What Is Flow Matching in Robot Policies?

For humanoid and other robot foundation models, the action head is the part that turns what the model sees into motion. Flow matching is how three published systems train that part, including two that learn from egocentric human video. Knowing the term explains what a policy’s “action targets” are being used for, and why the labels attached to video, such as wrist motion and hand pose, matter to the people training them.

Firsthand does not train flow-matching policies and does not deliver robot actions. It captures 3D hand pose, wrist trajectories, contact events and discrete action segments directly at the time of recording, in the same session as the egocentric video and depth, collected fast, on demand and to a written spec, and delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. A buyer who trains a flow-matching action head builds the action targets from those labels; this site states no flow-matching-specific export.

Flow matching is a way of training the part of a robot policy that outputs actions. Rather than regressing one action directly, the model is trained to predict a vector field: a direction that moves a noised version of an action chunk back toward the real one. At run time the policy starts from random noise and follows that field for a small number of steps to produce the chunk.

This page calls that component a flow-matching action head. The papers name it differently: GR00T N1 calls it an action module, and EgoVerse calls it an action decoder. Every technical statement about flow matching below comes from three official papers listed under Sources: NVIDIA’s GR00T N1, NVIDIA’s EgoScale and the EgoVerse consortium paper. Where a paper does not say something, this page does not either.

The GR00T N1 paper (arXiv 2503.14734) describes a vision-language-action model with a dual-system design. Its introduction says a Diffusion Transformer “trained with action flow-matching” serves as the System 1 action module, and its overview calls that module “a DiT-based flow-matching policy that outputs high-frequency actions.” The vision-language module runs at 10 Hz, and the action module generates closed-loop motor actions at a higher frequency of 120 Hz.

The paper writes the recipe out. Given a ground-truth action chunk, a flow-matching timestep between 0 and 1 and sampled noise, it forms a noised action chunk as the timestep times the true chunk plus one minus the timestep times the noise. The network is trained to approximate “the denoising vector field,” which is the noise minus the true chunk, by minimizing a squared-error loss.

At inference the paper states that it generates action chunks with K-step denoising. It samples random noise, then uses forward Euler integration to build the chunk iteratively, and reports: “we found K=4 inference steps to work well across all embodiments.” The chunk length H = 16 is covered on the action chunking page.

The objective does not fix where the action targets come from, and the cited papers show two different answers. GR00T N1 has no recorded actions for human video, so it states: “For human videos, in the absence of ground-truth actions, we extract learned latent actions and use them as flow-matching targets.” For robot datasets it uses both ground-truth robot actions and learned latent actions as flow-matching targets.

EgoScale (arXiv 2602.16710) instead supervises the head with recorded-style motion. Its architecture section states: “The model then predicts a chunk of future actions using a flow-matching objective.” The Figure 1 caption describes a flow-based policy pretrained on 20,854 hours of egocentric human video “using wrist motion and retargeted dexterous hand actions.” For validation, the paper draws 16 samples from the flow-matching policy at each sampled timestep, averages the predicted action chunks and computes the mean squared error against ground-truth wrist and hand actions.

The same objective therefore sits on top of different targets: learned latent actions in one paper and wrist and hand motion in the other. See the pages on latent actions and action representation for how those targets differ.

EgoVerse (arXiv 2604.07607) has a section headed “Flow Matching Action Decoder.” It states that the action decoder “is parameterized by a multi-block transformer decoder trained with a flow matching objective.” Images are processed by a ResNet-18 backbone, and proprioceptive inputs are encoded with an MLP before a shared transformer encoder conditions the decoder.

The training objective is co-training on human and robot data. The paper states that, per training step, it computes the conditional flow matching loss on a mini-batch of human samples and a mini-batch of robot samples, and the total is the robot loss plus the human loss. It is explicit about scope: its study “mainly focused on human-and-robot co-training,” with pre-train-then-fine-tune left to future work.

Flow matching describes how the action head is trained and sampled. Action chunking describes what it outputs: a block of consecutive future actions. In GR00T N1 and EgoScale the two appear together, since the flow-matching head predicts a chunk. Neither term fixes the other.

On diffusion, the sources use the vocabulary loosely together. GR00T N1 calls its action module a diffusion transformer trained with action flow-matching, and describes sampling as iterative denoising, while its baseline, Diffusion Policy, “models action distributions through a diffusion-based generative process.” This page does not rank flow matching against other objectives. In the passages cited here, none of the three papers reports an ablation of flow matching against another action objective, so no performance claim about the objective itself is made.

This site states that Firsthand captures 3D hand pose, wrist trajectories, contact events and discrete action segments, each segment with a start time, end time, verb and noun, directly at the time of recording and in the same session as the egocentric video and depth. A wrist trajectory is the position and orientation of the wrist over time, the kind of quantity EgoScale uses as supervision. The site does not state that Firsthand delivers robot actions, flow-matching targets or an action representation of its own.

The collection is fast and on demand. The site states that brief to written spec takes the same week, and that the published clock starts at the signed spec: a median of 48 hours to the first validated batch where coverage is already deep, and three to four weeks where contributors must be recruited first in a thinner environment. Delivery is in RLDS, LeRobot format or WebDataset under a buyer-owned commercial license. If your policy uses a flow-matching head, name the labels you need and the delivery format in the spec, and keep the target construction on your side of the pipeline.

Several things a team might assume are not stated and should be raised in the spec conversation: that Firsthand trains, evaluates or uses flow-matching policies; a recommended number of denoising steps; any flow-matching-ready export; delivery of robot joint or gripper commands; and any result showing Firsthand data improves a flow-matching policy. The only flow-matching figures on this page, the 10 Hz and 120 Hz module rates, K = 4 inference steps and 16 validation samples, belong to GR00T N1 and EgoScale and are not Firsthand specifications.

What is flow matching in robot policies?

Flow matching is a training objective for a policy’s action head. The model learns a vector field that turns random noise into an action chunk, and samples the chunk by following that field for a few iterative steps. GR00T N1, EgoScale and EgoVerse each use it.

Is flow matching the same as action chunking?

No. Flow matching is how the action head is trained and sampled; action chunking is the block of consecutive future actions it outputs. GR00T N1 and EgoScale use both together.

Which robot models use a flow-matching action head?

The three official sources cited here: NVIDIA’s GR00T N1, which trains its action module with action flow-matching; NVIDIA’s EgoScale, which predicts a chunk of future actions with a flow-matching objective; and EgoVerse, which uses a flow-matching action decoder.

Can a flow-matching policy learn from egocentric human video?

Yes, in the cited work. EgoScale pretrains a flow-based policy on 20,854 hours of egocentric human video using wrist motion and retargeted hand actions. GR00T N1 uses learned latent actions as flow-matching targets for human video.

Does Firsthand train flow-matching policies or deliver robot actions?

No. Firsthand captures 3D hand pose, wrist trajectories, contact events and discrete action segments at the time of recording. A buyer builds the action targets and trains the policy in their own pipeline.

How do I get a custom dataset for a flow-matching policy fast?

Write the labels and delivery format you need into a skill spec. The site states brief to spec takes the same week, with a median of 48 hours to the first validated batch where coverage is deep and three to four weeks where contributors must be recruited first, both from the signed spec.