Helper method to apply gradient checkpointing if enabled.

Validate that all statistics have correct types and shapes. Args: stats_list: List of statistics dictionaries to validate Raises:

Close and cleanup the parquet writer if it exists.

Preprocess images for the model. Images from LeRobot are typically in [B, C, H, W] format and normalized to [0, 1]. PaliGemma expects

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,

Check if delta timestamps are multiples of 1/fps +/- tolerance. This ensures that adding these delta timestamps to any existing timestamp in

Create a default SmolVLA configuration for testing.

Create dummy data for testing both implementations.

Apply one denoising step of the noise `x_t` at a given timestep.

https://en.wikipedia.org/wiki/Signed_number_representations#Sign%E2%80%93magnitude

https://en.wikipedia.org/wiki/Signed_number_representations#Two%27s_complement

Establish a connection with the learner. Args: stub (services_pb2_grpc.LearnerServiceStub): The stub to use for the connection. s

Run the batch through the model and compute the loss for training. Args: batch: Training batch containing observations and action

Run the batch through the model and compute the loss for training. Args: batch: Training batch containing observations and action

Given a string, return a torch.device with checks on whether the device is available.

Loads the step's state from a state dictionary. Args: state: A dictionary of state tensors.

Create optimizers for the SAC policy.

Builds a 'Ping' status packet. Args: scs_id (int): ID of the servo responding. error (int, optional): Error to be ret

Save optimizer state to disk. Args: optimizer: Either a single optimizer or a dictionary of optimizers. save_dir: Directory to sa

Set a new UART baud-rate on the port. Args: baudrate (int): Desired baud-rate in bits / second. Raises: Runt

Convert model state dict to flat array for transmission

Update the target model's parameters with an EMA step.

Update configuration parameters. Args: **kwargs: Key-value pairs to update Returns: X2RDataProcessingConfig:

Saves a NumPy array or PIL Image to a file. This function handles both NumPy arrays and PIL Image objects, converting the former to a PI

Write data to a JSON file. Creates parent directories if they don't exist. Args: data (dict): The dictionary to write. fpath

Predict state-action value for all of the learned Q functions. Args: z: (*, latent_dim) tensor for the current state's latent rep

GPT model gets hyperparameters from a config object. Please refer configuration_vqbet.py for more details.

Create mock lerobot features mapping similar to what hw_to_dataset_features returns.

Create mock policy image features with different resolutions.

Create a mock robot observation with motor positions and camera images.

Convert float to unsigned integer for CAN transmission.

Get motor name from name or ID.

Generate `count` consecutive TimedAction objects starting at timestep `start_t`.

Normalize possible return types from hub `make_env` into the mapping: { suite_name: { task_id: vector_env } } Accepts: - dict (as

Resolve step class from registry or import path. This method implements a two-tier resolution strategy: **Tier 1: Registry-based res

Converts an unsigned integer value into a list of byte-sized integers to be sent via a communication protocol. Depending on the proto

Generate rewind frame indices for temporal augmentation. Rewind simulates going backwards through previously seen frames, starting from

Applies RoPE positions [B, L] to x [B, L, H, D].

Build multiple Adam optimizers. Args: params: Must be a dict[str, Iterable[Parameter]] mapping parameter group names

Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset. Parameters: - hf_datase

Compute absolute frame indices with clamping for bidirectional observation sequence. Bidirectional sampling centered on target frame: - Befor

Compute ACP indicator ratio from dataset metadata or by scanning the indicator column.

Connects to the RealSense camera specified in the configuration. Initializes the RealSense pipeline, configures the required streams

Create default dataset statistics for testing.

Enable torque on selected motors.

Run the batch through the model and compute the loss for training.

Check if debug tracking is enabled. Returns True if tracker exists and is enabled.

Returns a list of all registered processor step names.

Logs observation and action data to Rerun for real-time visualization. This function iterates through the provided observation and action di

Converts a policy's output tensor into a dictionary of named actions. This function translates the numerical output from a policy model into a

Create robot environment from configuration. Args: cfg: Environment configuration. Returns: Tuple of (gym environment, teleo

Forward pass of the classifier for inference.