Processes an 'info' dictionary. Subclasses must implement this method. Args: info: The input 'info' dictionary from the transitio

Get the next action from the queue. Returns: Tensor | None: The next action (action_dim,) or None if queue is empty.

Create an `EnvTransition` dictionary with sensible defaults. Args: observation: Observation dictionary. action: Action dicti

Override to method to ensure all components move together.

Override to return trainable parameters from both models.

Check if the queue is empty. Returns: bool: True if no actions remain, False otherwise.

Saves a transition, ensuring tensors are stored on the designated storage device.

Retrieves a processor step class from the registry by its name. Args: name: The name of the step to retrieve. Returns:

Return the maximum latency or None if empty.

Return the number of cached decoders.

Set evaluation mode for both models.

Fixture to provide the XVLA preprocessor for tests.

Convert an `EnvTransition` back to the canonical batch format used in LeRobot. This is the inverse of `batch_to_transition`. Args:

This function only adds the frame to the episode_buffer. Apart from images — which are written in a temporary directory — nothing is

This will save to disk the current episode in self.episode_buffer. Video encoding is handled automatically based on batch_encoding_s

Write a value to a single motor.

Merge new actions into the queue. This method operates differently based on RTC mode: - RTC enabled: Replaces the queue, accounting f

Saves the pipeline's configuration and state to a directory. This method creates a JSON configuration file that defines the pipeline's struct

Get the number of remaining actions in the queue. Returns: int: Number of unconsumed actions.

Update the running statistics with a batch of vectors. Args: batch: An array where all dimensions except the last are batch dimen

Convert various data types to PyTorch tensors with configurable options. This is a unified tensor conversion function using single dispatch

Track debug information for a denoising step at a given time. If a step with the given time already exists, it will be updated with the new d

Clear the cache and close file handles.

PolicyFeature factory

Save object in local directory. Args: save_directory (`str` or `Path`): Path to directory in which the o

Set seed for reproducibility.

Establish communication with the robot. Args: calibrate (bool): If True, automatically calibrate the robot after connect

Close the parquet writers. This function needs to be called after data collection/conversion, else footer metadata won't be written to the pa

Disconnect from the robot and perform any necessary cleanup.

Initialize logging configuration for LeRobot. In multi-GPU training, only the main process logs to console to avoid duplicate output. Non-mai

Upload model checkpoint to the Hub. Use `allow_patterns` and `ignore_patterns` to precisely filter which files should be pushed to t

Create a LeRobot Dataset from scratch in order to record data.

Create or load pre- and post-processor pipelines for a given policy. This function acts as a factory. It can either load existing processor

Download the object from the Huggingface Hub and instantiate it. Args: pretrained_name_or_path (`str`, `Path`):

Normalize stage+tau reward to [0, 1] with custom breakpoints. Maps stage index + within-stage tau to normalized progress [0, 1]. The bre

Set training mode for both models.

Sample a random batch of transitions and collate them into batched tensors.

This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to the docstring of this

Encodes an observation into its latent representation.

Get all recorded debug steps. Returns: List of all DebugStep objects (may be empty if disabled).

Moves the processor's normalization stats to the specified device. Args: device: The target PyTorch device. Ret

Find stage and within-stage progress (tau) for a frame. Args: current_frame: Frame index relative to episode start episode_length

Builds a "Write" instruction. The parameters for Write are: param[0] = start_address L param[1] = start_

Compute the loss for the given model Args: batch: Dictionary containing: - action: Action tensor

Wait for `seconds` with better precision than time.sleep alone at the expense of more CPU usage. Parameters: - seconds: duration to wa

Read a value from multiple motors.

Write a value to a single motor. Positions are always in degrees. Can write 'Goal_Position', 'Kp', or 'Kd'.

Write calibration parameters to the motors.

Aggregates and filters pipeline features to create a dataset-ready features dictionary. This function transforms initial features using the

Compute comprehensive statistics for array features along specified axes. This function calculates min, max, mean, std, and quantiles (1%, 10%, 5

Convert hardware-specific features to a LeRobot dataset feature dictionary. This function takes a dictionary describing hardware outputs (like jo

Fixture to provide the PI0Fast preprocessor for tests.

Close environment and disconnect robot.

Disable torque on selected motors.

Add multiple features to a LeRobotDataset in a single pass. This is more efficient than calling add_feature() multiple times, as it only copi

Heuristic to estimate the number of samples based on dataset size. The power controls the sample growth relative to dataset size. Lower the po