mirror of
https://github.com/huggingface/lerobot.git
synced 2026-08-08 17:39:44 +00:00
CarolinePascal
06364ce436
Documents every remaining public class/function across the 15 files in lerobot.processor (pipeline, batch/device/normalize/rename/observation steps, env-specific and HIL steps, action bridges, tokenization, language, converters, factory functions, and the normalization migration script), reformats pre-existing docstrings to the machine-checked Args:/Returns:/Raises: standard, and expands docs/source/api/processor.mdx from 3 documented classes to the full public surface. Adds lerobot.processor to check_docstrings.py's MODULES_TO_CHECK ratchet and removes the module's ruff D-ignore. Docstrings and doc comment reformatting only; no behavioral changes.
256 lines
8.2 KiB
Plaintext
256 lines
8.2 KiB
Plaintext
# Processors
|
|
|
|
Processors are the data transformation layer between a robot, a dataset and a policy. A pipeline is a chain
|
|
of [`ProcessorStep`]s; each step declares how it transforms both the data and the feature contract.
|
|
|
|
See [Introduction to Robot Processors](../introduction_processors) for the concepts,
|
|
[Implement your own processor](../implement_your_own_processor) to write a step, and
|
|
[Debug your processor pipeline](../debug_processor_pipeline) when a pipeline misbehaves.
|
|
|
|
## Core pipeline
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.DataProcessorPipeline
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.PolicyProcessorPipeline
|
|
|
|
`RobotProcessorPipeline` is a type alias for `DataProcessorPipeline[TInput, TOutput]`, used for the
|
|
teleop-action, robot-action and robot-observation pipelines that don't go through a policy.
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ProcessorKwargs
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ProcessorStepRegistry
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ProcessorMigrationError
|
|
|
|
### Typed base steps
|
|
|
|
Each subclasses `ProcessorStep` to implement one part of an `EnvTransition` (observation, action, reward,
|
|
etc.), leaving the rest of the transition untouched by default.
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ObservationProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ActionProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.RobotActionProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.PolicyActionProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.RewardProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.DoneProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.TruncatedProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.InfoProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.ComplementaryDataProcessorStep
|
|
|
|
[[autodoc]] lerobot.processor.pipeline.IdentityProcessorStep
|
|
|
|
## Factory functions
|
|
|
|
Build the canonical processor pipelines used by policies and robots.
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_processors
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_teleop_action_processor
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_robot_action_processor
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_robot_observation_processor
|
|
|
|
[[autodoc]] lerobot.processor.factory.DefaultPolicyProcessorSteps
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_policy_processor_steps
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_policy_processor_pipelines
|
|
|
|
[[autodoc]] lerobot.processor.factory.make_default_pre_post_processors
|
|
|
|
## Batch dimension
|
|
|
|
[[autodoc]] lerobot.processor.batch_processor.AddBatchDimensionProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.batch_processor.AddBatchDimensionObservationStep
|
|
|
|
[[autodoc]] lerobot.processor.batch_processor.AddBatchDimensionActionStep
|
|
|
|
[[autodoc]] lerobot.processor.batch_processor.AddBatchDimensionComplementaryDataStep
|
|
|
|
## Device
|
|
|
|
[[autodoc]] lerobot.processor.device_processor.DeviceProcessorStep
|
|
- all
|
|
|
|
## Normalization
|
|
|
|
[[autodoc]] lerobot.processor.normalize_processor.NormalizerProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.normalize_processor.UnnormalizerProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.normalize_processor.hotswap_stats
|
|
|
|
## Renaming
|
|
|
|
[[autodoc]] lerobot.processor.rename_processor.RenameObservationsProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.rename_processor.rename_stats
|
|
|
|
## Vanilla observation processing
|
|
|
|
Converts standard Gymnasium observations (`pixels`, `agent_pos`, `environment_state`) to the LeRobot format.
|
|
|
|
[[autodoc]] lerobot.processor.observation_processor.VanillaObservationProcessorStep
|
|
- all
|
|
|
|
## Environment-specific observation processing
|
|
|
|
[[autodoc]] lerobot.processor.env_processor.LiberoProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.env_processor.IsaaclabArenaProcessorStep
|
|
- all
|
|
|
|
## NumPy / PyTorch action conversion
|
|
|
|
[[autodoc]] lerobot.processor.gym_action_processor.Torch2NumpyActionProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.gym_action_processor.Numpy2TorchActionProcessorStep
|
|
- all
|
|
|
|
## Delta and relative actions
|
|
|
|
[[autodoc]] lerobot.processor.delta_action_processor.MapTensorToDeltaActionDictStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.delta_action_processor.MapDeltaActionToRobotActionStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.relative_action_processor.RelativeActionsProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.relative_action_processor.AbsoluteActionsProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.relative_action_processor.to_relative_actions
|
|
|
|
[[autodoc]] lerobot.processor.relative_action_processor.to_absolute_actions
|
|
|
|
## Policy/robot action bridge
|
|
|
|
Converts between a robot's per-motor action dict and a policy's stacked action tensor.
|
|
|
|
[[autodoc]] lerobot.processor.policy_robot_bridge.RobotActionToPolicyActionProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.policy_robot_bridge.PolicyActionToRobotActionProcessorStep
|
|
- all
|
|
|
|
## Tokenization
|
|
|
|
[[autodoc]] lerobot.processor.tokenizer_processor.TokenizerProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.tokenizer_processor.ActionTokenizerProcessorStep
|
|
- all
|
|
|
|
## Language
|
|
|
|
[[autodoc]] lerobot.processor.newline_task_processor.NewLineTaskProcessorStep
|
|
- all
|
|
|
|
`RenderMessagesStep` requires the `[dataset]` extra and is not re-exported from `lerobot.processor`; import
|
|
it directly from `lerobot.processor.render_messages_processor`.
|
|
|
|
[[autodoc]] lerobot.processor.render_messages_processor.RenderMessagesStep
|
|
- all
|
|
|
|
## Human-in-the-loop (HIL)
|
|
|
|
Steps supporting human-in-the-loop RL: teleop event/action bookkeeping, time limits, image preprocessing,
|
|
the `gym-hil` adapter, gripper penalties, intervention handling, and a learned reward classifier.
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.HasTeleopEvents
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.AddTeleopActionAsComplimentaryDataStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.AddTeleopEventsAsInfoStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.ImageCropResizeProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.TimeLimitProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.GymHILAdapterProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.GripperPenaltyProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.InterventionActionProcessorStep
|
|
- all
|
|
|
|
[[autodoc]] lerobot.processor.hil_processor.RewardClassifierProcessorStep
|
|
- all
|
|
|
|
## Transition converters
|
|
|
|
Convert between an `EnvTransition` and the raw dict formats used by robots, policies, and dataset batches.
|
|
|
|
[[autodoc]] lerobot.processor.converters.create_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.identity_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.to_tensor
|
|
|
|
[[autodoc]] lerobot.processor.converters.from_tensor_to_numpy
|
|
|
|
[[autodoc]] lerobot.processor.converters.robot_action_to_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.robot_action_observation_to_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.observation_to_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.policy_action_to_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.batch_to_transition
|
|
|
|
[[autodoc]] lerobot.processor.converters.transition_to_robot_action
|
|
|
|
[[autodoc]] lerobot.processor.converters.transition_to_policy_action
|
|
|
|
[[autodoc]] lerobot.processor.converters.transition_to_observation
|
|
|
|
[[autodoc]] lerobot.processor.converters.transition_to_batch
|
|
|
|
## Migrating legacy policies
|
|
|
|
Standalone script to migrate a pretrained policy with built-in normalization layers to the processor
|
|
pipeline system. See the module docstring for CLI usage.
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.extract_normalization_stats
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.detect_features_and_norm_modes
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.remove_normalization_layers
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.clean_state_dict
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.load_state_dict_with_missing_key_handling
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.convert_features_to_policy_features
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.display_migration_summary_with_warnings
|
|
|
|
[[autodoc]] lerobot.processor.migrate_policy_normalization.load_model_from_hub
|