Convert 6D rotation representation to rotation matrix. Args: rot6d: 6D rotation as (6,) array - first two rows of rotati

Set up TRT engines for action head only (N1.7 mode). Backbone (Qwen3-VL) stays in PyTorch. Only the 4 action head components (State Encoder,

Set up TRT engine for DiT-only acceleration (backward compatible). Only replaces the DiT model in the action head. The backbone and other act

Set up TRT engines for N1.7: ViT TRT + LLM TRT + Action Head TRT. The Qwen3-VL backbone's vision encoder and text model are both replaced wit

Set up TRT engines for ViT + LLM only; action head stays in PyTorch. Use this on platforms where DiT cannot be exported with dynamic vl_seq_len

ONNX-compatible causal mask without complex type casts.

Recursively turn a (nested) defaultdict into a regular dict.

Convert a single observation into a VLAStepData object for processing. Args: observation: Single observation dict with video, sta

Patch Gr00tPolicy to drop DiT blocks down to _TRUNCATED_DIT_BLOCKS after init.

Unbatch a batched observation into a list of single observations. Args: value: Batched observation with shape (B, ...) for each m

Validate the API token in the request.

Validate source and destination paths to prevent security issues.

Apply both state and action processing together. Convenience method that processes state and action in one call, automatical

Apply sin/cos encoding to values. Args: values: Array of shape (..., D) containing values to encode Returns: Array of shape

Build all TRT engines for the full pipeline. Shape profiles are automatically derived from the ONNX models. Dynamic sequence dimensions use h

Build HF cache environment variables for a cache key.

Check if the action is valid. Args: action: Dictionary containing the action to be executed Raises: Assertio

Check if the action is valid. Args: action: Dictionary containing the action to be executed Raises: Assertio

Dimension A — processor_config.json checks for a single model.

Dimension B — config.json checks for a single model.

Dimension C — embodiment_id.json checks.

Dimension D — statistics.json checks.

Dimension E — README & Documentation Consistency.

Dimension F1 — config.json ↔ processor_config.json agreement.

Dimension F2 — MODALITY_CONFIGS ↔ examples/*/modality.json.

Dimension F — Cross-file consistency (source-of-truth only).

Dimension J — Verify enum serialization uses names not values.

Check if the observation is valid. Args: observation: Dictionary containing the current state/observation of the environment

Check if the observation is valid. Args: observation: Dictionary containing the current state/observation of the environment

Check the test fixture processor_config against source of truth.

Args: ds: list of dicts with the same keys and the same value type recursive: whether to recursively compress nested dictionaries

Convert cartesian_position (XYZ + euler 3D) to eef_9d (XYZ + rot6d). Args: cartesian_position: (..., 6) array [x, y, z, euler_x, euler_y,

Convert cartesian_position (XYZ + euler 3D) to eef_9d (XYZ + rot6d). Uses extrinsic XYZ Euler convention (scipy ``"XYZ"``, equivalent to ``tf

For video, the observation space will be (video_horizon,) + original shape For state (if not None), the observation space will be (st

Recursively convert tensors to lists in nested dictionaries/lists.

Recursively convert tensors to lists in nested dictionaries/lists.

Create a synthetic test video for benchmarking.

Decode the action from the model output.

chunk["single_arm"]: (B, T, 5) chunk["gripper"]: (B, T, 1) Convert to: { "shoulder_pan.pos":

Simulates codebase pattern: exception during decode → container never closed.

Fixed pattern: context manager ensures close even on exception.

Delete the current cached shard to free memory.

Download the dataset from HuggingFace Hub into dest. Uses hf_hub_download for each file in hf_files to avoid enumerating the entire r

Download minimal files from the DROID v3.0 dataset.

Download a small sample from a SimplerEnv dataset.

Convert euler angles (3D) to rotation 6D representation. Uses extrinsic XYZ Euler convention (scipy ``"XYZ"``, equivalent to ``tfg.rotation_m

Main entry point for real-robot policy evaluation.

Expand action dimension by copying weights from existing dimensions. Args: old_action_dim: Original action dimension

Export the Qwen3-VL text model (LLM) to ONNX. The LLM receives inputs_embeds (with vision tokens already scattered in), pre-computed 3D posit

Export Qwen3-VL Vision Model to ONNX. Pre-computes position/rotary embeddings for the captured grid_thw to avoid ComplexDouble ops. Monkey-pa

Export the vl_self_attention (SelfAttentionTransformer) to ONNX. This module sits between VLLN and the DiT, transforming backbone embeddings.

Convert downloaded v3.0 data to GR00T LeRobot v2.0 format.

Return the first NVIDIA EGL vendor JSON file found, or raise FileNotFoundError.

Wait for the background caching job to complete and retrieve the shard.

Execute TRT inference with the given input tensors. Accepts both positional and keyword arguments. Returns a dict of output tensors b

Forward pass through the action head. Args: backbone_output: Output from the backbone model containing:

For each dt window, choose the first timestamp in the window. Assumes timestamps sorted. One timestamp might be chosen multiple times due to

Compute and return the next action based on current observation with validation. This is the main public interface. It validates the observat

Generate actions using the flow matching diffusion process. Args: backbone_features: [B, seq_len, backbone_embedding_dim

Extract and process a single timestep from episode data. Converts raw episode data into a VLAStepData structure and applies

Get the merged dataset statistics.

Generate DeepSpeed configuration.