Undo action normalization and convert relative→absolute. Args: action: Normalized action array of shape (..., action_horizon, act

Validate configuration.

Run the verification. Returns True if with_correction is the better match.

Run ONNX Runtime inference and compare against PyTorch output. Returns dict of {output_name: cosine_similarity}. Requires onnxruntime-gpu; sk

Forward pass through TensorRT DiT.

Pad image to square dimensions by adding black bars to the smaller dimension. Args: img: Image tensor of shape (..., C,

Process a list of messages and return a dictionary of model inputs. Args: messages (list[dict[str, Any]]): List of messa

Cleanup resources on destruction

Cleanup resources on destruction

Get a pose by index

Load complete episode data as a processed DataFrame. Combines parquet data loading and video decoding to create a unified DataFrame

Initialize the wrapper around a Gr00tPolicy instance. Args: policy: The Gr00tPolicy instance to wrap strict: Whether

Initialize the Replay Policy. Args: dataset_path: Path to the LeRobot-style dataset directory modality_configs: Dicti

Qwen3Backbone is to generate n_queries to represent the future action hidden states. Args: model_name: nvidia/Cosmos-Reas

Initialize an end-effector pose. Args: translation: Translation vector [x, y, z] rotation: Rotation in speci

Initialize a joint trajectory from a list of joint poses. Args: poses: Sequence of JointPose objects times:

Initialize an end-effector trajectory from a list of end-effector poses. Args: poses: Sequence of EndEffectorPose object

Initialize unified state and action processor. Args: modality_configs: Nested dict with structure: {embo

Initialize mixture dataset with datasets, weights, and configuration.

Initialize single-step dataset with sharding configuration.

Initialize LeRobot episode loader with dataset path and modality configurations. The initialization process involves: 1. Loa

When file_path is None, don't record.

video_delta_indices: np.ndarray[int], please check `assert_delta_indices` to see the requirements state_delta_indices: np.ndarray[int

Args: run_name: Name of the experiment run exp_cfg_dir: Path to the directory containing all experiment metadata

Initialize the trainer. Args: *args: Positional arguments forwarded to ``Trainer``.

Iterate over the mixture dataset with background shard caching. Implements an efficient iteration strategy: 1. Filter shards

Return the number of shards.

Allow len(): len(joint_pose) returns number of joints

Return the number of poses in the action chunking

Return the number of shards in the dataset.

Return number of episodes in dataset.

Set default values for action-related fields if not specified.

Validate fields and set default values.