Get the video path pattern for the LeRobot dataset.

Factory function that can be used with copytree() ignore parameter. Arguments define a sequence of glob-style patterns that are used to speci

Initialize per-horizon normalizers from statistics.

Create a dummy observation matching AR_droid expectations. AR_droid expects: - 2 external cameras (exterior_image_0_left, exterior_im

Patch older checkpoint configs with locally-added embodiment definitions.

Pad to max_action_dim, return masks.

Tokenize data['language'] (or default_instruction if missing).

Gathers final state from data['state'], then pads to max_state_dim. Return (state, state_mask, n_state_tokens).

Process, stack, and pad images from data['video'].

Process clean image blocks with causal attention pattern - OPTIMIZED First frame: conditioning, cannot attend to anything (self-atten

Process noisy action blocks with teacher forcing pattern - OPTIMIZED First action (for first frame): cannot attend to anything (self-

Process noisy image blocks with teacher forcing pattern - OPTIMIZED First frame: conditioning, cannot attend to anything (self-attent

Process state blocks: self-attention only - OPTIMIZED OPTIMIZATION: State blocks only do self-attention within each block. In

Receive batch data from rank 0 using torch.distributed primitives.

on_red -> red

Apply the focus rectangle transformation to a video. Args: video (np.ndarray): Video tensor of shape [T, H, W, C]

Uniformly sample from language-consistent ranges based on the first index's language. Args: step_indices (np.ndarray): Or

Create and print a visualization of the attention mask pattern. Returns a binary mask [total_len, total_len] where 1 = can attend, 0

Wrap each generator invocation with the context manager

Return running totals # _accumulate([1,2,3,4,5]) --> 1 3 6 10 15 # _accumulate([1,2,3,4,5], operator.mul) --> 1 2 6 24 120

Activate specific components of the action head model.

Activate a model - for VLM this is a no-op, for action_head this manages components.

Describe type, shape, device, data type (of np array/tensor) Very useful for debugging

Apply the transformation to the data corresponding to keys matching the `apply_to` regular expression and return the processed data.

Apply action horizon overrides if action_horizon is specified in config. This function should be called after the config is loaded but before

Check if input is batched and add batch dimension if needed

Build episodes.jsonl entries.

Build the frame index schedule for multi-frame chunks. Returns a list of frame-index lists. Each inner list has 4 indices.

Build the modality.json structure expected by GEAR/DreamZero.

Build an obs dict matching what GrootSimPolicy.forward() expects.

Build tasks.jsonl entries from the dataset.

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

Colorize text. Available text colors: red, green, yellow, blue, magenta, cyan, white. Available text highlights: on_red, on_

Compute relative-action statistics: (action - reference_state) for each key. This replicates the logic in groot/vla/data/dataset/lerobot.py _

Compute mean/std/min/max/q01/q99 for numeric columns across all episodes.

Compute mean/std and min/max statistics of all data keys in a LeRobotDataset.

Convert DROID 1.0.1 RLDS dataset to LeRobot format with idle filtering. Args: raw_dir: Path to raw DROID RLDS dataset (e.g., ./data/

r""" Convert the model output to the corresponding type the UniPC algorithm needs. Args: model_output (`torch.Tensor`):

Deepcopy the nested structure, but does NOT copy the leaf values like Tensors

Delete the cached shard.

Return categorised feature names from info.json.

Detect padding in the image by finding non-black regions. Returns slices for the content region (y_slice, x_slice).

Enable vram management for the action head model.

Match OpenVLA LIBERO action encoding: 0 = close, 1 = open. Raw LIBERO demos store the gripper command in the last action dimension with appro

Encode a sequence of frames to a video file using PyAV.

Ensure the model is loaded on GPU before inference.

Supports both dir and file. Supports glob wildcard.

If exist, remove. Supports both dir and file. Supports glob wildcard.