Compute relative transformation between two poses. For EndEffectorPose: Computes the relative transformation from other to self.

Compute relative joint displacement. Args: other: Reference joint pose Returns: JointPose represent

Compute relative transformation from other to self. The result represents the transformation needed to go from other's frame to self

Internal method to compute actions from observations. Pipeline: 1. Unbatch observations into individual samples 2. Convert ea

Transform Gr00t sim observation format and compute actions. This method transforms the flat observation format from Gr00t sim environments

Replay the next action chunk from the dataset. Args: observation: Optional batched observation dictionary (used for validation, n

Simple ping handler that returns a success message.

Kill the server.

Replace ActionHead.get_action() with TRT-accelerated inference. VLLN (LayerNorm) stays in PyTorch. State Encoder, Action Encoder, DiT, and Act

Assert that each (name, snippet) appears in the README (whitespace-normalised).

Validate that the action has the correct structure and types. This method ensures that all required action keys are present and that their

Validate action in Gr00t sim environment format. This validation is specific to the flat action format used by Gr00t sim environments.

Validate that the action has the correct structure and types. Expected action structure: - action: dict[str, np.ndarray[np.float3

Validate that the observation has the correct structure and types. This method ensures that all required modalities are present and that thei

Validate observation from Gr00t sim environment format. This validation is specific to the flat observation format used by Gr00t sim environm

Validate that the observation has the correct structure and types. This method ensures that all required modalities are present and that thei

Compute loss *and* log token-level accuracy every training step. We delegate the heavy-lifting (including label smoothing, custom loss

Converts an image to uint8 if it is a float image. This is important for reducing the size of the image when sending it over the network.

Create a deep copy of this joint pose. Returns: New JointPose instance with copied data

Create a deep copy of this end-effector pose. Returns: New EndEffectorPose instance with copied data

Create a single evaluation environment with wrappers. Args: env_name: Name of the gymnasium environment to use idx: Environment i

Undo action normalization and convert relative actions to absolute.

Common mistake: just let the container go out of scope without closing.

Mimics current codebase pattern: open -> decode -> close, no exception safety.

Proper pattern with context manager.

Fixed pattern: use av.open as context manager for guaranteed cleanup.

Compute the delta action chunking where each pose represents the relative transformation from the previous frame. If referen

Sorted parquet episode files from the bundled libero demo dataset.

Read an ONNX model and derive min/opt/max shape profiles. For each input: - Fixed dimensions (concrete values) are kept as-is across min/opt/

get_action_with_features with DiT replaced by TRT.

Mimics current codebase VideoRecorder/VideoWriter: open -> encode -> close, no exception safety.

Fixed pattern: use context manager for encoding.

Get rotation as Euler angles in xyz order (degrees)

Safely expand action dimension with explicit targeting. Args: old_action_dim: Original action dimension new_

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

actions: shape (B, T, action_dim) timesteps: shape (B,) -- a single scalar per batch item returns: shape (B, T, hidden_s

Args: x: [B, T, input_dim] input tensor cat_ids: [B] category/embodiment IDs Returns: [B, T, hidd

Args: x: [B, T, input_dim] input tensor cat_ids: [B] category/embodiment IDs Returns: [B, T, outp