Convert action chunking to the specified action format. Must be implemented by subclasses. Args: action_format:

Resolve a string to an EmbodimentTag, case-insensitively. Matches by enum **name** first (e.g. ``"xdof"`` -> ``XDOF``), then by enum

Load configuration from YAML file.

Set runtime input shape for dynamic dimensions.

Replace ``old`` with ``new`` in ``text``, asserting exactly one occurrence.

Return the Nth block (1-based) whose code contains keyword.

Return the root directory of the repository.

Deserialize and load a TensorRT engine from file.

Create a deep copy of this pose. Must be implemented by subclasses. Returns: New Pose instance with copied data

Context manager that prints the wall-clock duration of a labelled phase. Usage:: with timed("model load"): model = load_mode

Run a subprocess step with consistent timing/logging/failure formatting.

Extract all fenced code blocks from a markdown file with context.

Run bash/sh code blocks extracted from a README via subprocess. If ``force_yes`` is True, ``apt-get`` commands are automatically given the ``

Create a random EndEffectorPose with valid rotation.

Call an endpoint on the server. Args: endpoint: The name of the endpoint. data: The input data for the endpo

Create an EndEffectorPose from a flat array using the specified action format. This is the inverse of the xyz_rot6d / xyz_rotvec / h

Get the EmbodimentTag for a gym-registered environment name. Looks up the env_name prefix (before "/") in ENV_PREFIX_TO_EMBODIMENT_TAG. Falls

Build a synthetic observation that matches the policy's modality config. Uses a fixed seed so that failures are reproducible.

Create an EndEffectorActionChunk from a 2-D array using the specified action format. This is the inverse of ``.to(action_format)``.

Compute and return the next action based on current observation with validation. This is the main public interface. It validates the observat

Apply state processing (normalization, encoding). Args: state: Dict mapping joint_group -> raw state values

Save configuration to YAML file.

Get rotation in specified representation. Args: rotation_type: Desired type ("quat", "euler", "rotvec", "matrix", "rot6d

Load and check the exported ONNX model for validity.

Create dummy input tensors matching the model's expected format.

Create random state dict with dimensions matching statistics.

Calculate the dataset statistics of all columns for a list of parquet files. Args: parquet_paths (list[Path]): List of paths to parquet f

Load MODALITY_CONFIGS from embodiment_configs.py as serializable dicts.

Resolve the video backend. torchcodec is the only officially supported backend. Other backends (decord, ffmpeg, opencv, pyav, torchvision_av)

Load modality configs and statistics from test fixtures.

Create synthetic VLAStepData matching the fixture config.

Create random action dict with dimensions matching statistics.

Build the dataset. Returns a tuple of (train_dataset, eval_dataset).

Call an endpoint on the server. Args: endpoint: The name of the endpoint. data: The input data for the endpo

Return True if *path* looks like a bundled LeRobot demo dataset. Checks for ``meta/modality.json``, at least one ``.parquet`` under ``data/``,

Prepare inputs for the model, mimicking what happens inside _get_action. Returns collated_inputs that can be passed to model.get_action()

Register a new endpoint to the server. Args: name: The name of the endpoint. handler: The handler function t

Compute the relative action chunking with respect to a reference frame. If reference_frame is None, uses the first pose in the actio

Set dataset statistics for normalization. Args: statistics: Nested dict with structure: {embodiment_tag:

Recursively convert dataclasses and numpy arrays to JSON-serializable format. Args: obj: Object to convert (can be dataclass, numpy

Return True when the LIBERO dataset looks complete enough to reuse.

Lazily import decord, raising ImportError if unavailable.

Lazily import torchcodec, raising ImportError if unavailable.

Redirect stdout, stderr, and all logging to the log file. The caller's own progress prints should happen *outside* this context. If *tee* is

Apply action processing (absolute->relative conversion, normalization). Processing order: 1. Convert absolute actions to rel

Raise AssertionError if the port is already bound. Call this before starting a model server subprocess to catch port conflicts early (e.g. a

Return True if *path* looks like a HuggingFace ``transformers`` checkpoint dir.

Return the expected PyTorch dtype for a named input tensor.

Get default configuration.

Kill the server.

Compute overall statistics from per-dataset statistics using weighted averaging. This function combines statistics from multiple datasets ac

Prepare inputs for backbone and action head.

Return a shared model path, downloading once if not present. Models are stored at ``TEST_CACHE_PATH/models/<repo_name>/`` so all tests share

Map a tag value string back to its enum name, or return the value as-is.

Huggingface will call model.train() at each training_step. To ensure the expected behaviors for modules like dropout, batchnorm, etc.

Load TRT engines, delete PyTorch modules, and monkey-patch forward methods. Args: policy: Gr00tPolicy instance trt_engine_path: P