Configuration for a modality defining how data should be sampled and loaded. This class specifies which indices to sample relative to a base inde

TensorRT engine wrapper for loading and executing inference.

Core policy class for Gr00t model inference. This policy handles the end-to-end inference pipeline: 1. Validates input observations 2. Pr

Represents a robot configuration in joint space. This class stores joint angles/positions for a robot manipulator. Unlike end-effector p

Episode-level data loader for LeRobot format datasets. This class handles the loading and preprocessing of individual episodes from LeRobot

Represents action chunking in Cartesian space as a sequence of end-effector poses. Examples: # Create an end-effector action chunkin

Unified configuration for Gr00tN1d7 model with backbone and action head. Gr00tN1d7 uses the Cosmos-Reason2-2B (Qwen3-VL architecture) VLM backbon

Catalog entry describing a robotics dataset and how to obtain it. Datasets are resolved in order: shared drive path, in-repo path, then downl

Embodiment tags supported by the GR00T N1.7 checkpoint. Pretrain tags (baked into the base model nvidia/GR00T-N1.7-3B, inference-ready): - OX

Factory class for building training datasets. Model-agnostic.

Represents action chunking in joint space as a sequence of joint configurations. Examples: # Create a joint action chunking

An inference server that spin up a ZeroMQ socket and listen for incoming requests. Can add custom endpoints by calling `register_endpoint`.

Unified processor for robot state and action data. Handles: - State normalization (min/max, mean/std, sin/cos encoding) - Action nor

Linear layer with category-specific weights and biases for multi-embodiment support.

Represents a single end-effector pose with translation and rotation components. This class handles Cartesian space representations of robot

Action head component for flow matching diffusion policy.

Configuration for one embodiment variant of the inference smoke tests.

Minimal mock that satisfies BasePolicy interface without ABC enforcement.

Represents a single step of VLA (Vision-Language-Action) data. This is the core data structure returned by datasets, containing raw observat

Iterable dataset that combines multiple sharded datasets with configurable mixing ratios. This dataset provides the core functionality for m

Single-step dataset that creates shards from individual timesteps across episodes. This dataset implementation provides step-level data acce

Configuration for building TensorRT engines from ONNX models.

Two-layer MLP with category-specific weights for multi-embodiment support.

Dataset configuration (supports single or multiple datasets).

Common Euler angle conventions

Wrapper for Gr00tPolicy to enable compatibility with existing Gr00t simulation environments. This wrapper is specifically designed for retro-fitt

Pad non-square images to square by adding black bars (letterboxing). This is the albumentations equivalent of LetterBoxTransform (torchvision).

Custom transform to pad non-square images to square by adding black bars. Works with any tensor shape where the last 3 dimensions are (C, H, W).

Minimal ShardedDataset for testing mixture logic.

Quaternion ordering conventions

Supported rotation representation types

Write to multiple streams simultaneously.

Alternate Vision-Language DiT that separates image and non-image tokens during cross-attention processing.

Replace specified mask regions with random noise. This transform replaces pixels where mask value matches target_mask_values with random RGB nois

Configuration for GR00T inference benchmarking.

This callback saves the best checkpoint based on the metric.

This callback format checkpoint to make them standalone. For now, it copies all config files to /checkpoint-{step}/experiment_cfg/: - conf.yam

Complete configuration.

Capture DiT forward pass inputs during inference via a pre-forward hook.

Configuration for exporting GR00T N1.7 model to ONNX.

Crop the center part of the input based on fractions while maintaining aspect ratio. Args: crop_fraction: Fraction of the image to crop (

Crop a random part of the input based on fractions while maintaining aspect ratio. Args: crop_fraction: Fraction of the image to crop (0.

Gr00tN1d7: VLA model with Cosmos-Reason2-2B (Qwen3-VL) backbone.

Trainer that bypasses torch dataloader and makes data collator async.

ONNX-exportable wrapper for Qwen3-VL text model with deepstack. Key adaptations from the original Qwen3VLTextModel: 1. Eager attentio

Capture LLM (Qwen3VLTextModel) inputs during inference via a pre-forward hook. Captures inputs_embeds, position_ids, attention_mask, visual_pos_m

Holds the pre-loaded policy and dataset loader for one variant.

Apply random tint to specific mask regions. Args: target_mask_values: List of mask values to apply the transform to alpha_range:

Configuration for a modality defining how data should be sampled and loaded. This class specifies which indices to sample relative to a base inde

Action encoder with multi-embodiment support and sinusoidal positional encoding.

Configuration for multi-step environment settings. Attributes: video_delta_indices: Indices of video observations to stack state_

Unified TensorRT deployment pipeline configuration.

ONNX-exportable wrapper for Qwen3-VL Vision Model. Pre-computes position embeddings and rotary embeddings for a fixed grid_thw to avoid Compl

Policy that replays recorded actions from a LeRobot-style dataset. This policy loads actions from a dataset and replays them in sequence. It vali

Configuration for a single dataset in a mixed-training setup. A list of these objects can be supplied in ``DataConfig.datasets`` to mix multi

Produces a sinusoidal encoding of shape (B, T, w) given timesteps of shape (B, T).

Produces a sinusoidal encoding of shape (B, T, w) given timesteps of shape (B, T).

Adapter between: • Raw robot observation dictionary • GR00T VLA input format • GR00T action chunk → robot joint commands

Pipeline steps. Values are the user-facing CLI tokens for --steps.

Wrapper for TensorRT DiT engine.

Minimal stand-in for Gr00tN1d6: accepts any kwargs, returns zero action_pred.

Training configuration.

Configuration for TRT verification.

Capture ViT (VisionModel) forward inputs/outputs during inference.

Configuration for video recording settings. Attributes: video_dir: Directory to save videos (if None, no videos are saved) steps_

Container for various environment wrapper configurations. Attributes: video: Configuration for video recording multistep: Configu

AsyncVectorEnv that tolerates variable-shaped info arrays across envs. Gymnasium's default _add_info pre-allocates a numpy array based on the

Abstract base class for robot action chunking. This class provides common functionality for different action chunking types including re

Configuration for evaluating a policy.

Configuration for evaluating a policy.

Abstract base class for robotic control policies. This class defines the interface that all policies must implement, including methods for ac

Abstract base class for robotic control policies. This class defines the interface that all policies must implement, including methods for ac