Action-to-Action Flow Matching in Robotics

• The paper introduces A2A, a method that reduces control latency by generating actions via learned velocity fields instead of iterative denoising.
• A2A is formulated using differential equations in both action and latent spaces, enabling streaming execution and receding-horizon control.
• Empirical results demonstrate significant latency improvements and robust performance compared to traditional diffusion-based approaches.

Searching arXiv for papers on Action-to-Action Flow Matching and closely related flow-matching policy work. Action-to-Action Flow Matching (A2A) is a family of flow-matching formulations for action generation in robotics and Vision-Language-Action systems in which future actions, action trajectories, or latent action codes are produced by integrating a learned velocity field, rather than by running long diffusion-style denoising chains. Across recent work, A2A has been instantiated as direct action-space transport around the previously executed action, latent transport from encoded action history to future-action codes, manifold-valued flow on $SE(3)$, and chunk-level flow-matching action experts inside large VLAs. The unifying objective is reduced control latency while preserving multimodal behavior and closed-loop responsiveness (Jiang et al., 28 May 2025, Jia et al., 7 Feb 2026, Funk et al., 2024, Shi et al., 30 Mar 2026).

1. Emergence and scope

A2A emerged from dissatisfaction with diffusion-style robot policies that start from pure Gaussian noise and require many iterative denoising steps before a usable action sequence is available. In the streaming-flow formulation, diffusion and flow-matching policies are described as sampling a “trajectory of trajectories”: a diffusion or flow trajectory over action trajectories, with intermediate trajectories discarded and no action executable until the entire sampling process completes (Jiang et al., 28 May 2025). The explicit A2A formulation in latent space makes the same critique more directly: it questions the necessity of uninformed noise sampling and instead initializes generation from encoded historical proprioceptive sequences, arguing that the source and target distributions are then much closer and can often be connected in one or a few flow steps (Jia et al., 7 Feb 2026).

A parallel line of work uses closely related language inside VLAs. StreamingVLA replaces chunk-wise denoising with “action flow matching,” enabling one-by-one execution and asynchronous overlap between observation, action generation, and action execution (Shi et al., 30 Mar 2026). SnapFlow, by contrast, keeps the standard flow-matching VLA backbone but compresses multi-step denoising into a single forward pass through self-distillation (Luan et al., 7 Apr 2026). This suggests that “A2A” has become a family resemblance term for fast continuous-time action generation rather than a single canonical algorithm.

2. Core mathematical formulations

One canonical A2A formulation treats the future action trajectory itself as the flow state. Let the observation history be $h=\{o_1,\dots,o_K\}\in\mathcal O^K$, and let the target future trajectory be $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$. A history-conditioned velocity field $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$ defines the ODE

$$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$

For each demonstration $\xi$, the analytic conditional field is

$v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$

with initialization $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$. The resulting marginal is

$$p_\xi(a\mid t)=\mathcal N\!\Bigl(a\,\Big|\,\xi(t),\,\sigma_0^2e^{-2kt}I\Bigr),$$

and training minimizes the conditional flow-matching loss

$$\mathcal L(\theta)=\mathbb E_{(h,\xi)\sim p_D}\mathbb E_{t\sim{\rm Unif}[0,1]}\mathbb E_{a\sim p_\xi(a|t)} \bigl\|v_\theta(a,t\mid h)-v_\xi(a,t)\bigr\|_2^2.$$

At the optimum, the induced marginal equals

$h=\{o_1,\dots,o_K\}\in\mathcal O^K$0

which yields per-timestep matching of multimodal targets (Jiang et al., 28 May 2025).

A second formulation operates in a shared latent action space. Historical actions $h=\{o_1,\dots,o_K\}\in\mathcal O^K$1 are encoded as $h=\{o_1,\dots,o_K\}\in\mathcal O^K$2, future actions $h=\{o_1,\dots,o_K\}\in\mathcal O^K$3 as $h=\{o_1,\dots,o_K\}\in\mathcal O^K$4, and the linear interpolant

$h=\{o_1,\dots,o_K\}\in\mathcal O^K$5

is paired with the target vector field $h=\{o_1,\dots,o_K\}\in\mathcal O^K$6. The training loss is

$h=\{o_1,\dots,o_K\}\in\mathcal O^K$7

usually combined with an action autoencoder term and an inference-consistency term. In this formulation, A2A denotes transport from a past-action latent to a future-action latent rather than from Gaussian noise to action (Jia et al., 7 Feb 2026).

In many VLA implementations, the same flow-matching machinery is applied to entire action chunks. A standard interpolation is

$h=\{o_1,\dots,o_K\}\in\mathcal O^K$8

with target velocity $h=\{o_1,\dots,o_K\}\in\mathcal O^K$9, or equivalently $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$0 with target $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$1. These formulations keep Gaussian endpoints but relocate action generation to a learned flow field over continuous action chunks (Yang et al., 25 Jun 2026).

3. Streaming execution and receding-horizon control

The most distinctive operational feature of direct-action A2A is streaming execution. Instead of initializing from pure noise in trajectory space, the policy starts in action space from

$\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$2

where $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$3 is the last executed action. Numerical integration then produces actions incrementally: $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$4 typically via Euler or RK4. Because each new action depends only on past actions, actions can be sent to the robot during the sampling process itself, making the method well suited to receding-horizon control (Jiang et al., 28 May 2025).

The stabilizing term $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$5 is central in this variant. It exponentially contracts the flow toward the demonstration trajectory, yielding marginals $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$6 and reducing drift and distribution shift. Empirically, an ablation with $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$7 loses several percent of task success rate, which the paper uses to support the utility of stabilization (Jiang et al., 28 May 2025).

StreamingVLA generalizes the same execution logic to large VLAs. Its action flow matching overlaps generation and execution, while an action saliency-aware adaptive observation mechanism overlaps execution and observation. The reported result is a $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$8 latency speedup and a $\xi:[0,1]\to\mathcal A\subseteq\mathbb R^d$9 reduction in execution halting, without sacrificing performance; in LIBERO, StreamingVLA (AFM) reports $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$0 success with $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$1 ms per action and StreamingVLA (AFM+AEO) reports $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$2 success with $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$3 ms per action (Shi et al., 30 Mar 2026).

4. Representational and architectural variants

A2A has diversified rapidly across representation spaces. Some systems operate directly in Euclidean action space; others move to latent action manifolds or rigid-body pose groups in order to encode geometry, temporal coherence, or cross-embodiment structure.

Paper	Representation	Distinguishing mechanism
ActionFlow (Funk et al., 2024)	$$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$4	SE(3)-Invariant Transformer and local-frame ODE
LG-Flow Policy (Songwei et al., 30 Jan 2026)	chunk latents $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$5	GRU/VAE latent flow with geometry-aware point-cloud FiLM
RotVLA (Li et al., 13 May 2026)	latent actions in $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$6	triplet-frame composition and unified latent+robot action flow
NORA-1.5 (Hung et al., 18 Nov 2025)	$$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$7 action chunks	parallel transformer action expert with $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$8

ActionFlow formulates A2A directly on $$v_\theta:\mathcal A\times[0,1]\times\mathcal H\to\mathbb R^d$$9. A single pose $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$0 evolves through translation and rotation velocities in local coordinates, while the reference flow uses straight geodesic interpolation and a pure $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$1 flow-matching loss. Coupled with Invariant Point Attention, this yields globally $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$2-equivariant action generation (Funk et al., 2024).

LG-Flow Policy moves the flow to a temporally regularized latent action trajectory. Demonstration trajectories are chunked, encoded with temporal convolution and GRU, projected into latent codes $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$3, and then transported from Gaussian noise to target latent sequences. The decoder, modulated by wrist-camera features through FiLM, reconstructs low-level control while preserving a smooth latent plan (Songwei et al., 30 Jan 2026).

RotVLA uses a continuous rotational latent action space $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$4, with latent actions extracted by projecting unconstrained matrices onto the nearest rotation through SVD. Its triplet-frame learning objective combines single-step reconstruction, composition consistency, and a SoftVQ codebook regularizer. In downstream control, latent actions serve explicitly as a planner conditioning a joint latent-and-robot-action flow head (Li et al., 13 May 2026).

NORA-1.5 attaches a transformer-based flow-matching action expert to an autoregressive VLA backbone. The expert receives a noisy action sequence as query input, attends to key/value pairs from the VL backbone, predicts the flow field over a horizon $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$5, and is trained jointly with the FAST+ autoregressive head under

$$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$6

This hybridization treats flow matching as both an action generator and a source of gradient signal back into the vision-language stack (Hung et al., 18 Nov 2025).

5. Empirical performance, acceleration, and reliability

Direct streaming A2A reports favorable comparisons to prior diffusion and flow-matching baselines on Push-T and RoboMimic. The reported inference latency per action is typically $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$7–$$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$8 ms for A2A, versus $$\frac{d}{dt}\,a(t)=v_\theta\bigl(a(t),t\mid h\bigr).$$9–$\xi$0 ms for $\xi$1-step diffusion and $\xi$2–$\xi$3 ms for streaming diffusion, while imitation accuracy often matches or exceeds full diffusion policies with $\xi$4 DDPM steps (Jiang et al., 28 May 2025).

The latent-history A2A paper reports one-step generation in $\xi$5 ms per action chunk, with simulated-task results of $\xi$6 on Close Box, $\xi$7 on Pick Cube, $\xi$8 on Stack Cube, $\xi$9 on Open Drawer, and $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$0 on Pick-Place Bowl under the setting “100 demonstrations, 30 epochs, 6 inference steps.” It also reports $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$1 success on real Pick Cube and Open Drawer, and substantially stronger robustness to visual perturbations than the listed baselines (Jia et al., 7 Feb 2026).

Acceleration inside large VLAs has produced additional gains. SnapFlow compresses $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$2-step denoising into $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$3-NFE self-distilled generation; on $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$4 across four LIBERO suites it reports $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$5 average success, compared with $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$6 for the $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$7-step teacher, with denoising reduced from $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$8 ms to $v_\xi(a,t)=\dot\xi(t)-k\bigl(a-\xi(t)\bigr),$9 ms and end-to-end latency from $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$0 ms to $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$1 ms (Luan et al., 7 Apr 2026). This is not a direct-action streaming formulation, but it occupies the same efficiency frontier that motivated A2A.

Reliability-oriented extensions operate during or around flow generation rather than replacing it. Neuro-symbolic safety guidance for flow-matching VLAs interleaves denoising with a minimum-norm constrained optimization enforcing discrete-time exponential CBF constraints on predicted end-effector trajectories. On SafeLIBERO it reports $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$2 collision avoidance and $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$3 task success, with improvements of $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$4 and $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$5 over single-step methods (English et al., 1 Jul 2026). PAMAE replaces the single action expert with a phase-aware sparse MoE; on multi-stage manipulation simulation tasks it improves task success by up to $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$6 over strong VLA baselines, and PAMAE$a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$7 reports an average success of $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$8 across Table-Cleaning, Drawer-Cycle, Lid-Open, Shelf-Insert, and Cup-Upright (Yang et al., 25 Jun 2026).

6. Limitations, misconceptions, and open directions

A common misconception is that A2A guarantees full trajectory fidelity. In the streaming-flow formulation, the theoretical guarantee is only on per-timestep marginals: $a_0\sim\mathcal N(\xi(0),\sigma_0^2I)$9 The paper states explicitly that this does not guarantee the full joint trajectory distribution; trajectories may splice segments from different modes. It further notes that this compositionality can be advantageous in robotics, but global constraints not expressible per timestep can still be violated (Jiang et al., 28 May 2025).

Another misconception is that A2A always means one-step generation from the previous action. The literature is broader. Some formulations do exactly that or approximate it in latent space (Jia et al., 7 Feb 2026), whereas manifold-based and streaming formulations still integrate an ODE over several steps (Jiang et al., 28 May 2025, Funk et al., 2024), and many VLA action heads labeled as A2A or action flow matching still begin from $$p_\xi(a\mid t)=\mathcal N\!\Bigl(a\,\Big|\,\xi(t),\,\sigma_0^2e^{-2kt}I\Bigr),$$0 over action chunks (Hung et al., 18 Nov 2025, Yang et al., 25 Jun 2026).

Formal guarantees remain uneven. The neuro-symbolic safety-guided sampler gives a discrete safety guarantee only so long as the QP is feasible at every denoising step, and the paper explicitly does not provide a convergence proof, explicit safety-margin bounds, Lagrange-multiplier analysis, Lipschitz assumptions on $$p_\xi(a\mid t)=\mathcal N\!\Bigl(a\,\Big|\,\xi(t),\,\sigma_0^2e^{-2kt}I\Bigr),$$1, or spectral-norm bounds (English et al., 1 Jul 2026). PAMAE reports that its phase labels are rule-based and coarse, all experiments are in simulation, and the current design fixes both the number of experts and the number of phases to three (Yang et al., 25 Jun 2026).

Data efficiency is also unresolved. NORA-1.5 reports that with only $$p_\xi(a\mid t)=\mathcal N\!\Bigl(a\,\Big|\,\xi(t),\,\sigma_0^2e^{-2kt}I\Bigr),$$2k frames on Galaxea, pure flow matching underperforms autoregressive decoding, suggesting a need to pretrain the flow expert (Hung et al., 18 Nov 2025). The latent-history A2A formulation notes sensitivity to noisy proprioceptive feedback, although injecting small Gaussian noise into the historical action sequence improves robustness (Jia et al., 7 Feb 2026).

Open directions named in the literature include hierarchical A2A experts spanning multiple temporal scales, uncertainty-aware or energy-based flow matching, real-robot validation of phase-aware or safety-guided variants, self-supervised phase discovery, dynamic expert allocation, reinforcement-learning integration, and extension beyond control to temporal video generation (Hung et al., 18 Nov 2025, Yang et al., 25 Jun 2026, Jia et al., 7 Feb 2026). Collectively, these directions indicate that A2A is evolving from a latency-reduction technique into a broader design pattern for temporally structured generative control.