PAPER REPORTENAll readings ↗

World Action Models: The Next Frontier in Embodied AI

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Siyin Wang; Junhao Shi; Zhaoyang Fu; Xinzhe He; Feihong Liu; Chenchen Yang; Yikang Zhou; Zhaoye Fei; Jingjing Gong; Jinlan Fu; Mike Zheng Shou; Xuanjing Huang; Xipeng Qiu; Yu-Gang Jiang

Affiliations: Fudan University; Shanghai Innovation Institute; National University of Singapore

Source: 2605.12090 ↗ · Project page ↗ · Catalog record

Reading: 197 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: The survey organizes predictive robot policies by how future states and actions are coupled, while exposing the lack of controlled evidence that distinguishes useful foresight from auxiliary supervision. e-definitione-cascadede-jointe-datae-evaluatione-controlled

At a glanceWhat to know
Research problem
Author claim

Reactive observation-to-action policies can exploit semantic priors without explicitly forecasting environmental change. The authors argue that embodied control needs predictive physical reasoning, but the literature mixes video backbones, simulators, planners, and policies under overlapping names. The survey asks what qualifies as a WAM, where prediction and control interact, and which evidence could distinguish useful foresight from visually plausible generation. e-identitye-definitione-boundaries

Core mechanism
Source description

A modality-independent definition requires both a quantifiable future-state representation and coupled action generation. Video Action Models occupy a narrower video-oriented scope; inheriting a video backbone alone does not satisfy the proposed predictive commitment. e-definitione-boundaries

Reading caution
Reader analysis

The authors state that cascaded, joint, discrete, and latent designs lack systematic comparisons under matched scale, data, and evaluation. Accordingly, this survey offers no controlled architecture ranking, new benchmark score table, or original ablation result. Its cited success and efficiency claims are not a common-protocol meta-analysis. e-controllede-scope

Core contributions

  • Source description

    A modality-independent definition requires both a quantifiable future-state representation and coupled action generation. Video Action Models occupy a narrower video-oriented scope; inheriting a video backbone alone does not satisfy the proposed predictive commitment. e-definitione-boundaries

  • Source description

    An architectural taxonomy distinguishes cascaded future-to-action pipelines from joint predictive systems. It further separates explicit and latent plans, autoregressive representations, and diffusion systems with a unified stream or multiple coupled streams. e-cascadede-autoregressivee-joint

  • Source description

    A data-and-evaluation map connects action-grounded demonstrations and action-free video to distinct learning signals, then separates visual fidelity, physical commonsense, action plausibility, and downstream policy evaluation. e-datae-evaluatione-policy

Figure 3. The proposed WAM boundary depends on predictive commitment and action coupling. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the left panel from the input circles upward. The VLA receives current observation and language, then emits an action. The world model instead receives an action as an input and predicts the next observation. The WAM sits between those roles: its outputs include both the action and anticipated observation. In the figure, $O_t$ denotes the current observation, $L$ language, $A$ action, and $O_{t+1}$ the next observation. On the right, VAM lies within the larger WAM scope, while Video Policy overlaps it. Section 2.2 explains why inheriting a video-generation backbone does not automatically imply the predictive objective required by this definition. e-definitione-boundariese-controlled

What it supports. The survey defines a WAM by what is predicted and how that prediction is connected to control. Its scope extends beyond video to other representations of future physical state. The conceptual circles help separate architectural ancestry from an explicitly supervised prediction of what happens next.

Where the evidence stops. These are the authors' proposed conceptual boundaries, not evidence that every included model learns causal physics or generalizes better. The drawing also abstracts away the training-versus-inference distinctions that become important in Sections 4 and 7.

2. Motivation

2.1 The problem and the proposed response

Author claim

Reactive observation-to-action policies can exploit semantic priors without explicitly forecasting environmental change. The authors argue that embodied control needs predictive physical reasoning, but the literature mixes video backbones, simulators, planners, and policies under overlapping names. The survey asks what qualifies as a WAM, where prediction and control interact, and which evidence could distinguish useful foresight from visually plausible generation. e-identitye-definitione-boundaries

2.2 What this reading follows

A robot can predict a convincing future without knowing how to reach it, and it can complete a task without demonstrating that its imagined future guided the action. This survey makes that gap the organizing problem for World Action Models. Begin with its input/output definition, follow the information path through cascaded and joint architectures, then examine what their training data can actually supervise. The final step is to choose an evaluation that tests the intended capability. The figures below are the survey's own conceptual diagrams and resource tables; they map a research landscape rather than report a new robot system's experimental victory. e-definitione-cascadede-jointe-datae-evaluatione-controlled

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryFoundational work
ArchitectureNot applicable
Prediction paradigmNot applicable
QuadrantNot applicable

This table preserves the labels recorded at reading time. The current major category is Related resources. View the current classification.

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The recorded foundational survey/resource classification is supported. This paper synthesizes multiple architectures rather than proposing one WAM implementation, so architecture, prediction paradigm, and quadrant are not applicable to the paper itself. Its 'Joint WAM' umbrella includes multiple streams and shared-encoder/separate-decoder systems; it must not be translated automatically into a catalog 'One Model' judgment for every cited work. e-scopee-jointe-shared

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • For the surveyed WAM formulation: observation o, including vision, proprioception, or other sensors; language instruction l.
  • For learning: aligned observation–action–future-observation examples, potentially supplemented with action-free observation sequences.
  • A predicted or utilized representation of future observation o′ and associated action a; these describe a model family, not a new executable system introduced by the survey.

4.2 Equations and their role

p(o,ao,l)=p(ao,o,l)p(oo,l)p(o',a\mid o,l)=p(a\mid o',o,l)\,p(o'\mid o,l)
The survey's cascaded factorization. Observation o contains current sensory information, l is the language instruction, o′ is the subsequent observation, and a is the action. The future predictor supplies the intermediate state used by the action model. The equation describes a probabilistic decomposition; it does not by itself certify physical feasibility or execution success. e-definitione-cascaded
VLA: p(ao,l),WM: p(oo,a),WAM: p(o,ao,l)\text{VLA: }p(a\mid o,l),\qquad \text{WM: }p(o'\mid o,a),\qquad \text{WAM: }p(o',a\mid o,l)
The three source-defined prediction targets distinguish choosing an action, forecasting a supplied intervention's consequence, and jointly characterizing future observations and actions. The labels identify the three model families; all variables retain the definitions above. e-definition

5. Method in detail

5.1 Identify the future that an action is supposed to realize

Reader analysis

The definition becomes useful when applied to an information path. A conventional world model receives an action and predicts its consequence; a policy chooses the action. In the survey's cascaded formulation, a predicted future bridges these roles. First locate the forecast: is it an RGB frame, geometric motion, or a latent representation? Then locate the mechanism that turns the current-to-future transition into control. UniPi's frame-pair IDM and VPP's latent-conditioned policy illustrate different interfaces. Reader interpretation: merely drawing an arrow from a future predictor to a policy does not prove that the policy relies on it. The conditioning path must be present, and its contribution must survive appropriate controls. This is why the probability factorization explains the design but cannot certify a physically reachable plan or successful execution. e-definitione-cascadede-latente-controlled

Figure 5. A cascaded WAM turns a predicted transition into commands through a separate decoder. Original paper, p. 15 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Compare the intermediate representation and the blue action block in each panel. In 1(a), Video Gen predicts RGB frames and the dotted route brings future and current observations to an inverse dynamics model, abbreviated IDM. Panel 1(b) retains the explicit visual plan but substitutes geometric extraction for learned inverse dynamics. Panel 2(a) changes the planning carrier to a latent future representation. The arrows still terminate at an action-producing stage. Read these pathways alongside the factorization on page 3: a future is predicted first, and action generation is conditioned on that future. Section 4.1 gives concrete examples, including UniPi's frame-pair IDM and latent planning in VPP. e-cascadede-definitione-latent

What it supports. The important interface is the representation passed from forecasting to control. Pixel plans permit direct visual inspection; latent plans avoid reconstructing the full visual output. Both require an action extraction mechanism, and the survey distinguishes learning that mapping from deriving motion geometrically.

Where the evidence stops. The blocks are family-level schematics. They do not specify a universal action horizon, feedback loop, or guarantee that a generated plan is executable. The IDM label in the latent panel summarizes an action-decoding role rather than prescribing every method's decoder architecture.

5.2 Separate future supervision from future computation at deployment

Source description

The joint family introduces a second distinction: what teaches the model need not be what it computes when acting. The survey describes FLARE's future tokens receiving targets from a frozen encoder of actual future observations. Those observations provide training supervision; they are not observations of the future available to a deployed robot. Fast-WAM provides a sharper example: it trains with a future-video objective, blocks action attention to future video latents, and removes future-video generation at inference. DreamZero instead retains joint prediction and replaces imagined frames with real observations after action chunks execute. These examples motivate reading the training graph, inference graph, and feedback schedule separately. Their inclusion under the same WAM umbrella does not make their control mechanisms equivalent, nor does joint supervision alone establish online planning. e-implicite-hiddene-feedbacke-controlled

Figure 6. Joint supervision can use one shared trunk or several explicitly coupled computational paths. Original paper, p. 21 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start at 1(a), where vision and action encoders feed a common Diffusion Transformer, or DiT. The V and A circles mark the world/video and action outputs. Move right to compare where information is shared. In 2(a), the Cross-Attention block spans separate video and action DiTs. In 2(b), the dashed Hidden State arrow points from the video DiT to the action DiT; it is not a bidirectional exchange. In 2(c), fusion occurs in the Unified Encoder before the two decoding paths. The caption and Section 4.2.2 support these distinctions, while page 24 explains how hidden-state transfer can persist even when full future-video sampling is bypassed. e-jointe-hiddene-sharede-implicit

What it supports. 'Joint' identifies a coupled predictive system, not a claim that all modalities traverse one homogeneous backbone. The survey's structural subdivisions make the coupling interface inspectable. They also show why a catalog judgment about one model versus separate models needs architecture evidence beyond a shared training objective.

Where the evidence stops. Figure 6 does not show method-specific masks, noise schedules, frozen teachers, or deployment pruning. Some surveyed models remove future-video generation at inference. The drawing therefore cannot establish that every action is conditioned on a freshly generated future.

5.3 Connect data supervision to a falsifiable evaluation claim

Reader analysis

The data diagram asks what information each source can teach. An action-free human video may reveal how objects move, yet converting that motion into robot commands still requires grounding. UMI-style demonstrations and aligned robot transitions help supply that connection, while simulation offers controllable geometric supervision. Reader interpretation: evaluation should test whichever connection the training recipe claims to improve. Perceptual similarity is relevant if the claim concerns visual fidelity; executable controls and their realized consequences matter if the claim concerns action grounding. Table 8 provides the metric families, and Table 9 supplies policy testbeds. Section 7 identifies the missing link between them: whether changing an imagined future changes actions appropriately, and whether the resulting trajectory actually follows the plan. More data and better-looking predictions cannot answer that question by themselves. e-datae-umie-simulatione-humane-evaluatione-policye-coupled-evaluation

Figure 7. Training sources offer different combinations of physical grounding and collection scale. Original paper, p. 26 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Use the endpoint annotations to orient this diagram. Although the axes are titled Scaling Difficulty and Transfer Difficulty, moving right reaches Cheap/Scalable, while moving upward reaches Easy to Match. Thus the robot-teleoperation point combines easier matching with a complex setup. Portable human demonstrations move toward easier collection while retaining useful physical grounding; simulation emphasizes scale with a domain gap; human and egocentric video extends coverage but is harder to match to robot control. Section 5 connects this picture to the supervision available: paired observation–action–future transitions couple prediction and action, while action-free observation sequences can supervise world prediction without directly supplying executable robot commands. e-datae-umie-humane-limitations

What it supports. The authors motivate mixing complementary sources, rather than treating raw volume as sufficient. High-quality paired transitions provide control grounding, while broader observations expand the dynamics and situations encountered during learning. The figure organizes that rationale; it supplies no measured optimal mixture.

Where the evidence stops. The axis titles invite an increasing-difficulty reading that conflicts with their endpoint directions. Follow the endpoints as described here. The curve and distances are qualitative: there are no calibrated costs, transfer scores, uncertainty estimates, or fitted scaling law.

5.4 Training and inference

During training

Source description

The survey supplies no universal optimizer or training recipe. Cascaded modules are described as separately optimized; joint architectures supervise world and action prediction together. FLARE, as summarized here, aligns internal future tokens to future-observation embeddings from a frozen teacher and can apply that alignment alone to action-free video. UWM instead handles missing actions by fully noising the action side. These are examples from the survey, not independently verified implementations. e-cascadede-implicite-noise

Source description

Data sources trade action grounding against scale and transfer difficulty. Portable UMI demonstrations add tracking and retargeting; simulation supplies privileged geometry; human video may supply visual dynamics without executable robot actions. Table 5 reports FastUMI-100K as 100K+ trajectories, RealOmin as 1M trajectories across 3000+ environments, and RDT2 as 10,000 hours. These are surveyed dataset descriptors with incompatible scale units, not performance results or a validated mixture recipe. e-datae-umie-simulatione-human

During inference

Source description

Future prediction need not mean rendering and executing an entire imagined video. The survey describes DreamZero replacing imagined frames with newly observed frames after executing an action chunk. Conversely, Fast-WAM removes future-video generation at inference and uses current-context features to condition action denoising. The deployed information path therefore needs a separate audit from the training objective. e-feedbacke-hidden

Reader analysis

Reader interpretation: joint supervision can improve learned representations without proving that predicted futures causally determine each deployed action. The survey's strict predictive definition and its inclusion of training-only future-generation designs create a boundary requiring method-specific treatment; neither a shared loss nor the label 'Joint WAM' establishes one homogeneous backbone. e-definitione-boundariese-jointe-hiddene-controlled

5.5 Implementation flow

  1. Trace prediction and control separately

    Section 3 distinguishes external uses of world models—demonstration synthesis, imagined reinforcement learning, reward modeling, and virtual evaluation—from the internal policy architectures of Section 4. A simulator takes an intervention and predicts its consequence; an action-generating agent must also determine the intervention. This distinction prevents a training aid from being mistaken for an inference-time planner. e-supporte-definition

  2. Follow a cascaded plan into an action decoder

    The world stage predicts a future from the observation and instruction. The second stage maps that future, together with current context, into commands. Explicit plans use images or geometric motion information; action extraction can use learned inverse dynamics or geometric computation. Implicit plans pass predicted latent features to a policy, avoiding full pixel reconstruction. UniPi illustrates frame-pair inverse dynamics; VPP illustrates prediction of future latent sequences for policy conditioning. e-cascadede-latent

  3. Locate the coupling inside joint prediction

    Autoregressive systems serialize world and action variables, with separate output heads, unified discrete tokens, or predictive latent representations. Diffusion and flow-based systems refine world and action variables through a shared trunk or coordinated branches. Figure 6 distinguishes cross-attention, one-way hidden-state conditioning, and shared encoding before separate decoding. These are different computational interfaces, even when all receive joint supervision. e-autoregressivee-jointe-hiddene-shared

  4. Evaluate the claimed capability at the appropriate level

    Visual reconstruction and perceptual metrics assess appearance; physical tests examine object continuity, interactions, and motion. Action plausibility asks whether an inverse dynamics model can recover executable controls. Policy benchmarks separately cover general, bimanual/humanoid, mobile, contact/deformable, and real-robot manipulation. Success on one axis cannot establish the others; the survey explicitly calls for evaluation of their causal alignment. e-evaluatione-actione-policye-coupled-evaluation

6. Experiments & results

This survey defines World Action Models through future-state prediction coupled to action generation, then organizes the literature by architecture, data, and evaluation. Its useful contribution is a vocabulary for tracing how imagined futures influence control. It does not establish a winning architecture: the authors identify missing matched comparisons and missing tests of whether executed actions actually follow predicted futures.

Source and visual limitations
Reader analysis

This is a conceptual and literature survey, not a new empirical method paper. Its original tables summarize architectures, dataset scales, and evaluation resources, rather than report common-protocol control scores. It contains no original controlled ablation or diagnostic experiment to crop; Section 7 explicitly identifies the absence of matched architecture comparisons. Table 5 supplies quantitative resource descriptors, and Table 8 supplies an evaluation map, with neither presented as a performance result. No featured result is selected because the base report records no new empirical results. e-scopee-controllede-umie-evaluation

6.1 Read the original evidence

Table 5. Portable-demonstration scale is reported in heterogeneous units, with missing fields preserved. Original paper, p. 30 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read across a row before comparing its size with another dataset. Scale reports trajectories for FastUMI-Data, FastUMI-100K, RealOmin, and Hoi!, but hours for RDT2. The Task and Env columns describe coverage rather than held-out evaluation success. The final column describes observation modalities, with the retained caption defining P as proprioception, D as depth, and T as tactile. Dashes remain unspecified entries; they do not mean zero tasks or zero robots. Section 5.2 explains the collection idea behind this table: portable human demonstrations use tracking and retargeting to connect diverse physical interactions to robot-executable actions. e-umie-data

What it supports. As reported by this survey, FastUMI-100K contains 100K+ trajectories, RealOmin contains 1M trajectories across 3000+ environments, and RDT2 contains 10,000 hours. These descriptors illustrate the scale of portable demonstrations, but they do not provide a common denominator for ranking data efficiency.

Where the evidence stops. This is a survey's dataset summary, not a newly measured control-results table. It provides no common train/test split, collection-cost comparison, or matched policy score. RDT2's hours cannot be converted to trajectories using information supplied here.

Table 8. Appearance, physical behavior, and recoverable control require different evidence. Original paper, p. 36 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the gray category bands down the table. Visual Fidelity includes pixel, perceptual, semantic, and distributional comparisons; these assess different aspects of generated appearance. Object Dynamics moves to physical interactions and violations. Motion and Trajectory Plausibility focuses on how motion evolves, including end-effector trajectories in EWMBench. The final band asks whether generated transitions preserve enough information to recover actions. Read the rightmost column carefully: WorldSimBench uses an implicit manipulative evaluation, while Wow, wo, val! uses an IDM Turing Test with downstream real-world execution. The accompanying discussion on pages 34–37 keeps these world-model tests distinct from direct policy benchmarking. e-evaluatione-actione-coupled-evaluation

What it supports. A convincing video is only one part of the evidence needed for control. The table makes action plausibility a separate requirement, and Section 7 goes further by calling for tests of the causal alignment between imagined and executed futures. It is an evaluation map, not a comparative leaderboard.

Where the evidence stops. The survey reports no common experiment applying every listed metric to the same models and tasks. Reader caution: an IDM execution test also depends on the decoder, so a fixed decoder and a real-video control are needed to interpret a generator's failure.

6.2 Results and evaluation conditions

No quantitative results are included in this reading.

6.3 Ablations and diagnostic examples

No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

The authors state that cascaded, joint, discrete, and latent designs lack systematic comparisons under matched scale, data, and evaluation. Accordingly, this survey offers no controlled architecture ranking, new benchmark score table, or original ablation result. Its cited success and efficiency claims are not a common-protocol meta-analysis. e-controllede-scope

Author claim

Long rollouts can accumulate prediction drift and action errors, while prediction latency constrains feedback. Contact forces, tactile state, and material compliance remain poorly represented by RGB futures. The optimal mixture of human video and robot demonstrations is unresolved, including whether transfer benefits are semantic or dynamical. e-limitationse-latency

Source description

The proposed Counterfactual Consistency and Foresight-Conditioned Success concepts are research directions, without implemented scoring protocols or validation in this survey. Uncertainty-aware checks on imagined futures are likewise proposed deployment safeguards, not demonstrated safety guarantees. e-coupled-evaluatione-safety

Reader analysis

Figure 7 is qualitative: its arrow directions lead toward easier transfer and cheaper scaling despite axis titles phrased as 'difficulty.' Table 5 mixes trajectory counts and hours and leaves several embodiment/task cells unspecified. Neither visual supports numerical tradeoff estimation or filling missing metadata. e-datae-umi

7.2 Questions for discussion

  1. What intervention would distinguish a useful future representation from auxiliary regularization?
  2. How should a WAM definition handle models that remove future generation at deployment?
  3. Which physical variables must be predicted for contact-rich control when RGB appearance is ambiguous?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproducing the survey's synthesis requires version-specific records of predictive targets, world-to-action interfaces, frozen modules, and training versus inference paths. The supplied v1 is the evidence anchor. Architectural sketches and dataset summaries do not provide a single runnable model, a shared training split, complete hyperparameters, or a survey-level compute budget. e-identitye-scopee-jointe-umi

Reader analysis

Reader-proposed checks: compare action-only, auxiliary-future, and actively future-conditioned policies under matched data and compute; separately compare visual-quality rankings against fixed-decoder execution and plan adherence. Both require explicit held-out tasks, camera/action conventions, execution horizons, feedback schedules, and uncertainty reporting. The illustrated edition specifies controls and falsifiable outcomes; neither check has been run. e-controllede-actione-coupled-evaluatione-policy

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Disentangle auxiliary supervision from active future conditioning

Reader-proposed experiment, not run: use a fixed LIBERO split and identical pretrained initialization, observation history, action horizon, parameter budget, and training budget. Compare an action-only policy, a policy with future-representation supervision but blocked future-to-action access, and a policy that actively conditions actions on predicted future representations. Train each configuration with its intended mask to avoid relying solely on an out-of-distribution test-time ablation. Across multiple seeds, report closed-loop success, prediction quality, actual inference latency, and uncertainty. If the auxiliary-only model matches the actively conditioned model, gains would not establish a need for online foresight; a reproducible conditioned-model advantage would support that narrower mechanism. e-policye-implicite-hiddene-controllede-latency

Check 2: Check whether visual rankings predict executable plan adherence

Reader-proposed experiment, not run: from identical held-out initial states and instructions, collect generated plans from several predictors and matched recorded plans as a positive control. Keep camera conventions, horizon, action frequency, and a single real-trajectory-trained IDM fixed. Compare perceptual rankings with execution success and deviation between planned and realized object/end-effector trajectories. For a counterfactual subset, use two feasible alternative goal placements allowed by the instruction and test whether decoded actions and realized endpoints follow the selected plan. If visually strong plans fail while recorded controls execute, appearance is an inadequate surrogate; if both fail, decoder validity remains unresolved. e-actione-evaluatione-coupled-evaluatione-policy

8.3 Reading coverage

Visual audit: Visually inspected the title/author/version page; Figures 3–7; Tables 5, 8, and 9; and all declared pages supporting retained definitions, architectures, training/inference examples, data, evaluation, limitations, and proposed checks. All six final original crops were inspected. Figure 7 was cropped from a 400-DPI render and retains its unusual endpoint directions; the other crops use 200 DPI. Diagram arrows were checked against captions and adjacent text. Figure 7's axis-title ambiguity is disclosed in its guide. Figure 6 omits deployment pruning rather than specifying a universal inference graph. All 25 text chunks, including references, were read; Figures 1–2 and Tables 1–4, 6–7 were not visually inspected. No appendix is present; separate supplements remain unverified.

PDF pages inspected for this edition: 1, 3, 5, 6, 12, 13, 14, 15, 17, 18, 20, 21, 22, 24, 26, 29, 30, 32, 34, 35, 36, 37, 38, 40, 41, 42, 43, 44. Appendix coverage: not present.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Title page and Abstract (p. 1)
  • 1 Introduction (pp. 1–4)
  • 2 Definitions and Formalism, including 2.1–2.2 (pp. 3–6)
  • 3 VLAs and World Models: Foundations and Early Integration, including 3.1–3.3.2 (pp. 6–13)
  • 4 Architecture: Cascaded and Joint WAMs, including all subsections (pp. 13–26)
  • 5 Training data: robot teleoperation, portable demonstrations, simulation, human and egocentric data (pp. 26–34)
  • 6 Evaluation: world modeling and action policy, including all subsections (pp. 34–41)
  • 7 Open challenges and Opportunities (pp. 41–44)
  • 8 Conclusions (p. 44)
  • References [1]–[366] (pp. 45–69); all 25 supplied text chunks read individually

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • The extraction's image omission was addressed by inspecting original PDF pages and final crops. Figures 3–7 and Tables 5, 8, and 9 were visually reviewed; Figures 1–2 and Tables 1–4, 6–7 were read through their supplied text and captions only.
  • Identity/version: the inspected title page states arXiv:2605.12090v1 [cs.RO], 12 May 2026. The exact title and all 14 authors agree with the catalog. No other revision or edition was supplied, so cross-version differences could not be assessed.
  • The bibliography was read as part of the survey; the cited papers, linked homepage, repository, and datasets were not independently inspected. No code was inspected and no experiments were reproduced.
  • No appendix is present in the supplied PDF; the conclusion is followed by references.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title/author block, arXiv margin, and AbstractInspect

Exact observed title; authors Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, Mike Zheng Shou, Xuanjing Huang, Xipeng Qiu, and Yu-Gang Jiang. Affiliations: Fudan University; Shanghai Innovation Institute; National University of Singapore. Margin identifies arXiv:2605.12090v1 [cs.RO], 12 May 2026; the abstract identifies the work as a survey.

Go to primary source ↓
e-scopePDF p. 3, Section 1 roadmap; p. 44, Section 8 ConclusionsInspect

The paper organizes definitions, background, architectures, training datasets, evaluation, and open challenges as a synthesis of the WAM literature.

Go to primary source ↓
e-definitionPDF p. 3, introductory factorization and Section 2 notation; p. 5, Figure 3 and Section 2.1 unnumbered objectivesInspect

Defines observation o, language l, action a, subsequent observation o′; gives cascaded factorization and VLA, WM, WAM probability targets. WAM criteria are forward predictive modeling and coupled action generation.

Go to primary source ↓
e-boundariesPDF p. 6, Section 2.2, items 1–3Inspect

WAM is modality-independent; VAM is video-oriented; a video backbone without active predictive supervision does not suffice. AWM is described as functionally similar terminology.

Go to primary source ↓
e-supportPDF p. 12, Figure 4 and Section 3.3.1; p. 13, Sections 3.3.2 and 4 openingInspect

World models support imitation learning, reinforcement learning, reward modeling, and evaluation as external tools; Section 4 turns to integration within policy architectures.

Go to primary source ↓
e-cascadedPDF p. 14, Section 4 opening and Section 4.1.1; p. 15, Figure 5 and Geometric Action ExtractionInspect

Cascaded systems predict an intermediate future before decoding actions. Explicit routes use learned inverse dynamics or geometric extraction; Figure 5 also depicts a latent carrier. UniPi regresses actions between generated frame pairs.

Go to primary source ↓
e-latentPDF p. 17, Section 4.1.2, opening and VPP/Video Policy paragraphsInspect

Implicit planning replaces pixel-space output with latent future features. VPP predicts future latent sequences; the described Video Policy freezes its video U-Net before training a feature-conditioned action U-Net.

Go to primary source ↓
e-autoregressivePDF p. 18, Sections 4.2 and 4.2.1Inspect

Autoregressive joint modeling serializes heterogeneous variables; the survey distinguishes decoupled heads, unified discrete representations, and predictive latent representations.

Go to primary source ↓
e-jointPDF p. 20, Section 4.2.2; p. 21, Figure 6 and captionInspect

Diffusion-based joint modeling divides into unified-stream and multi-stream architectures. Figure 6 shows cross-attention, a video-to-action hidden-state arrow, and shared representation before separate decoding.

Go to primary source ↓
e-noisePDF p. 21, Section 4.2.2, UWM paragraphInspect

UWM independently controls world/action noise levels to select functional modes and handles action-free video by fully noising missing actions.

Go to primary source ↓
e-implicitPDF p. 22, Section 4.2.2, Implicit Future Prediction and FLARE paragraphsInspect

Future tokens/prefixes can receive latent alignment supervision. FLARE aligns projected future-token activations to embeddings of future observations from a frozen teacher; video-only data uses alignment without the action objective.

Go to primary source ↓
e-feedbackPDF p. 22, Section 4.2.2, DreamZero paragraphInspect

The survey describes joint video/action denoising and replacement of imagined frames with ground-truth observations after executed chunks via the KV cache.

Go to primary source ↓
e-hiddenPDF p. 24, Section 4.2.2, Hidden-State Coupling and Fast-WAM paragraphsInspect

Hidden-state coupling transfers video-branch representations to the action branch. Fast-WAM uses a mask blocking attention to future video latents during training and removes future-video generation at inference, retaining current-context feature conditioning.

Go to primary source ↓
e-sharedPDF p. 26, Section 4.2.2 continuation, UVA paragraphInspect

UVA uses a shared representation with two jointly trained diffusion heads; action decoding can bypass video generation at inference.

Go to primary source ↓
e-dataPDF p. 26, Section 5 opening, Figure 7 and captionInspect

Contrasts aligned observation/action/future triplets with action-free observation sequences. Figure 7 positions robot teleoperation, portable demonstrations, simulation, and human/ego data qualitatively. Endpoint labels indicate easier matching upward and cheaper scaling rightward; no numerical calibration is supplied.

Go to primary source ↓
e-umiPDF p. 29, Section 5.2; p. 30, Table 5, all rows and captionInspect

Portable demonstrations use tracking and retargeting. Table 5 lists FastUMI-Data: 10K+ trajectories, 22 tasks; FastUMI-100K: 100K+ trajectories, 54 tasks, 5 environments; RealOmin: 1M trajectories, 30 tasks, 3000+ environments; Hoi!: 3,048 trajectories, 38 environments; RDT2: 10,000 hours, 52+ tasks, 100+ environments. Dashes and absent embodiment entries remain unspecified. The caption defines P, D, and T modalities.

Go to primary source ↓
e-simulationPDF p. 30, Section 5.3 openingInspect

The survey motivates simulation by controllability, scalability, and privileged geometric information such as depth, object poses, collision boundaries, and multi-view state.

Go to primary source ↓
e-humanPDF p. 32, Sections 5.4.1–5.4.2Inspect

Distinguishes learning passive visual dynamics from human video and extracting action-conditioned information through pose and motion tracking.

Go to primary source ↓
e-evaluationPDF p. 34, Sections 6–6.1.1; p. 35, Section 6.1.1; p. 36, Table 8Inspect

The survey states that established evaluations separate world modeling and action policy. Table 8 maps visual fidelity, object dynamics, motion/trajectory plausibility, and action plausibility to metrics and benchmarks. FVD is distributional; DINO uses feature cosine similarity; physics and execution tests assess different properties.

Go to primary source ↓
e-actionPDF p. 37, Sections 6.1.2 continuation and 6.1.3Inspect

EWMBench uses end-effector trajectory metrics HSD, nDTW, and DYN. WorldSimBench evaluates control-relevant information; Wow, wo, val! applies an IDM to generated video and evaluates inferred actions through real-world execution.

Go to primary source ↓
e-policyPDF p. 38, Section 6.2 introduction; p. 40, Table 9Inspect

Policy benchmarks are grouped into general manipulation, bimanual/humanoid, mobile, contact/deformation, and real-device evaluation. Table 9 lists LIBERO with a Franka Panda and MuJoCo/robosuite; benchmark descriptors are not policy success measurements.

Go to primary source ↓
e-controlledPDF p. 41, Section 7, Architectural CouplingInspect

Authors state that systematic matched comparisons of architectural paradigms are absent and question whether useful world modeling comes from training-time gradients or inference-time future generation.

Go to primary source ↓
e-limitationsPDF p. 42, Section 7, Multimodal Physical State Representation, Data Utilization and Mixture Design, and Long-Horizon PlanningInspect

Open issues include invisible contact variables, poorly understood data mixtures, prediction drift, compounded action errors, and temporal abstraction.

Go to primary source ↓
e-latencyPDF p. 43, Section 7, Inference Latency and Computational EfficiencyInspect

World prediction adds latency; authors ask how much predictive fidelity downstream control actually needs and propose task-adaptive prediction.

Go to primary source ↓
e-coupled-evaluationPDF p. 43, Section 7, Evaluation MethodologyInspect

Calls for joint tests of the causal connection between imagined futures and actions; Counterfactual Consistency and Foresight-Conditioned Success are proposed directions.

Go to primary source ↓
e-safetyPDF pp. 43–44, Section 7, Safety and Reliable Physical DeploymentInspect

Discusses extended erroneous action sequences and proposes uncertainty/physical-constraint checks on predicted futures; tractable, robust deployment verification remains unresolved.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.