Foresight Without Seeing: Latent Futures for World Action Models

# Foresight Without Seeing:Latent Futures for World Action Models
Source: [https://arxiv.org/html/2608.11605](https://arxiv.org/html/2608.11605)
Zhongbo WuAffiliation:Shanghai Jiao Tong UniversityAffiliation:ACE RoboticsZheng ZhangAffiliation:ACE RoboticsAffiliation:Nanyang Technological UniversityZihan WangThanks:Research intern\.Affiliation:Shanghai Jiao Tong UniversityShan YouAffiliation:ACE RoboticsTao HuangThanks:Corresponding author\.Affiliation:Shanghai Jiao Tong University

###### Abstract

World Action Models \(WAMs\) couple future visual prediction with robot action generation, enabling policies to model how the physical world evolves during interaction\. Existing WAMs differ primarily in how such predictive dynamics are exposed to the action pathway\. Explicit\-future WAMs provide direct access to predicted scene evolution through future generation, but incur substantial inference costs from iterative video denoising\. In contrast, direct\-policy WAMs skip future generation and efficiently predict actions from the current observation, but lack an explicit inference\-time interface for exposing predictive dynamics to the Action DiT\. To bridge this gap, we proposeForeWAM, a dynamics\-conditioned direct\-policy WAM that provides predictive context for action generation without decoding future videos\. At its core,Future\-KVperforms a single Video DiT prefill over the clean current visual latent and stochastic future slots, and reuses the resulting layer\-wise key\-value states throughout action denoising\. This allows the Action DiT to access predictive context formed by the video backbone without iterative future generation\. We further introduce dynamics registers supervised by a frozen latent action teacher, encouraging the implicit future states to capture interaction\-induced transitions, including object motion, contact changes, and task progress\. Ground\-truth future observations and the teacher are used only during training; deployment requires neither future observations nor the teacher and performs no future video generation\. Without embodied robot data pretraining, the standard and accelerated variants of ForeWAM achieve average success rates of 96\.7% and 96\.9% on LIBERO, respectively\. The standard variant further achieves 61\.6% success on LIBERO\-Plus\. These results demonstrate that direct\-policy WAMs can retain efficient action prediction while exposing predictive dynamics to the action pathway, without explicitly generating future observations\.

## 1Introduction

Vision\-Language\-Action \(VLA\) models offer a promising approach to Physical AI by predicting robot actions from visual observations and language instructions\. However, they primarily learn reactive observation\-to\-action mappings without explicitly modeling how the physical world evolves through interaction\. World Action Models \(WAMs\) have emerged as a new paradigm that couples future visual prediction with action generation, enabling policies to capture interaction\-induced scene dynamics\([14](https://arxiv.org/html/2608.11605#bib.bib1);[18](https://arxiv.org/html/2608.11605#bib.bib2);[33](https://arxiv.org/html/2608.11605#bib.bib14);[46](https://arxiv.org/html/2608.11605#bib.bib19);[52](https://arxiv.org/html/2608.11605#bib.bib21)\)\.

![Refer to caption](https://arxiv.org/html/2608.11605v1/paradigm.png)Figure 1:World Action Model paradigms\.\(a\) Cascaded WAMs first generate future observations and then condition action prediction on them\. \(b\) Joint WAMs generate future observations and actions within a unified generative process\. \(c\) Direct\-policy WAMs skip future rollout at inference and condition action prediction on a latent world representation extracted from the current observation\. \(d\) Our ForeWAM retains direct action prediction while additionally exposing action\-relevant predictive dynamics through hidden future\-slot K/V states and dynamics registers\. Hatched tokens denote noisy variables; future slots are stochastic internal states rather than observed future frames\.WAM designs differ primarily in how predictive visual context reaches the action pathway, as summarized in Figure[1](https://arxiv.org/html/2608.11605#S1.F1)\. Figure[1](https://arxiv.org/html/2608.11605#S1.F1)\(a\) first generate future observations and then condition action prediction on them, whereas Figure[1](https://arxiv.org/html/2608.11605#S1.F1)\(b\) denoise future video and actions together\([14](https://arxiv.org/html/2608.11605#bib.bib1);[46](https://arxiv.org/html/2608.11605#bib.bib19);[2](https://arxiv.org/html/2608.11605#bib.bib5)\)\. Both expose predicted scene changes to the action pathway, but iterative video denoising adds inference cost and generation errors may propagate into action prediction\. Figure[1](https://arxiv.org/html/2608.11605#S1.F1)\(c\), represented by Fast\-WAM, avoid future\-video generation at inference while retaining future\-video modeling during training\([48](https://arxiv.org/html/2608.11605#bib.bib20)\)\. This improves efficiency, but leaves open how the Action DiT can access predictive, action\-relevant context without a future rollout\. Together, these designs expose a trade\-off between predictive context and inference efficiency, raising a central question:

*How can a direct\-policy WAM enable its Action DiT to access predictive dynamics without explicitly generating future observations?*

We address this question withForeWAM, aForesight\-without\-Seeing World Action Modelthat learns to act from latent futures without video rollouts\. As shown in Figure[1](https://arxiv.org/html/2608.11605#S1.F1)\(d\), ForeWAM preserves the direct\-policy inference structure while replacing explicit future\-observation generation with a latent future interface exposed to the Action DiT\. At its core isFuture\-KV, an implicit interface that transfers predictive context from the Video DiT to the Action DiT\. Future\-KV preserves the clean visual latent of the current observation, initializes unobserved future slots with noise, and processes them through a single Video DiT prefill\. The resulting layer\-wise key–value states are cached and reused throughout action denoising, allowing the action pathway to access predictive context over both the current observation and latent future slots without iteratively generating or decoding future video\.

To further encourage these implicit future states to focus on scene transitions induced by robot interaction, we introducedynamics registerssupervised by a frozen LaWM latent\-action teacher\([11](https://arxiv.org/html/2608.11605#bib.bib10)\)\. During training, the teacher extracts compact, non\-executable latent\-action representations from pairs of real visual observations before and after a transition\. These representations supervise the dynamics registers to encode state\-transition information, including object motion, contact changes, and task progress\. Future\-KV thus establishes a predictive information pathway from the Video DiT to the Action DiT, while latent\-action supervision further strengthens the interaction\-relevant dynamics represented along this pathway\. Ground\-truth future observations and the latent\-action teacher are used only during training\. At deployment, ForeWAM requires neither future observations nor the teacher and performs no future\-video generation\.

As a result, ForeWAM achieves competitive performance while substantially improving both training and inference efficiency, using only a compact Wan2\.1\-T2V\-1\.3B Video DiT and eliminating the need for embodied robot\-data pretraining\. To further accelerate inference, we apply OneDP\([38](https://arxiv.org/html/2608.11605#bib.bib4)\)to distill the action\-denoising process into a reduced\-step schedule, yielding an accelerated variant termedForeWAM\-Flash\. On our observed LIBERO\-Plus subset, ForeWAM and ForeWAM\-Flash achieve success rates of 61\.6% and 58\.2%, respectively, surpassing the reported Fast\-WAM result of 51\.5% by 10\.1 and 6\.7 percentage points\. ForeWAM reduces the mean action\-generation latency from 667 ms to 568 ms, a 14\.8% reduction relative to Fast\-WAM, while ForeWAM\-Flash further lowers it to 220 ms, corresponding to a 67\.0% reduction\. Moreover, ForeWAM uses approximately one\-third of the policy parameters of Fast\-WAM \(2B versus 6B\)\.

Our main contributions are summarized as follows:

- •We identify a key interface problem in direct\-policy WAMs: removing future\-video generation improves efficiency but eliminates the explicit pathway through which predictive dynamics reach the Action DiT\.
- •We proposeForeWAM, combining Future\-KV with latent\-action\-supervised dynamics registers\. A single Video DiT prefill produces layer\-wise K/V states for action denoising, while a frozen LaWM teacher encourages the registers to capture interaction\-induced scene transitions\.
- •Without embodied robot\-data pretraining, ForeWAM achieves up to 10\.1 percentage points higher LIBERO\-Plus success and 67\.0% lower action\-generation latency than the reported Fast\-WAM configuration, while using approximately one\-third of its policy parameters\. Matched component comparisons further validate the proposed design\.

## 2Related Work

#### Vision\-language\-action policies\.

VLA models map visual observations and language instructions to executable robot actions\([6](https://arxiv.org/html/2608.11605#bib.bib50);[5](https://arxiv.org/html/2608.11605#bib.bib22);[24](https://arxiv.org/html/2608.11605#bib.bib13);[36](https://arxiv.org/html/2608.11605#bib.bib48);[29](https://arxiv.org/html/2608.11605#bib.bib49);[19](https://arxiv.org/html/2608.11605#bib.bib51);[44](https://arxiv.org/html/2608.11605#bib.bib43)\), commonly by attaching an action decoder to a pretrained vision\-language backbone\([21](https://arxiv.org/html/2608.11605#bib.bib23);[22](https://arxiv.org/html/2608.11605#bib.bib18);[49](https://arxiv.org/html/2608.11605#bib.bib46)\)\. Diffusion and flow objectives support multimodal continuous action generation\([13](https://arxiv.org/html/2608.11605#bib.bib11);[27](https://arxiv.org/html/2608.11605#bib.bib15);[4](https://arxiv.org/html/2608.11605#bib.bib7)\), while large\-scale robot pretraining can improve transfer across tasks and embodiments\([3](https://arxiv.org/html/2608.11605#bib.bib6);[7](https://arxiv.org/html/2608.11605#bib.bib8);[50](https://arxiv.org/html/2608.11605#bib.bib42)\)\. These methods establish strong direct policies, but do not by themselves provide an explicit action\-facing interface through which predictive visual dynamics can be accessed during control\.

#### World\-action models\.

World Action Models \(WAMs\) augment direct action prediction with predictive world dynamics\. Existing future\-modeling WAMs broadly follow cascaded and joint paradigms\. Cascaded approaches follow an*imagine\-then\-act*structure, predicting future observations or intermediate representations before extracting actions\. Some methods explicitly generate future visual observations as intermediate plans\([14](https://arxiv.org/html/2608.11605#bib.bib1);[15](https://arxiv.org/html/2608.11605#bib.bib24);[18](https://arxiv.org/html/2608.11605#bib.bib2);[20](https://arxiv.org/html/2608.11605#bib.bib32)\), whereas others use structured or compressed predictive representations, such as correspondences, point tracks, motion fields, masks, or distilled foresight\([1](https://arxiv.org/html/2608.11605#bib.bib26);[25](https://arxiv.org/html/2608.11605#bib.bib27);[41](https://arxiv.org/html/2608.11605#bib.bib29);[51](https://arxiv.org/html/2608.11605#bib.bib30);[30](https://arxiv.org/html/2608.11605#bib.bib31);[42](https://arxiv.org/html/2608.11605#bib.bib33)\)\. Joint WAMs instead co\-model future states and actions within a shared architecture, allowing world and action representations to interact during generation\. Autoregressive variants organize visual states and actions within a unified generative sequence\([9](https://arxiv.org/html/2608.11605#bib.bib9);[8](https://arxiv.org/html/2608.11605#bib.bib35);[10](https://arxiv.org/html/2608.11605#bib.bib34);[40](https://arxiv.org/html/2608.11605#bib.bib47)\), whereas diffusion\- and flow\-based variants jointly model world dynamics and action trajectories, with some recent approaches using latent or implicit representations for greater efficiency\([2](https://arxiv.org/html/2608.11605#bib.bib5);[46](https://arxiv.org/html/2608.11605#bib.bib19);[52](https://arxiv.org/html/2608.11605#bib.bib21);[17](https://arxiv.org/html/2608.11605#bib.bib25);[34](https://arxiv.org/html/2608.11605#bib.bib28);[23](https://arxiv.org/html/2608.11605#bib.bib36);[39](https://arxiv.org/html/2608.11605#bib.bib37);[43](https://arxiv.org/html/2608.11605#bib.bib38);[12](https://arxiv.org/html/2608.11605#bib.bib39);[26](https://arxiv.org/html/2608.11605#bib.bib40);[47](https://arxiv.org/html/2608.11605#bib.bib41);[35](https://arxiv.org/html/2608.11605#bib.bib52);[31](https://arxiv.org/html/2608.11605#bib.bib44)\)\. Although these approaches expose future scene evolution to action prediction, iterative future generation or tightly coupled world–action computation introduces substantial inference overhead\. Direct\-policy WAMs such as Fast\-WAM avoid future generation by predicting actions from the current observation representation\([48](https://arxiv.org/html/2608.11605#bib.bib20);[45](https://arxiv.org/html/2608.11605#bib.bib45)\)\. However, future dynamics are not explicitly exposed to the Action DiT under this direct\-policy interface\.

In contrast, our method retains direct\-policy inference while exposing predictive dynamics to the Action DiT through a hidden future\-slot K/V interface and dynamics registers supervised by a LaWM latent\-action target\([11](https://arxiv.org/html/2608.11605#bib.bib10)\)\. The intended contribution is therefore the complementary composition of these two conditioning paths, rather than no\-rollout inference or future\-aware representation learning in isolation\.

## 3Method

Our goal is to expose predictive visual context to a direct action policy without decoding a future video at deployment\. The proposed model combines a video diffusion transformer, a dedicated Action DiT, a hidden future\-slot K/V cache, and latent\-action\-supervised dynamics registers \(Figure[2](https://arxiv.org/html/2608.11605#S3.F2)\)\. The cache preserves distributed visual context, whereas the registers provide a compact transition\-oriented pathway\. We first formalize the deployment interface, then describe token routing and the two conditioning paths, and finally specify the joint training objective\.

![Refer to caption](https://arxiv.org/html/2608.11605v1/architecture_1.png)Figure 2:Dynamics\-conditioned Action DiT\. During training, demonstrated future frames supervise the video flow objective and a frozen latent\-action encoder supplies the LaWM target\. At inference, the future frames and teacher path are absent: the current latent is retained, future slots are initialized with noise, and one video prefill produces the per\-layer K/V cache read during action denoising\.### 3\.1Problem Formulation

We consider language\-conditioned chunk\-level control\. At control time, the policy receives a synchronized multi\-camera observationoo, an instructionll, and a proprioceptive statepp\. It predicts an executable action chunka1:H∈ℝH×dacta\_\{1:H\}\\in\\mathbb\{R\}^\{H\\times d\_\{\\mathrm\{act\}\}\}of horizonHH\. A direct policy models

pθ\(a1:H∣o,l,p\)\.p\_\{\\theta\}\(a\_\{1:H\}\\mid o,l,p\)\.\(1\)At inference, future observations, privileged simulator state, and teacher outputs are unavailable\.

Letu1:Tu\_\{1:T\}denote a future visual trajectory or its latent representation\. An explicit\-future WAM may factorize action prediction conceptually as

p\(a1:H∣o,l,p\)=∫pϕ\(u1:T∣o,l,p\)pθ\(a1:H∣o,l,p,u1:T\)du1:T\.p\(a\_\{1:H\}\\mid o,l,p\)=\\int p\_\{\\phi\}\(u\_\{1:T\}\\mid o,l,p\)p\_\{\\theta\}\(a\_\{1:H\}\\mid o,l,p,u\_\{1:T\}\)\\,\\mathrm\{d\}u\_\{1:T\}\.\(2\)This factorization is commonly approximated by generating a future representation before or together with the action\. It exposes temporal context, but couples control latency to future generation\. A direct\-policy WAM can instead retain a future\-video training objective while omitting future rollout at inference\([48](https://arxiv.org/html/2608.11605#bib.bib20)\)\. Our problem is to retain this direct policy while giving its Action DiT an explicit route to predictive visual context\.

We distinguish the teacher\-forced training target from the deployment\-time interface\. Letz1:Tz\_\{1:T\}denote the VAE encoding of the demonstrated video segment used by the video flow\-matching loss\. During training, the video branch uses this target; at deployment, we construct a stochastic substratez~Fsub\\widetilde\{z\}^\{\\mathrm\{Fsub\}\}without observing the future segment, and expose its hidden per\-layer K/V stateℋKV\\mathcal\{H\}\_\{\\mathrm\{KV\}\}together with its dynamics\-register sliceDθD\_\{\\theta\}to the Action DiT\. Given a current\-frame latentzcur​\(o\)z\_\{\\mathrm\{cur\}\}\(o\), the substrate is

z~1:TFsub=concat\(zcur\(o\),ϵF\),ϵF∼𝒩\(0,I\)\\widetilde\{z\}^\{\\mathrm\{Fsub\}\}\_\{1:T\}=\\operatorname\{concat\}\\\!\\left\(z\_\{\\mathrm\{cur\}\}\(o\),\\epsilon\_\{F\}\\right\),\\qquad\\epsilon\_\{F\}\\sim\\mathcal\{N\}\(0,I\)\(3\)where the current latent occupies the first position andϵF\\epsilon\_\{F\}fills the future slots\. A single video prefill produces the dynamics\-register states and their per\-layer cache:

\(Dθ,ℋKV\)=KVPrefillϕ\(z~1:TFsub,l,p\)\\left\(D\_\{\\theta\},\\mathcal\{H\}\_\{\\mathrm\{KV\}\}\\right\)=\\operatorname\{KVPrefill\}\_\{\\phi\}\\left\(\\widetilde\{z\}^\{\\mathrm\{Fsub\}\}\_\{1:T\},l,p\\right\)\(4\)whereDθD\_\{\\theta\}denotes the dynamics\-register slice of the prefetched video state\. The resulting deployment\-time policy is

pθ\(a1:H∣o,l,p,Dθ\(o,l,p,ϵF\),ℋKV\(o,l,p,ϵF\)\)p\_\{\\theta\}\\\!\\left\(a\_\{1:H\}\\mid o,l,p,D\_\{\\theta\}\(o,l,p,\\epsilon\_\{F\}\),\\mathcal\{H\}\_\{\\mathrm\{KV\}\}\(o,l,p,\\epsilon\_\{F\}\)\\right\)\(5\)Equation[5](https://arxiv.org/html/2608.11605#S3.E5)remains a direct action policy: it conditions on neither a ground\-truth future nor a decoded video\. The stochastic future slots are an internal conditioning substrate, and their usefulness is learned from the joint video–action objective rather than from future observations at deployment\.

### 3\.2Model Architecture

#### Design rationale\.

Direct\-policy WAMs eliminate the iterative cost of generating future video, but this efficiency also leaves the action expert without an explicit, action\-facing representation of how the scene may evolve\. When the Action DiT is conditioned primarily on features of the current observation, it must infer both the present scene configuration and the consequences of candidate actions from the same visual context\. This is particularly challenging for interaction\-dependent behaviors, such as grasping, pushing, and placing, in which the appropriate action depends on the state transition induced by physical contact\. We therefore seek to retain direct action prediction while providing the action expert with hidden features that encode task\-relevant temporal structure, without access to future observations or decoded future video at inference time\.

Our model addresses this challenge through two complementary context pathways\. First, Future\-KV provides distributed visual context over the current frame and future latent slots\. The video backbone preserves the clean latent of the current frame, initializes the future slots with noise, and performs a single prefill\. The resulting layer\-wise keys and values are cached and made available to the Action DiT throughout action denoising\. Because the cache is maintained in feature space, Future\-KV exposes spatiotemporal context without requiring an iterative future\-video rollout or pixel\-space reconstruction\.

Second, we apply latent\-action \(LA\) supervision to a compact set of dynamics registers\. A frozen LaWM teacher maps the demonstrated visual transition to a latent\-action target, and a trainable projection head encourages the dynamics registers to match this target\. This supervision biases the registers towards interaction\-relevant changes, rather than requiring the action expert to recover such information solely from a generic future\-video objective\. The latent\-action target serves as a non\-executable transition cue and is used only during training\.

The two pathways impose different inductive biases\. Future\-KV preserves rich, distributed visual information, whereas the LA\-supervised dynamics registers provide a compact, action\-oriented summary of transition structure\. The Action DiT reads both pathways through the structured attention routing described below\. At inference, actions are predicted directly from the current observation and these hidden representations; neither future observations nor the LaWM teacher is available, and no future video is decoded\. Sec\.[4\.4](https://arxiv.org/html/2608.11605#S4.SS4)evaluates the corresponding component configurations, including a coverage\-distinct base\-policy reference without Future\-KV or LA supervision\. The observed complementarity is therefore a configuration\-level result rather than a fully matched causal conclusion\.

#### Token groups and routing\.

The reported configuration uses four token groups: current\-frame tokensCC, dynamics registersD=\{Di\}i=1NDD=\\\{D\_\{i\}\\\}\_\{i=1\}^\{N\_\{D\}\}, future\-slot tokensFF, and action tokensAA\. Readability registers are disabled\. The current observation is encoded intoCC, andFFoccupies the latent positions initialized in Eq\.[3](https://arxiv.org/html/2608.11605#S3.E3)\. The Action DiT receives a noisy action chunk and predicts its flow\. Both branches use the Wan2\.1 text condition; the proprioceptive state is projected into the conditioning space as an additional context token\.

The structured attention mask routes information as Figure[3](https://arxiv.org/html/2608.11605#S3.F3)\.

Figure 3:The structured mask routes current tokensCC, dynamics registersDD, future\-slot tokensFF, and action tokensAA\.Thus, future\-slot tokens can integrate the current frame and dynamics registers, and action tokens can read the complete video sequence together with the registers\. In the implementation, each action query concatenates the cached video keys and values with the keys and values computed from the current action tokens at that denoising step\. The mask defines architectural routing; it is not by itself evidence of disentanglement or causal sufficiency\.

#### Future\-KV prefill\.

During training, the video branch receives demonstrated future latents and learns a future\-latent flow objective, so its intermediate states receive a temporal learning signal\. At inference, we preserve the clean current latent, place pure noise in future slots, and run the video branch once at the prefill levelσ=1\.0\\sigma=1\.0\. We cache the resulting key and value tensors at every layer and reuse them throughout action denoising\. Future\-KV therefore incurs one video prefill per action query instead of an iterative future\-video rollout\. The cached states are hidden conditioning features, not realized future frames; no future observation is decoded or fed back into the control loop\. In the end\-to\-end configuration, gradients from the action loss remain connected to this prefill during training\.

#### Latent\-action\-supervised dynamics registers\.

Generic video supervision need not preferentially retain interaction\-relevant change\. We therefore use a frozen LaWM latent\-action encoder, trained as an inverse\-dynamics component, to encode the demonstrated visual transition as a quantized latent\-action targetzLAz\_\{\\mathrm\{LA\}\}during training\. The mean\-pooled dynamics registers pass through a trainable projectiongψg\_\{\\psi\}into the teacher space\. This target describes a visual transition; it is neither passed to the policy at deployment nor interpreted as a motor command\. Executable actions remain the output of the Action DiT\. The LA path is thus a training\-time shaping signal for a compact register interface, not a second action decoder\.

At inference, the policy encodes the current observation, builds the stochastic future substrate, prefills the cache once, and denoises the action chunk while readingCC,DD, andℋKV\\mathcal\{H\}\_\{\\mathrm\{KV\}\}\. The teacher and observed future transition are absent from this computation\.

### 3\.3Training Objective

We train the video and action branches with continuous flow matching\([27](https://arxiv.org/html/2608.11605#bib.bib15)\)\. For a targetyy, either a future video latent or an action chunk, we draw noiseϵ\\epsilonand a time variablett, and form

yt=\(1−t\)​y\+t​ϵy\_\{t\}=\(1\-t\)y\+t\\epsilon\(6\)The target velocity isϵ−y\\epsilon\-y, giving

ℒFM​\(y\)=𝔼y,ϵ,t​\[‖fθ​\(yt,t,o,l,p\)−\(ϵ−y\)‖22\]\\mathcal\{L\}\_\{\\mathrm\{FM\}\}\(y\)=\\mathbb\{E\}\_\{y,\\epsilon,t\}\\left\[\\left\\\|f\_\{\\theta\}\(y\_\{t\},t,o,l,p\)\-\(\\epsilon\-y\)\\right\\\|\_\{2\}^\{2\}\\right\]\(7\)
The video and action losses are

ℒvideo=ℒFM\(z1:T\),ℒaction=ℒFM\(a1:H\),\\mathcal\{L\}\_\{\\mathrm\{video\}\}=\\mathcal\{L\}\_\{\\mathrm\{FM\}\}\(z\_\{1:T\}\),\\qquad\\mathcal\{L\}\_\{\\mathrm\{action\}\}=\\mathcal\{L\}\_\{\\mathrm\{FM\}\}\(a\_\{1:H\}\),\(8\)wherez1:Tz\_\{1:T\}is the demonstrated video\-latent target anda1:Ha\_\{1:H\}is the demonstrated executable action chunk\. The frozen teacher supplies a detached targetzLAz\_\{\\mathrm\{LA\}\}\. With mean\-pooled dynamics registers, the distillation loss is

ℒLA=‖gψ​\(1ND​∑i=1NDDi\)−sg⁡\(zLA\)‖22\\mathcal\{L\}\_\{\\mathrm\{LA\}\}=\\left\\\|g\_\{\\psi\}\\\!\\left\(\\frac\{1\}\{N\_\{D\}\}\\sum\_\{i=1\}^\{N\_\{D\}\}D\_\{i\}\\right\)\-\\operatorname\{sg\}\(z\_\{\\mathrm\{LA\}\}\)\\right\\\|\_\{2\}^\{2\}\(9\)The stop\-gradient applies to the teacher target only\. In the reported end\-to\-end configuration, gradients from the action objective can flow through the video\-to\-action K/V interface\.

The total objective is

ℒ=ℒvideo\+ℒaction\+λLA​ℒLA\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{video\}\}\+\\mathcal\{L\}\_\{\\mathrm\{action\}\}\+\\lambda\_\{\\mathrm\{LA\}\}\\mathcal\{L\}\_\{\\mathrm\{LA\}\}\(10\)The three terms train future latent prediction, executable action generation, and the transition\-oriented register bottleneck, respectively\. The objective does not establish that the registers are causally necessary or that the latent action is executable; those properties require targeted interventions\.

## 4Experiments

### 4\.1Experimental Setup

#### Benchmarks and evaluation protocol\.

We evaluate in\-distribution control on the four standard LIBERO suites: Spatial, Object, Goal, and Long\([28](https://arxiv.org/html/2608.11605#bib.bib16)\)\. We report task success rate over 50 rollouts per task\. We evaluate out\-of\-distribution robustness on LIBERO\-Plus\([16](https://arxiv.org/html/2608.11605#bib.bib12)\), which perturbs the original tasks along seven dimensions: camera viewpoint, robot initial state, language instruction, lighting, background texture, sensor noise, and object layout\.

### 4\.2Main Results

#### Results on LIBERO\.

Both variants retain strong in\-distribution performance without embodied pretraining \(Table[4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1)\)\. Ours achieves 96\.7% overall, ranging from 92\.8% on Long to 99\.6% on Object\. Ours\-Flash reaches 96\.9% overall and differs from Ours by at most 0\.8 percentage points on any suite\. The two variants are 0\.9 and 0\.7 points below Fast\-WAM, respectively\. Thus, the accelerated variant preserves the standard\-LIBERO performance of the full inference configuration\.

Table 1: Success rate \(%\) on the standard LIBERO suites, evaluated with 50 rollouts per task\.

#### Robustness on LIBERO\-Plus\.

On the observed LIBERO\-Plus subset \(Table[4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2)\), Ours reaches 61\.6% overall, with the highest rates under lighting and language perturbations and the lowest rate under robot\-initial\-state shifts\. Compared with Fast\-WAM, Ours is 10\.1 points higher overall; the largest gains are on camera viewpoint \(\+46\.1 points\) and sensor noise \(\+21\.1 points\), with smaller gains on object layout, language, and background texture but lower success on robot\-initial\-state shifts and lighting\. Ours\-Flash reaches 58\.2%, 3\.4 points below Ours and 6\.7 points above Fast\-WAM overall, while remaining lower than Fast\-WAM on robot\-initial\-state, language, lighting, and background shifts\. Because external results come from different sources, these cross\-method differences are descriptive rather than coverage\-matched causal estimates\.

Table 2: Observed success rate \(%\) on seven LIBERO\-Plus perturbation categories\.

### 4\.3Inference Efficiency

#### Action\-denoising latency\.

Table[4\.4](https://arxiv.org/html/2608.11605#S4.SS4.SSS0.Px1)reports standalone action\-generation inference latency measured on a single NVIDIA A800 GPU with 80 GB of memory\. Ours lowers the 10\-step latency from 667 ms for FastWAM to 568 ms\. Distilling the Ours action branch from 10 to 2 denoising steps yields Ours\-Flash, which reaches 220 ms while retaining the Future\-KV and dynamics\-register interface, a 61% reduction relative to Ours\. These are standalone inference measurements, not average task\-completion times or LIBERO\-Plus rollout statistics\.

### 4\.4Ablation Study

#### Component comparison on LIBERO\-Plus\.

Among the three coverage\-matched configurations, each with 1,482 observed evaluations, Ours achieves the strongest overall result at 61\.6% \(Table[4\.4](https://arxiv.org/html/2608.11605#S4.SS4.SSS0.Px1)\)\. It exceeds Future\-KV only \(58\.5%\) and LA supervision only \(58\.0%\) by 3\.1 and 3\.6 percentage points, respectively\. The Base policy uses neither Future\-KV nor LA supervision and reaches 53\.6% over 10,027 observed evaluations under a different coverage profile\. We therefore include it as a contextual reference rather than a matched estimate of the gain from adding both components\. All configurations use no embodied pretraining, and the aggregate comparison does not by itself establish the causal contribution of either pathway\.

Table 3: Standalone action\-generation inference latency\.

Table 4: Overall observed success rate \(%\) for the LIBERO\-Plus ablation\.

## 5Limitations and Discussion

Our evaluation is currently limited to the standard LIBERO suites and LIBERO\-Plus\. Although these benchmarks cover a range of manipulation tasks and robustness perturbations, they do not fully capture the diversity of embodiments, interaction dynamics, visual conditions, and long\-horizon behaviors encountered in broader robotic settings\. It therefore remains unclear how well ForeWAM generalizes to different robot morphologies, task distributions, or real\-world deployment scenarios\. In particular, the robustness gains observed on LIBERO\-Plus should be interpreted within the evaluated subset rather than as evidence of universal out\-of\-distribution generalization\.

## 6Conclusion

We introducedForeWAM, a dynamics\-conditioned direct\-policy World Action Model that provides predictive context for action generation without explicit future\-video rollout\. ForeWAM combinesFuture\-KVwithlatent\-action\-supervised dynamics registers, enabling the Action DiT to access distributed future context and compact transition cues\. Future observations and the latent\-action teacher are used only during training\.

ForeWAM achieves 96\.7% average success on LIBERO and 61\.6% on LIBERO\-Plus, while ForeWAM\-Flash reaches 96\.9% on LIBERO with substantially lower action\-generation latency\. Component comparisons further show that combining the two pathways performs better than either alone\. These results suggest that predictive dynamics can benefit direct action policies without being explicitly materialized as future observations\.

## References

- Bharadhwajet al\.\(2024\)H\. Bharadhwaj, R\. Mottaghi, A\. Gupta, and S\. TulsianiTrack2act: predicting point tracks from internet videos enables generalizable robot manipulation\.InEuropean Conference on Computer Vision,pp\. 306–324\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Biet al\.\(2026\)H\. Bi, H\. Tan, S\. Xie, Z\. Wang, S\. Huang, H\. Liu, R\. Zhao, Y\. Feng, C\. Xiang, Y\. Rong,et al\.Motus: a unified latent action world model\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 35101–35113\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p2.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Bjorcket al\.\(2025\)J\. Bjorck, F\. Castañeda, N\. Cherniadev, X\. Da, R\. Ding, L\. Fan, Y\. Fang, D\. Fox, F\. Hu, S\. Huang,et al\.Gr00t n1: an open foundation model for generalist humanoid robots\.arXiv preprint arXiv:2503\.14734\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Blacket al\.\(2024\)K\. Black, N\. Brown, D\. Driess, A\. Esmail, M\. Equi, C\. Finn, N\. Fusai, L\. Groom, K\. Hausman, B\. Ichter,et al\.π0\\pi\_\{0\}: A vision\-language\-action flow model for general robot control\.arXiv preprint arXiv:2410\.24164\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.3.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.3.1)\.
- Brohanet al\.\(2023\)A\. Brohan, N\. Brown, J\. Carbajal, Y\. Chebotar, X\. Chen, K\. Choromanski, T\. Ding, D\. Driess, A\. Dubey, C\. Finn,et al\.Rt\-2: vision\-language\-action models transfer web knowledge to robotic control\.arXiv preprint arXiv:2307\.15818\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Brohanet al\.\(2022\)A\. Brohan, N\. Brown, J\. Carbajal, Y\. Chebotar, J\. Dabis, C\. Finn, K\. Gopalakrishnan, K\. Hausman, A\. Herzog, J\. Hsu,et al\.Rt\-1: robotics transformer for real\-world control at scale\.arXiv preprint arXiv:2212\.06817\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Buet al\.\(2025\)Q\. Bu, Y\. Yang, J\. Cai, S\. Gao, G\. Ren, M\. Yao, P\. Luo, and H\. LiUnivla: learning to act anywhere with task\-centric latent actions, 2025\.URL https://arxiv\. org/abs/2505\.06111\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.6.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.6.1)\.
- Cenet al\.\(2025a\)J\. Cen, S\. Huang, Y\. Yuan, K\. Li, H\. Yuan, C\. Yu, B\. Hou, Y\. Jiang, J\. Guo, X\. Li,et al\.Rynnvla\-002: a unified vision\-language\-action and world model\.arXiv preprint arXiv:2511\.17502\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Cenet al\.\(2025b\)J\. Cen, C\. Yu, H\. Yuan, Y\. Jiang, S\. Huang, J\. Guo, X\. Li, Y\. Song, H\. Luo, F\. Wang,et al\.Worldvla: towards autoregressive action world model\.arXiv preprint arXiv:2506\.21539\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.7.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.7.1)\.
- Cheanget al\.\(2024\)C\. Cheang, G\. Chen, Y\. Jing, T\. Kong, H\. Li, Y\. Li, Y\. Liu, H\. Wu, J\. Xu, Y\. Yang,et al\.Gr\-2: a generative video\-language\-action model with web\-scale knowledge for robot manipulation\.arXiv preprint arXiv:2410\.06158\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Chenet al\.\(2026a\)J\. Chen, K\. Wang, K\. Chen, S\. Chen, F\. Gao, W\. Tang, Z\. Li, W\. Liu, Z\. Yao, B\. Li,et al\.Lawam: latent world action models for efficient dynamics\-aware robot policies\.arXiv preprint arXiv:2606\.15768\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p5.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p2.1)\.
- Chenet al\.\(2026b\)J\. Chen, W\. Song, P\. Ding, Z\. Zhou, H\. Zhao, B\. Tang, D\. Wang, and H\. LiUnified diffusion vla: vision\-language\-action model via joint discrete denosing diffusion process\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 139291–139311\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Chiet al\.\(2025\)C\. Chi, Z\. Xu, S\. Feng, E\. Cousineau, Y\. Du, B\. Burchfiel, R\. Tedrake, and S\. SongDiffusion policy: visuomotor policy learning via action diffusion\.The International Journal of Robotics Research44\(10\-11\),pp\. 1684–1704\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Duet al\.\(2023\)Y\. Du, S\. Yang, B\. Dai, H\. Dai, O\. Nachum, J\. Tenenbaum, D\. Schuurmans, and P\. AbbeelLearning universal policies via text\-guided video generation\.Advances in neural information processing systems36,pp\. 9156–9172\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p1.1),[§1](https://arxiv.org/html/2608.11605#S1.p2.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Duet al\.\(2024\)Y\. Du, S\. Yang, P\. Florence, F\. Xia, A\. Wahid, P\. Sermanet, T\. Yu, P\. Abbeel, J\. B\. Tenenbaum, L\. Kaelbling,et al\.Video language planning\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 31138–31155\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Feiet al\.\(2025\)S\. Fei, S\. Wang, J\. Shi, Z\. Dai, J\. Cai, P\. Qian, L\. Ji, X\. He, S\. Zhang, Z\. Fei,et al\.Libero\-plus: in\-depth robustness analysis of vision\-language\-action models\.arXiv preprint arXiv:2510\.13626\.Cited by:[§4\.1](https://arxiv.org/html/2608.11605#S4.SS1.SSS0.Px1.p1.1)\.
- Guoet al\.\(2024\)Y\. Guo, Y\. Hu, J\. Zhang, Y\. Wang, X\. Chen, C\. Lu, and J\. ChenPrediction with action: visual policy learning via joint denoising process\.Advances in Neural Information Processing Systems37,pp\. 112386–112410\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Huet al\.\(2024\)Y\. Hu, Y\. Guo, P\. Wang, X\. Chen, Y\. Wang, J\. Zhang, K\. Sreenath, C\. Lu, and J\. ChenVideo prediction policy: a generalist robot policy with predictive visual representations\.arXiv preprint arXiv:2412\.14803\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p1.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Huang and Zheng \(2025\)J\. Huang and W\. ZhengSize\-aware contrastive imitation learning for language\-conditioned multi\-task robotic manipulation\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Huanget al\.\(2024\)S\. Huang, M\. Levy, Z\. Jiang, A\. Anandkumar, Y\. Zhu, L\. Fan, D\. Huang, and A\. ShrivastavaArdup: active region video diffusion for universal policies\.In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems \(IROS\),pp\. 8465–8472\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Intelligenceet al\.\(2026\)P\. Intelligence, B\. Ai, A\. Amin, R\. Aniceto, A\. Balakrishna, G\. Balke, K\. Black, G\. Bokinsky, S\. Cao, T\. Charbonnier,et al\.π\\pi0\. 7: a steerable generalist robotic foundation model with emergent capabilities, 2026\.URL https://arxiv\. org/abs/2604\.15483\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Intelligenceet al\.\(2025\)P\. Intelligence, K\. Black, N\. Brown, J\. Darpinian, K\. Dhabalia, D\. Driess, A\. Esmail, M\. Equi, C\. Finn, N\. Fusai,et al\.π\\pi0\. 5: a vision\-language\-action model with open\-world generalization, 2025\.URL https://arxiv\. org/abs/2504\.160541\(2\),pp\. 3\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.4.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.4.1)\.
- Kimet al\.\(2026\)M\. J\. Kim, Y\. Gao, T\. Lin, Y\. Lin, Y\. Ge, G\. Lam, P\. Liang, S\. Song, M\. Liu, C\. Finn,et al\.Cosmos policy: fine\-tuning video models for visuomotor control and planning\.arXiv preprint arXiv:2601\.16163\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Kimet al\.\(2024\)M\. J\. Kim, K\. Pertsch, S\. Karamcheti, T\. Xiao, A\. Balakrishna, S\. Nair, R\. Rafailov, E\. Foster, G\. Lam, P\. Sanketi,et al\.Openvla: an open\-source vision\-language\-action model\.arXiv preprint arXiv:2406\.09246\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.2.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.2.1)\.
- Koet al\.\(2024\)P\. Ko, J\. Mao, Y\. Du, S\. Sun, and J\. B\. TenenbaumLearning to act from actionless videos through dense correspondences\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 40938–40958\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2026\)R\. Li, H\. Zhang, J\. Jin, Q\. Zeng, Z\. Zhuang, Y\. Tang, S\. Lyu, and D\. WangWorld\-value\-action model: implicit planning for vision\-language\-action systems\.arXiv preprint arXiv:2604\.14732\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Lipmanet al\.\(2022\)Y\. Lipman, R\. T\. Chen, H\. Ben\-Hamu, M\. Nickel, and M\. LeFlow matching for generative modeling\.arXiv preprint arXiv:2210\.02747\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1),[§3\.3](https://arxiv.org/html/2608.11605#S3.SS3.p1.1)\.
- Liuet al\.\(2023\)B\. Liu, Y\. Zhu, C\. Gao, Y\. Feng, Q\. Liu, Y\. Zhu, and P\. StoneLibero: benchmarking knowledge transfer for lifelong robot learning\.Advances in Neural Information Processing Systems36,pp\. 44776–44791\.Cited by:[§4\.1](https://arxiv.org/html/2608.11605#S4.SS1.SSS0.Px1.p1.1)\.
- Liuet al\.\(2025\)S\. Liu, L\. Wu, B\. Li, H\. Tan, H\. Chen, Z\. Wang, K\. Xu, H\. Su, and J\. ZhuRdt\-1b: a diffusion foundation model for bimanual manipulation\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 29982–30009\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Louet al\.\(2026\)Y\. Lou, X\. Chi, X\. Zhang, Z\. Qian, C\. Li, R\. Zhang, Y\. Lyu, G\. Song, C\. Fu, H\. Xu,et al\.Mask world model: predicting what matters for robust robot policy learning\.arXiv preprint arXiv:2604\.19683\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Lyuet al\.\(2026\)J\. Lyu, K\. Liu, X\. Zhang, H\. Liao, Y\. Feng, W\. Zhu, T\. Shen, J\. Chen, J\. Zhang, Y\. Dong,et al\.Lda\-1b: scaling latent dynamics action model via universal embodied data ingestion\.arXiv preprint arXiv:2602\.12215\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Pertschet al\.\(2025\)K\. Pertsch, K\. Stachowicz, B\. Ichter, D\. Driess, S\. Nair, Q\. Vuong, O\. Mees, C\. Finn, and S\. LevineFast: efficient action tokenization for vision\-language\-action models\.arXiv preprint arXiv:2501\.09747\.Cited by:[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.5.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.5.1)\.
- \[33\]S\. L\. Y\. G\. D\. Sadigh and S\. SongUnified video action model\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p1.1)\.
- Shenet al\.\(2026\)Y\. Shen, F\. Wei, Z\. Du, Y\. Liang, Y\. Lu, J\. Yang, N\. Zheng, and B\. GuoVideovla: video generators can be generalizable robot manipulators\.Advances in neural information processing systems38,pp\. 95597–95621\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Teamet al\.\(2026\)K\. Team, F\. Wang, S\. You, Q\. Zhang, T\. Huang, Z\. Fu, Z\. Zheng, Y\. Xi, F\. Lv, X\. Wu, Z\. Liu, C\. Wan, P\. Li, R\. Yang, X\. Li, W\. Wang, K\. Zhu, Y\. Zhang, S\. Fu, Z\. Zhang, X\. Wu, X\. Fan, D\. Tao, and X\. WangKairos: a regret\-aware native world\-action model stack for physical ai\.External Links:2606\.16533,[Link](https://arxiv.org/abs/2606.16533)Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Teamet al\.\(2024\)O\. M\. Team, D\. Ghosh, H\. Walke, K\. Pertsch, K\. Black, O\. Mees, S\. Dasari, J\. Hejna, T\. Kreiman, C\. Xu,et al\.Octo: an open\-source generalist robot policy\.arXiv preprint arXiv:2405\.12213\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Wan Team \(2025\)Wan TeamWan: open and advanced large\-scale video generative models\.arXiv preprint arXiv:2503\.20314\.Cited by:[Appendix A](https://arxiv.org/html/2608.11605#A1.SS0.SSS0.Px1.p1.1)\.
- Wanget al\.\(2024\)Z\. Wang, Z\. Li, A\. Mandlekar, Z\. Xu, J\. Fan, Y\. Narang, L\. Fan, Y\. Zhu, Y\. Balaji, M\. Zhou,et al\.One\-step diffusion policy: fast visuomotor policies via diffusion distillation\.arXiv preprint arXiv:2410\.21257\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p6.1)\.
- Wonet al\.\(2025\)J\. Won, K\. Lee, H\. Jang, D\. Kim, and J\. ShinDual\-stream diffusion for world\-model augmented vision\-language\-action model\.arXiv preprint arXiv:2510\.27607\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Wuet al\.\(2024\)H\. Wu, Y\. Jing, C\. Cheang, G\. Chen, J\. Xu, X\. Li, M\. Liu, H\. Li, and T\. KongUnleashing large\-scale video generative pre\-training for visual robot manipulation\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 10641–10662\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Xuet al\.\(2024\)M\. Xu, Z\. Xu, Y\. Xu, C\. Chi, G\. Wetzstein, M\. Veloso, and S\. SongFlow as the cross\-domain manipulation interface\.arXiv preprint arXiv:2407\.15208\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yanet al\.\(2026\)H\. Yan, Z\. Zhong, J\. Zhu, J\. He, W\. Yuan, W\. Song, X\. Gong, Y\. Cai, G\. Zhao, X\. Yan,et al\.S\-vam: shortcut video\-action model by self\-distilling geometric and semantic foresight\.arXiv preprint arXiv:2603\.16195\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yanget al\.\(2025\)L\. Yang, Y\. Bai, G\. Eskandar, F\. Shen, M\. Altillawi, D\. Chen, Z\. Liu, and A\. ValadaCoVAR: co\-generation of video and action for robotic manipulation via multi\-modal diffusion\.arXiv preprint arXiv:2512\.16023\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yanget al\.\(2026\)Y\. Yang, S\. Zeng, T\. Lin, X\. Chang, D\. Qi, J\. Xiao, H\. Liu, R\. Chen, Y\. Chen, D\. Huo,et al\.Abot\-m0: vla foundation model for robotic manipulation with action manifold learning\.arXiv preprint arXiv:2602\.11236\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Yeet al\.\(2026a\)A\. Ye, B\. Wang, C\. Ni, G\. Huang, G\. Zhao, H\. Li, H\. Li, J\. Li, J\. Lv, J\. Liu,et al\.GigaWorld\-policy: an efficient action\-centered world–action model\.arXiv preprint arXiv:2603\.17240\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yeet al\.\(2026b\)S\. Ye, Y\. Ge, K\. Zheng, S\. Gao, S\. Yu, G\. Kurian, S\. Indupuru, Y\. L\. Tan, C\. Zhu, J\. Xiang,et al\.World action models are zero\-shot policies\.arXiv preprint arXiv:2602\.15922\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p1.1),[§1](https://arxiv.org/html/2608.11605#S1.p2.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yuanet al\.\(2026a\)G\. Yuan, Q\. Qiao, J\. Zhang, and D\. XuAdaWorldPolicy: world\-model\-driven diffusion policy with online adaptive learning for robotic manipulation\.arXiv preprint arXiv:2602\.20057\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Yuanet al\.\(2026b\)T\. Yuan, Z\. Dong, Y\. Liu, and H\. ZhaoFast\-wam: do world action models need test\-time future imagination?\.arXiv preprint arXiv:2603\.16666\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p2.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.11605#S3.SS1.p2.2),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px1.tab1.2.8.1),[§4\.2](https://arxiv.org/html/2608.11605#S4.SS2.SSS0.Px2.tab1.2.8.1)\.
- Zhaoet al\.\(2025\)Q\. Zhao, Y\. Lu, M\. J\. Kim, Z\. Fu, Z\. Zhang, Y\. Wu, Z\. Li, Q\. Ma, S\. Han, C\. Finn,et al\.Cot\-vla: visual chain\-of\-thought reasoning for vision\-language\-action models\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 1702–1713\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Zhenget al\.\(2026\)J\. Zheng, J\. Li, Z\. Wang, D\. Liu, X\. Kang, Y\. Feng, Y\. Zheng, J\. Zou, Y\. Chen, J\. Zeng,et al\.X\-vla: soft\-prompted transformer as scalable cross\-embodiment vision\-language\-action model\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 60580–60606\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px1.p1.1)\.
- Zhiet al\.\(2025\)H\. Zhi, P\. Chen, S\. Zhou, Y\. Dong, Q\. Wu, L\. Han, and M\. Tan3dflowaction: learning cross\-embodiment manipulation from 3d flow world model\.arXiv preprint arXiv:2506\.06199\.Cited by:[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.
- Zhuet al\.\(2025\)C\. Zhu, R\. Yu, S\. Feng, B\. Burchfiel, P\. Shah, and A\. GuptaUnified world models: coupling video and action diffusion for pretraining on large robotic datasets\.arXiv preprint arXiv:2504\.02792\.Cited by:[§1](https://arxiv.org/html/2608.11605#S1.p1.1),[§2](https://arxiv.org/html/2608.11605#S2.SS0.SSS0.Px2.p1.1)\.

## Appendix AImplementation Details

#### Architecture and inputs\.

We initialize the visual branch from Wan2\.1\-T2V\-1\.3B, retaining its video DiT, text encoder, and video VAE\([37](https://arxiv.org/html/2608.11605#bib.bib3)\)\. We precompute instruction embeddings with the corresponding Wan2\.1 text encoder\. Both the video DiT and the action expert comprise 30 transformer blocks\. The video branch uses hidden dimensiondv=1536d\_\{v\}=1536, whereas the action expert usesda=1024d\_\{a\}=1024and is initialized from a linearly interpolated Wan2\.1 ActionDiT checkpoint\. The action horizon isH=32H=32\.

Each training example contains 33 observation frames\. A temporal ratio of 4 between the action and video streams maps each 32\-step action chunk to 9 video frames\. We concatenate the two synchronized camera views along the image width before VAE encoding, producing a224×448224\\times 448image composed of two224×224224\\times 224views\. The policy additionally receives an 8\-dimensional proprioceptive state\. Each action is seven\-dimensional, comprising a 6\-DoF end\-effector pose and one gripper\-control dimension\.

#### Optimization and inference\.

We train the video and action branches with continuous flow matching using a 1,000\-timestep schedule and a shift of 5\.0\. The standard policy uses 10 action denoising steps at inference; Ours\-Flash applies the accelerated variant of the same interface\. We disable readability registers and useND=16N\_\{D\}=16dynamics registers\. A frozen LaWAM teacher supplies a 32\-dimensional latent\-action target\. Gradients from the action objective propagate through the video\-to\-action interface without stop\-gradient\. At inference, we retain the current latent, initialize future slots with noise, and prefill the video K/V cache once atσ=1\.0\\sigma=1\.0\. This cache is reused across all action\-denoising steps\.

We optimize the joint objective with AdamW using a learning rate of1×10−41\\times 10^\{\-4\}, weight decay of 0\.01, cosine annealing, and gradient clipping at 1\.0\. None of the reported variants receives embodied pretraining before LIBERO training\.