Get short device name for table.

Get the length of a specific episode.

Compute original episode lengths. Returns: List of original episode lengths

Create Ray environment factory function without wrappers.

Load initial actions for policy initialization if available. Returns: List containing initial action dictionaries, or em

Get the maximum number of steps that we need to cache.

Get the modality configurations. Returns: dict[str, dict[str, ModalityConfig]]: The modality configurations, where

Get the number of timesteps in a specific shard.

Return True if any available GPU has RT cores (required for Vulkan ray tracing). Checks ``nvidia-smi`` GPU names against known RT-capable product

Build a ``uv run python -c`` command that downloads a file from HuggingFace. Reads HF_TOKEN from the environment and passes it explicitly so gate

Flips the sign of the gripper action (last dimension of action vector). This is necessary for some environments where -1 = open, +1 = close,

Invert a homogeneous transformation matrix. Args: T: A 4x4 homogeneous transformation matrix Returns: The inverse of th

Return a clean env for running isaaclab.sh commands. Strips VIRTUAL_ENV, UV_PROJECT_ENVIRONMENT, and the workspace venv bin directory from PA

Resolve the local dataset directory and return all video files within it.

Load the consolidated metadata rows stored in ``meta/episodes``.

initial_actions: list[dict[str, dict[str, np.ndarray]]] 0: (the first dataset) trajectory_name: action_key: act

Merge dataset statistics across all datasets, grouped by embodiment. Combines statistics from datasets with the same embodiment tag

Recursively converts bottom-level list of lists to NumPy arrays. Args: data: A nested dictionary where bottom nodes are list of list

Changes gripper action (last dimension of action vector) from [0,1] to [-1,+1]. Necessary for some environments (not Bridge) because the data

Convert raw robot observation dict into the structured GR00T VLA input.

Plot and save trajectory results comparing ground truth and predicted actions. Args: state_joints_across_time: Array of state joints

Plot and save trajectory results comparing ground truth and predicted actions. Args: state_joints_across_time: Array of state joints

Temporarily prevent keyboard interrupts by delaying them until after the protected code.

Display engine details (inputs/outputs) on rank 0 only.

Print results as a markdown table using median for robustness.

Copied from robosuite: https://github.com/ARISE-Initiative/robosuite/blob/eafb81f54ffc104f905ee48a16bb15f059176ad3/robosuite/utils/transform_util

Compute the relative transformation between two poses. Args: T0: Initial 4x4 homogeneous transformation matrix Tt: Current 4

Reset the policy to its initial state. Args: options: Dictionary containing the options for the reset Returns:

Reset the policy to its initial state. Args: options: Dictionary containing the options for the reset Returns:

Reset the random seed and regenerate sampling schedules. Used for deterministic training restarts or seed changes during training.

Resolve the LIBERO-finetuned GR00T-N1.7 checkpoint (``libero_10`` subfolder). Thin wrapper around :func:`resolve_model_checkpoint_path` kept for

Return a writable uv cache path, or None. Only redirects the uv cache when TEST_CACHE_PATH is set — on dev machines uv's default cache (~/.ca

Execute Python code blocks extracted from a README.

Run policy rollouts in parallel environments. Args: env_name: Name of the gymnasium environment to use policy: Policy instance

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

Set seed for all random number generators.

Set a global timeout (Linux only). Raises TimeoutError when exceeded.

Set a global timeout (Linux only). Raises TimeoutError when exceeded.

Configure logging.

Create balanced shards by distributing episode timesteps across shards. The sharding process: 1. Shuffle episode order for r

Swish activation function.

Convert a relative joint action chunking to an absolute action chunking by adding the relative joint positions to the reference frame

Convert trajectory to array of joint values. Returns: Array with shape (N, num_joints) where N is the number of poses

Convert trajectory to array of homogeneous transformation matrices. Returns: Array of homogeneous matrices with shape (N

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

Convert trajectory to array of translations and 6D rotations. Returns: Array with shape (N, 9) - 3 for xyz + 6 for rot6d

Convert trajectory to array of translations and rotation vectors. Returns: Array with shape (N, 6) - 3 for xyz + 3 for r

Convert TensorRT data type to PyTorch equivalent.