Resolve a GR00T model checkpoint, downloading from HuggingFace if needed. Resolution order: 1. Environment variable *path_override_env* (mus

Run decode+encode cycles and track memory.

Combine blocks (CodeBlock or raw string) and execute as a single Python script.

Run inference on a single trajectory. Args: skip_timing_steps: Number of initial inference steps to skip when calculating timing sta

Set random seed for reproducibility.

Set normalization statistics.

Convert a relative action chunking to an absolute action chunking by applying the relative poses on top of a reference frame.

Return a dictionary representation of this config, optionally excluding augmentation keys.

Return a JSON string of this config, optionally excluding augmentation keys.

Correctly set self.state from checkpoint so get_train_dataloader can read from it.

Reverse state processing (denormalization). Args: state: Dict mapping joint_group -> processed state values

Mean-std unnormalization (reverse z-score normalization). Args: normalized_values: Normalized input values (z-scores) -

Min-max unnormalization from [-1, 1] range back to original range. Args: normalized_values: Normalized input values in [-1, 1] range

Allow indexing: joint_pose[0] returns first joint value

Initialize the Gr00t Policy. Args: embodiment_tag: The embodiment tag defining the robot/environment type. Accept

Initialize Gr00tN1d7 model. Args: config: Model configuration transformers_loading_kwargs: Dict with transfo

Initialize a joint pose. Args: joints: Joint angles/positions as array-like of shape (n,) joint_names: Optio

input_pix_fmt: rgb24, bgr24 see https://github.com/PyAV-Org/PyAV/blob/bc4eedd5fc474e0f25b22102b2771fe5a42bb1c7/av/video/frame.pyx#L352

Assemble the downloaded files into a proper LeRobot v2 demo dataset.

Compute token-level accuracy, ignoring ``-100`` label positions. Args: preds: Predicted token ids of shape ``(batch, seq_len)``.

Remove the dataset directory created by the SO100 workflow.

Compute homogeneous transformation matrix

Compute and cache normalization parameters from statistics for all embodiments and modalities.

Internal method to compute relative transformation. Must be implemented by subclasses. Args: other: The referenc

Set shared cache env vars in os.environ once for the whole test session. HF cache dirs are content-addressed, so all test groups safely share one

Convert relative action to absolute action using reference state.

Convert absolute action to relative action using reference state.

Create appropriate dataset based on task and mode.

Create appropriate dataset based on task and mode.

Setup model with proper vocabulary expansion.

Check that the demo dataset exists and its parquet files are readable.

ONNX-friendly deepstack injection using torch.where. Original uses boolean indexing: hidden_states[visual_pos_masks, :] +

Return the cache root directory.

Download statistics.json and extract eef_9d stats for DROID.

Extract all frames and their timestamps using ffmpeg.

Extract frames at specific timestamps using ffmpeg.

Extract specific frames using ffmpeg.

Extract specific joint groups from data arrays based on modality metadata. Uses the modality metadata to slice the appropriate indic

Initialize from homogeneous transformation matrix

Compute and return the next action based on current observation. This method should be overridden by subclasses to implement policy-specific

Compute and return the next action based on current observation. This method should be overridden by subclasses to implement policy-specific

Check if a video backend is available without importing at module level.

Return True when the LIBERO submodule is properly git-initialized.

Load observation.state.cartesian_position from all episode parquets.

Load mask data for all configured mask views at specified indices.

Load masks from npz/npy file at specified indices.

Load all metadata files including dataset statistics. Parses the standard LeRobot metadata structure: - info.json: Dataset c

Load and process parquet data for a specific episode. Handles the complete data loading pipeline: 1. Load raw parquet file b

Load video data for all configured camera views at specified indices. Uses the configured video backend to decode video frames at th

Return a mock backbone that produces correctly-shaped outputs.

Create an ONNX-exportable attention forward for a single VisionAttention. Three key changes from the original: 1. Replaces cu_seqlens-based s

Return a minimal config for fast instantiation.

Convert rotation matrix to 6D rotation representation. Args: rotation_matrix: Rotation matrix (3, 3) Returns:

Return True if *path* is a real parquet file (not a Git LFS pointer stub).

Parse and validate modality configurations, filling in defaults where needed. For missing modality configs, creates default configur

Patch from_pretrained to prefer the local HF snapshot cache over network calls. When a HF repo ID is passed we try snapshot_download(local_files_

Suppress 429 / connection errors from the HuggingFace Hub in mistral regex patching. transformers calls model_info() inside a nested is_base_mist

Monkey-patch all VisionAttention.forward with ONNX-friendly versions. Args: vision_model: The vision model whose attention blocks to patc

Populate the LIBERO spatial dataset once on shared storage and reuse it.

Populate external_dependencies/LIBERO, reusing shared cache when available. The cache stores both the working tree (which includes the .git point

Set up the libero sim venv, using shared cache when available. setup_libero.sh always rm -rf's the venv before reinstalling, so this function

Resolve model weights (uses ``HF_TOKEN`` if the shared cache is empty).

Resolve model weights (uses ``HF_TOKEN`` if the shared cache is empty).

Download + convert the SO100 dataset, preferring shared cache when available.

Factory that creates an infinitely nestable defaultdict.

Recursively convert all floating point tensors to the given dtype.

Recursively convert all floating point tensors in a nested structure to the given dtype. Args: x: Input data structure (tensor, dict, lis

Resize all frames to have the same height for horizontal concatenation. Ensures both width and height are even numbers for H.264 comp

Replicates tf.image.resize_with_pad for one image using PIL. Resizes an image to a target height and width without distortion by padding with zero

Restore original attention forwards after export.