Flow Matching Loss

Training a network with mean squared error to predict the velocity that points from noise toward the real data.

Flow matching was introduced by Lipman and colleagues in 2022, treating “noise gradually turning into data” as continuous flow along a path, with the network learning the velocity field at every point along that path. Training doesn't need to simulate the whole trajectory: pick a random time τ, linearly interpolate between a real sample and Gaussian noise to get an intermediate point, and use mean squared error to have the network's predicted velocity match the direction of that straight-line path (the real sample minus the noise) — that's the flow matching loss. It looks a lot like a diffusion model's denoising loss in form, but the path is straighter, so fewer sampling steps are needed. At inference, starting from pure noise, a few steps of Euler integration along the predicted velocity produce a sample. The π0 family of VLAs uses it to train their action expert, generating continuous action chunks directly.

Exampleπ0 samples τ during training from a Beta distribution skewed toward the high-noise end, forms the noisy action τ·A + (1 − τ)·ε, and has the action expert regress toward A − ε; at inference it starts from pure noise and generates an action chunk with 10 steps of Euler integration.

Also called

Conditional Flow Matching Loss, CFM Loss, Velocity Field Regression Loss

Related

Flow Matching · Velocity Field · Rectified Flow · Mean Squared Error · Denoising Loss (Diffusion Loss) · π0

Sources

Flow Matching for Generative Modeling (arXiv:2210.02747)
π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv:2410.24164)