Start a model server subprocess with stderr captured to a temp file. Returns the Popen object and the path to the stderr log file. On failure

Reverse action processing (denormalization, relative->absolute conversion). Processing order: 1. Denormalize actions

Wait until the server accepts TCP connections, or raise if it dies/times out.

Merge scattered external-data files into a single .data file next to the ONNX model.

Return True when the converted SO100 dataset is present and non-empty.

Initialize or reinitialize the socket with current settings

Return True when the libero uv venv looks usable. Two subtleties: 1. Uses pyvenv.cfg rather than bin/python: uv creates bin/python as a

Create a minimal mock training Config.

Return all video files under directory, sorted by path.

Tile a single observation dict to batch size n.

Apply albumentations transforms to multiple images with replay functionality. When masks are provided, mask-based transforms run per-frame be

Benchmark component-wise timing. Returns dict with times for: data_processing, backbone, action_head Args: policy: The Gr00tPoli

Build albumentations-based image transformations equivalent to the torchvision version. Args: image_target_size: Target size for res

Compute E2E timing as sum of components (more stable than separate measurement).

Export the DiT (AlternateVLDiT) to ONNX. N1.7: image_mask and backbone_attention_mask are always present (from Qwen3Backbone output). In

Generate a schedule of (dataset_index, shard_index) pairs for shard sampling. For training: Uses weighted random sampling normalized

Returns a list of robot motor commands (one per model timestep).

Get the effective episode length accounting for action horizon.

Return current RSS in MB (Linux).

Load Gr00tN1d7Config defaults.

Prepare a single observation for inference.

Replicates tf.image.resize_with_pad for multiple images using PIL. Resizes a batch of images to a target height. Args: images: A batch of

Return the path to the LIBERO ``libero_demo`` dataset. Thin wrapper around :func:`resolve_demo_dataset` kept for backward compatibility with

Convert trajectory to the desired format. Args: action_format: The desired output format Returns: A

Initialize action chunking from a list of poses. Args: poses: Sequence of Pose objects times: Optional seque

Apply rotary position embeddings using only real-valued ops (no complex). Uses float32 internally to match transformers' apply_rotary_pos_emb_vis

Args: batch: video: [T, C, H, W] Returns: vlm_content format for collation

Read the tail of a server log file and return it as a string.

Encode features for the action head. Args: backbone_output: Output from the backbone model containing: -

Convert cartesian_position (XYZ + euler) to eef_9d (XYZ + rot6d).

Get the dimension of a key from its statistics (length of min array).

Output: For video: (video_horizon,) + obs_shape For state (if not None): (state_horizon,) + obs_shape

Get video metadata using ffprobe.

Resolve the .git/modules/<name> path from a submodule's .git file.

Initialize or reinitialize the socket with current settings

Point the repo-local dataset path at the shared cached dataset.

Symlink DATASET_ROOT → SHARED_DATASET_ROOT.

Preload all actions from the current episode for efficient replay.

Shared logic: ViT TRT + embed_tokens + scatter + get_rope_index. Returns all inputs needed by either PyTorch LLM or LLM TRT engine. These ops

Auto-detect embodiment tag from processor_config.json if not provided.

Shared helper: setup sim, start server, run rollout, assert results.

Internal method to set rotation from various representations

Convert numpy/pyarrow values into standard Python types for JSON dumps.

Benchmark data processing separately with proper warmup. Data processing is CPU-bound and needs more warmup iterations. Args: po

Build TensorRT engine from ONNX model. Args: onnx_path: Path to ONNX model engine_path: Path to save TensorRT engine prec

Build torchvision-based image transformations. Args: image_target_size: Target size for resizing (list of [height, width]) i

Start background caching of the next shard using ThreadPoolExecutor. Handles epoch transitions by regenerating the sampling schedule

Derive shapes from ONNX, with optional sequence length hints. Args: onnx_path: Path to ONNX model opt_seq_lens: Dict mapping dyna

Export the action decoder (CategorySpecificMLP) to ONNX. Input: model_output [B, sa_seq_len, hidden_size], embodiment_id [B] Output: [B, sa_s

Export the action encoder (MultiEmbodimentActionEncoder) to ONNX. Input: actions [B, action_horizon, max_action_dim], timesteps [B], embodiment_i

Export the state encoder (CategorySpecificMLP) to ONNX. N1.7 change: input_dim = max_state_dim * state_history_length. The state is reshaped

Filter the shard sampling schedule for distributed training. Distributes shards across world_size processes and num_workers per proc

Load configuration from YAML file.

Get total action dimension. Args: embodiment_tag: Embodiment identifier Returns: Total action dimen

Get the dataset statistics. This is only required for dataloaders for robtics datasets.

Return the local directory containing this dataset's videos. Checks the shared drive path then the in-repo path. If neither exists, d

Get the modality configuration from the underlying policy. Returns: Dictionary mapping modality names to their configurations

Normalize values using mean-std (z-score) normalization. Args: values: Input values to normalize - Shape: (T, D) or (B,

Normalize values using min-max normalization to [-1, 1] range. Args: values: Input values to normalize - Shape: (T, D) o

Normalize whitespace for robust command matching.

Recursively add an extra dim to arrays or scalars. GR00T Policy Server expects: obs: (batch=1, time=1, ...) Calling this functio

Replace the DiT forward method with TensorRT inference.

Resolve a bundled LeRobot demo dataset by *dataset_name*. Resolution order (first match wins): 1. *path_override_env* environment variable (