Generative Flow Matching for One-Step Robot Control

MeanFlow is a research framework for training robot manipulation policies using Flow Matching. Flow Matching has emerged as a faster, more stable alternative to diffusion models for continuous-time generative modeling.

This project focuses on compressing these trajectory distributions using MeanFlow architectures to generate actions in a single network function evaluation (1-NFE). This drastically reduces computational latency, enabling real-time deployment of generative policies directly on physical robot manipulators.