Action-to-Action Flow Matching

5 min read

Use the robot's recent motion, rather than random noise, as the starting point for generating its next action chunk.

SourceExecuted action historyFlow space512-D latentFlow net4 × AdaLN-MLP1-step latency0.56 ms

the key shift

Start close to the answer

A robot already has a structured prior: the motion it just executed. A2A promotes that history from a small conditioning signal to the source distribution itself.

Regression

ObservationsFuture actions

One deterministic mapping

Diffusion / flow

Gaussian noiseFuture actions

A long, iterative transport path

A2A

proposed
Past actionsFuture actions

A short, informed transport path

architecture

History latent → conditioned flow → future actions

Action history defines the source point z₀. Images tell the vector field where that point should move.

1

Encode

Executed actions a≤t

8-step history → 1D CNN ×3 → source latent z₀

Visual history I≤t

8 frames → ResNet-18 → linear projection → condition c

2

Transport

Flow net

AdaLN-MLP ×4

z₀512-D latent flowz₁

Learn vτ(zτ, τ, c), then integrate the ODE from the history latent to the future latent in as little as one Euler step.

3

Decode

Residual MLP ×4

Decode z₁ into the next 8-step action chunk a>t.

training

Three losses make the short path executable

The vector field is only one part of the job: the latent must preserve actions, and ODE sampling must land on a physically valid chunk.

Flow matching

LFM

Predict the target vector field between z₀ and z₁.

Reconstruction

LAE

Keep action chunks recoverable from the shared latent space.

Inference consistency

LIC

Align integrated latents and decoded actions with ground truth.

Ltotal = λ₁ LFM + λ₂ LAE + λ₃ LIC

evidence

Fast enough for one-step control

A2A was evaluated on five simulated manipulation tasks and two real-world Franka tasks.

0.56 ms

one-step latency

Measured on an RTX 5090

>90%

one-step success

Close Box after 32 epochs

30–40%

unseen visual scenes

Six-step A2A across levels 1–3

Success rate after 30 epochs / 100 demonstrations

Five simulation tasks · percentages

MethodStepsClose boxPick cubeStack cubeOpen drawerPick-place bowl
A2A69292869290
VITA68888809092
FM-UNet108270283468
ACT18286328060

takeaways

The useful intuition

Why it works

Adjacent action chunks are physically similar. Encoding both into one latent space aligns their distributions and straightens the transport path.

Why it generalizes

Action history and images travel through separate paths, so low-dimensional proprioception is not drowned out by visual features.

Where it bends

The continuity prior is weaker for binary gripper switches and is sensitive to uncertain histories. Small injected noise helps, but the balance is task-specific.

In one sentence: A2A turns the robot's recent executed motion into the initial state of a vision-guided flow, making the route to its next action chunk short enough for single-step inference.

The mental model

A2A starts from an observation about robot motion: the action chunk that just happened is usually close to the action chunk that should happen next. Instead of transporting Gaussian noise all the way to an executable action, it encodes the robot's executed action history into a latent starting point and learns a shorter, vision-conditioned flow to the future-action latent.

That shorter path is simple enough for a four-block AdaLN MLP and can be integrated in one step. The trade-off is equally direct: the method is strongest when actions change smoothly, while discrete switches and corrupted action histories weaken its continuity prior.

Sources

Built by Suveen.

www.suveenellawela.com v.2026.5