GR00TFlowMatchingActionHead<T>

Models & Types · AiDotNet.VisionLanguage.Robotics

Flow-matching action head used by GR00T N1 to denoise continuous joint commands conditioned on the System-2 vision-language latent.

How It Works

Per GR00T N1 (NVIDIA, 2025, “GR00T N1: An Open Foundation Model for Generalist Humanoid Robots”, arXiv:2503.14734), the System-1 action policy is a DiT-style transformer trained with the flow-matching objective (Lipman et al. 2023, arXiv:2210.02747). Inference reverses the learned vector field by Euler-integrating from Gaussian noise at t=0 to the data distribution at t=1; each integration step queries the network for the velocity field v_θ(x_t, t | latent) and updates x_{t+Δt} = x_t + Δt · v_θ.

This class implements the inference-time integrator. The per-step velocity network is supplied as a callback (the GR00T-N1 model wires its DiT decoder there), keeping this class composable.

References:

Constructors

Constructor Summary
GR00TFlowMatchingActionHead(Func<Tensor<T>, double, Tensor<T>, Tensor<T>> velocityNetwork, int numIntegrationSteps, int? seed) Builds an action head.

Properties

Property Summary
int NumIntegrationSteps { get; } Number of Euler integration steps for inference.

Methods

Method Summary
Tensor<T> Generate(int actionDimension, Tensor<T> system2Latent) Generates an action tensor of length actionDimension by Euler-integrating the flow-matching vector field from Gaussian noise at t=0 to the data distribution at t=1, conditioned on system2Latent.
Tensor<T> GenerateHorizon(int actionDimension, int horizon, Tensor<T> system2Latent) Generates a horizon of horizon action vectors, each of length actionDimension, concatenated into a flat tensor.