shapes`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -46,12 +46,21 @@ class VQBeTConfig(PreTrainedConfig):
             current step and additional steps going back).
         n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
         action_chunk_size: Action chunk size of each action prediction token.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy.
+            The key represents the input data name, and the value is a list indicating the dimensions
+            of the corresponding data. For example, "observation.image" refers to an input from
+            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
+            Importantly, shapes doesnt include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy.
+            The key represents the output data name, and the value is a list indicating the dimensions
+            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
+            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.

src/lerobot/processor/env_processor.py

@@ -17,7 +17,7 @@ from dataclasses import dataclass
 
 import torch
 
-from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.utils.constants import OBS_IMAGES, OBS_PREFIX, OBS_STATE, OBS_STR
 
 from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
@@ -92,7 +92,7 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # copy over non-STATE features
         for ft, feats in features.items():
-            if ft != PipelineFeatureType.STATE:
+            if ft != FeatureType.STATE:
                 new_features[ft] = feats.copy()
 
         # rebuild STATE features
@@ -100,13 +100,11 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # add our new flattened state
         state_feats[OBS_STATE] = PolicyFeature(
-            key=OBS_STATE,
+            type=FeatureType.STATE,
             shape=(8,),  # [eef_pos(3), axis_angle(3), gripper(2)]
-            dtype="float32",
-            description=("Concatenated end-effector position (3), axis-angle (3), and gripper qpos (2)."),
         )
 
-        new_features[PipelineFeatureType.STATE] = state_feats
+        new_features[FeatureType.STATE] = state_feats
 
         return new_features
 

src/lerobot/processor/hil_processor.py

@@ -314,7 +314,7 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
 
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
-class GripperPenaltyProcessorStep(ProcessorStep):
+class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
     """
     Applies a penalty for inefficient gripper usage.
 
@@ -329,27 +329,26 @@ class GripperPenaltyProcessorStep(ProcessorStep):
     penalty: float = -0.01
     max_gripper_pos: float = 30.0
 
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
+    def complementary_data(self, complementary_data: dict) -> dict:
         """
         Calculates the gripper penalty and adds it to the complementary data.
 
         Args:
-            transition: The incoming environment transition.
+            complementary_data: The incoming complementary data, which should contain
+                                raw joint positions.
 
         Returns:
-            The modified transition with the penalty added to complementary data.
+            A new complementary data dictionary with the `discrete_penalty` key added.
         """
-        new_transition = transition.copy()
-        action = new_transition.get(TransitionKey.ACTION)
-        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        action = self.transition.get(TransitionKey.ACTION)
 
         raw_joint_positions = complementary_data.get("raw_joint_positions")
         if raw_joint_positions is None:
-            return new_transition
+            return complementary_data
 
         current_gripper_pos = raw_joint_positions.get(GRIPPER_KEY, None)
         if current_gripper_pos is None:
-            return new_transition
+            return complementary_data
 
         # Gripper action is a PolicyAction at this stage
         gripper_action = action[-1].item()
@@ -365,12 +364,11 @@ class GripperPenaltyProcessorStep(ProcessorStep):
 
         gripper_penalty = self.penalty * int(gripper_penalty_bool)
 
-        # Update complementary data with penalty info
+        # Create new complementary data with penalty info
         new_complementary_data = dict(complementary_data)
         new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
-        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
 
-        return new_transition
+        return new_complementary_data
 
     def get_config(self) -> dict[str, Any]:
         """

src/lerobot/rl/gym_manipulator.py

@@ -412,10 +412,7 @@ def make_processors(
         if cfg.processor.observation.add_current_to_observation:
             env_pipeline_steps.append(MotorCurrentProcessorStep(robot=env.robot))
 
-    add_ee_pose = (
-        cfg.processor.observation is not None and cfg.processor.observation.add_ee_pose_to_observation
-    )
-    if kinematics_solver is not None and add_ee_pose:
+    if kinematics_solver is not None:
         env_pipeline_steps.append(
             ForwardKinematicsJointsToEEObservation(
                 kinematics=kinematics_solver,
@@ -438,12 +435,7 @@ def make_processors(
         )
 
     # Add gripper penalty processor if gripper config exists and enabled
-    # Only add if max_gripper_pos is explicitly configured (required for normalization)
-    if (
-        cfg.processor.gripper is not None
-        and cfg.processor.gripper.use_gripper
-        and cfg.processor.max_gripper_pos is not None
-    ):
+    if cfg.processor.gripper is not None and cfg.processor.gripper.use_gripper:
         env_pipeline_steps.append(
             GripperPenaltyProcessorStep(
                 penalty=cfg.processor.gripper.gripper_penalty,

src/lerobot/rl/wandb_utils.py

@@ -26,21 +26,8 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.constants import PRETRAINED_MODEL_DIR
 
 
-def cfg_to_group(
-    cfg: TrainPipelineConfig, return_list: bool = False, truncate_tags: bool = False, max_tag_length: int = 64
-) -> list[str] | str:
+def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[str] | str:
     """Return a group name for logging. Optionally returns group name as list."""
-
-    def _maybe_truncate(tag: str) -> str:
-        """Truncate tag to max_tag_length characters if required.
-
-        wandb rejects tags longer than 64 characters.
-        See: https://github.com/wandb/wandb/blob/main/wandb/sdk/wandb_settings.py
-        """
-        if len(tag) <= max_tag_length:
-            return tag
-        return tag[:max_tag_length]
-
     lst = [
         f"policy:{cfg.policy.type}",
         f"seed:{cfg.seed}",
@@ -49,8 +36,6 @@ def cfg_to_group(
         lst.append(f"dataset:{cfg.dataset.repo_id}")
     if cfg.env is not None:
         lst.append(f"env:{cfg.env.type}")
-    if truncate_tags:
-        lst = [_maybe_truncate(tag) for tag in lst]
     return lst if return_list else "-".join(lst)
 
 
@@ -98,7 +83,7 @@ class WandBLogger:
             entity=self.cfg.entity,
             name=self.job_name,
             notes=self.cfg.notes,
-            tags=cfg_to_group(cfg, return_list=True, truncate_tags=True),
+            tags=cfg_to_group(cfg, return_list=True),
             dir=self.log_dir,
             config=cfg.to_dict(),
             # TODO(rcadene): try set to True

src/lerobot/scripts/lerobot_record.py

@@ -184,9 +184,6 @@ class DatasetRecordConfig:
     vcodec: str = "libsvtav1"
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
-    # Expert noise injection scale. Noise is added to robot actions but not recorded in dataset.
-    # This forces recovery behavior for more robust learned policies. 0.0 means no noise. #https://arxiv.org/pdf/1703.09327, https://arxiv.org/abs/2507.09061
-    noise_scale: float = 0.0
 
     def __post_init__(self):
         if self.single_task is None:
@@ -286,7 +283,6 @@ def record_loop(
     single_task: str | None = None,
     display_data: bool = False,
     display_compressed_images: bool = False,
-    noise_scale: float = 0.0,
 ):
     if dataset is not None and dataset.fps != fps:
         raise ValueError(f"The dataset fps should be equal to requested fps ({dataset.fps} != {fps}).")
@@ -384,27 +380,18 @@ def record_loop(
             action_values = act_processed_teleop
             robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
 
-        # Write clean action to dataset (before noise injection)
-        if dataset is not None:
-            action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
-            frame = {**observation_frame, **action_frame, "task": single_task}
-            dataset.add_frame(frame)
-
-        # Expert noise injection: add noise to motor commands but not to recorded labels
-        if noise_scale > 0:
-            import torch
-
-            for key in robot_action_to_send:
-                if isinstance(robot_action_to_send[key], torch.Tensor):
-                    noise = torch.randn_like(robot_action_to_send[key]) * noise_scale
-                    robot_action_to_send[key] = robot_action_to_send[key] + noise
-
         # Send action to robot
         # Action can eventually be clipped using `max_relative_target`,
         # so action actually sent is saved in the dataset. action = postprocessor.process(action)
         # TODO(steven, pepijn, adil): we should use a pipeline step to clip the action, so the sent action is the action that we input to the robot.
         _sent_action = robot.send_action(robot_action_to_send)
 
+        # Write to dataset
+        if dataset is not None:
+            action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+            frame = {**observation_frame, **action_frame, "task": single_task}
+            dataset.add_frame(frame)
+
         if display_data:
             log_rerun_data(
                 observation=obs_processed, action=action_values, compress_images=display_compressed_images
@@ -523,7 +510,6 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     single_task=cfg.dataset.single_task,
                     display_data=cfg.display_data,
                     display_compressed_images=display_compressed_images,
-                    noise_scale=cfg.dataset.noise_scale,
                 )
 
                 # Execute a few seconds without recording to give time to manually reset the environment