Awesome-WAMs
GitHub
PAPER REPORTENAll readings ↗

World Action Models: A Survey

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

Authors: Qiuhong Shen, Shihua Zhang, Yue Liao, Qi Li, Zhenxiong Tan, Shizun Wang, Shuicheng Yan, Xinchao Wang

Affiliations: National University of Singapore

Source: 2606.20781 ↗ · Project page ↗ · Catalog record

Reading: 155 / 558 · 5 original figures & tables · ~20 min ·

1. Paper overview

In one sentence: The survey organizes WAMs by how predicted futures influence control, arguing that selective imagination can preserve useful dynamics while reducing the cost of acting. E02E03E04E06E15E18E20

At a glanceWhat to know
Research problem
Source description

The WAM label spans video planners, latent predictors and language-based policies with substantially different interfaces. A direct VLA need not predict consequences, while a world model need not select actions. The survey asks whether an explicit predictive pathway produces, scores, verifies or meaningfully trains the action pathway, then asks what that connection costs inside a control loop. E02E03

Core mechanism
Source description

A boundary definition admits cascades and coupled future/action models while excluding passive video generators and auxiliary future heads with no retained action use. It includes offline generated-future supervision, exemplified by DreamGen, so online imagination is not required in every included system. E03E05

Reading caution
Reader analysis

The census is a design map, not a leaderboard. Section 6 discusses disparate benchmarks and metrics but supplies no harmonized task/split/baseline table or original ablation. Its selective-imagination thesis is a synthesis of cited evidence, not a pooled causal estimate. E15E18E20

Core contributions

  • Source description

    A boundary definition admits cascades and coupled future/action models while excluding passive video generators and auxiliary future heads with no retained action use. It includes offline generated-future supervision, exemplified by DreamGen, so online imagination is not required in every included system. E03E05

  • Source description

    A three-family philosophy taxonomy complements a four-coordinate anatomy. The former concerns the generation requirement; the latter records predictive substrate, action coupling, backbone and deployment. Tables 1–2 instantiate the anatomy as a literature census. E04E06E14E15

  • Author claim

    The authors connect interactability, causality, persistence, physical plausibility and generalization to data and evaluation, arguing that the useful direction is less unnecessary future synthesis with capability reported alongside runtime cost. E16E18E20

Figure 1. The defining connection is the use of prediction in the action pathway. Original paper, p. 5 ↗

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

How to read it. Read the upper row as three different modeling obligations. The source defines o as an observation, l as language or goal context, a as action, and o′ as a future observation; o′ is deliberately broad and need not mean rendered RGB. The lower-left panel predicts a future and passes it into q, the action module. The middle panel conditions a predicted consequence on a candidate action; the prose adds the essential selection step, where that consequence determines execution. The lower-right panel predicts future and action together. These are alternative ways of retaining predictive information, rather than successive stages that every WAM must implement. E03E06E10

What it supports. The diagram makes a modular video planner plus inverse dynamics eligible for the WAM family, even when the modules are trained separately. It also requires more than a world predictor that simply renders consequences: the forecast must participate in producing, evaluating or training the action pathway.

Where the evidence stops. The word train in the boundary line needs §2.3's qualification: an auxiliary future loss alone is insufficient, although a retained future-trained action representation can qualify. A coupled distribution also does not establish one monolithic network.

2. Motivation

2.1 The problem and the proposed response

Source description

The WAM label spans video planners, latent predictors and language-based policies with substantially different interfaces. A direct VLA need not predict consequences, while a world model need not select actions. The survey asks whether an explicit predictive pathway produces, scores, verifies or meaningfully trains the action pathway, then asks what that connection costs inside a control loop. E02E03

2.2 What this reading follows

A robot can generate a convincing video and still fail to execute the imagined motion. This survey makes the connection between prediction and action the organizing question. Begin with its boundary diagram, then distinguish the requirement to generate pixels from the representation actually supplied to control. The four-axis anatomy adds the predictor family, action coupling and feedback schedule. Its census shows how these choices recombine across existing systems. The practical lesson is an evaluation question: how much future computation improves executed actions within a specified budget? The source provides a conceptual synthesis and research agenda, with classification tables rather than a new policy benchmark. E02E03E04E06E15E18E20

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 work / Surveys & technical resources classification is supported. Architecture, prediction paradigm and quadrant are correctly Not applicable for the survey itself. It analyzes both composed and shared predictive-action systems; neither its joint-distribution notation nor a surveyed joint loss establishes a One Model architecture for this resource. E02E06E14

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 abstraction: observation history, past action history, and language, goal-image or other task context (E06).
  • For the survey itself: the cited WAM literature organized in Tables 1–2 (E14, E15).
  • For the abstraction: an action-facing future substrate and associated action window, realized jointly or through composed modules (E06).
  • For the survey: classification, design tradeoffs and evaluation questions; no new executable policy (E02, E20).

4.2 Equations and their role

pΘ(st+1:t+H,at:t+H−1∣o≤t,a<t,l)p_{\Theta}(s_{t+1:t+H},a_{t:t+H-1}\mid o_{\leq t},a_{<t},l)
Equation (8): o is observation history, a denotes actions, l is task context, H is the prediction horizon, and s is the future in the chosen substrate space. The parameter collection Θ can comprise a shared model or a predictor plus actor; this abstract joint is not architecture evidence for one network. E06
pθ(st+1:t+H∣c)qψ(at:t+H−1∣st+1:t+H,c)p_{\theta}(s_{t+1:t+H}\mid c)q_{\psi}(a_{t:t+H-1}\mid s_{t+1:t+H},c)
Equation (12): c abbreviates observation history, past actions and task context. The predictor pθ produces the future first, and the action module qψ then uses it. This is the post-prediction factorization, not a guarantee of physical executability. E06E10
Ljoint(θ)=Lgen(s)+λLact(a)\mathcal{L}_{\mathrm{joint}}(\theta)=\mathcal{L}_{\mathrm{gen}}(s)+\lambda\mathcal{L}_{\mathrm{act}}(a)
Equation (19): the generative loss supervises the substrate and the action loss supervises controls, weighted by λ. The paper presents a common schematic objective, not one numerical setting shared by all WAMs. E09
Cchunk(T,K)=⌈T/K⌉Nfwd(K)C_{\mathrm{chunk}}(T,K)=\lceil T/K\rceil N_{\mathrm{fwd}}(K)
Equation (34): T is task length in control steps, K the chunk/replanning period, and Nfwd(K) the cost of producing one K-step forecast. The accompanying timing inequality compares Nfwd(K)/K with 1/fctrl, where fctrl is control frequency. Interpreting that comparison requires forward cost expressed as elapsed time on specified hardware. E13

5. Method in detail

5.1 Trace what turns a future into a control decision

Reader analysis

Start with the survey's UniPi description. Observations and task context condition a video predictor; a separate inverse-dynamics module recovers controls from the produced trajectory. The future is therefore upstream of the action. Now reverse that order for PointWorld: an external search supplies candidate controls, the point-flow predictor estimates consequences, and their scores determine selection. Finally, joint generation produces future and action through a coupled model. Reader interpretation: these pathways should be audited by tracing actual dependencies, not by looking for a video loss in a training diagram. A generated image can be action-relevant, but its visual plausibility does not establish executability. The survey also includes offline coupling, such as DreamGen's generated-future supervision; that inclusion should not be mistaken for an online planning call in the deployed policy. E03E05E08E09E10E18

Figure 2. Locate how far prediction must proceed before it becomes useful to action. Original paper, p. 7 ↗

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

How to read it. Follow each column downward from observation and instruction. The left column carries prediction to a pixel future; the center retains a video-derived predictor but routes action through a latent, feature, flow or mask; the right replaces the video-generation path with a compact predictive representation. The differences concern the generation requirement and lineage, not simply whether training data contain videos. Section 3 also allows the relevant connection to occur during action supervision. Its DreamGen example therefore belongs on the rendered side even though generated futures supply offline pseudo-actions for a downstream policy. Read the separate action boxes as a conceptual interface: joint future/action models also occur within these philosophies. E04E05E07E09

What it supports. The center column explains why dropping pixel decoding does not mean dropping the video prior. The right column explains a separate move: prediction can occur in features, geometry or task maps without a video-generation core. These are alternative representation and computation choices, not a measured ordering of robot success.

Where the evidence stops. Section 4 uses a different axis: a decoder-bound video latent still counts as pixel-grounded. Consequently, Latent-Only does not imply feature-substrate, and the drawing should not be read as proving separate trained modules or online generation in every example.

5.2 Separate a latent shortcut from a feature future

Reader analysis

Suppose a controller never renders an RGB forecast. That fact answers only part of the classification problem. The survey describes UWM as retaining a pixel-decodable latent future while allowing the visual branch to collapse at inference; its census places UWM in Latent-Only with a pixel-latent substrate. Fast-WAM instead uses an encoder-only hidden state shaped by future co-training, so the source calls its substrate a feature. FLARE provides another feature route: the action expert consumes predicted future embeddings aligned to a frozen teacher, without a video-generation core. Reader interpretation: record the fixed decoder, the retained action input and the training supervision separately. This prevents two mistakes: calling all latents equivalent, and interpreting an auxiliary objective as evidence that explicit future generation remains on the inference path. E03E05E07E11E15

Figure 4. Four questions describe a WAM more precisely than its backbone name. Original paper, p. 13 ↗

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

How to read it. Read the upper boxes as design coordinates, not a left-to-right execution graph. Substrate asks what future is represented; backbone asks how it is predicted; coupling asks how actions enter or leave that prediction; deployment asks when the model is invoked. Equation (37) orders the tuple as substrate, coupling, backbone, deployment, although this graphic places backbone before coupling. The abbreviated pre, mid, joint, head labels in the coupling box are not the formal categories: Equations (10)–(12) define action-conditioned rollout, joint generation and a post-prediction head. Use those factorizations to reconstruct the information flow. The lower loop denotes measured feedback becoming available after action. E06E07E13E14E23

What it supports. This anatomy separates changes that are easily conflated. Bypassing a decoder changes the rendering obligation without necessarily changing substrate; changing the replanning schedule changes deployment while the trained predictor can remain fixed. The useful comparison is therefore between specified implementations and control schedules, not model-family names alone.

Where the evidence stops. The return arrow visually points into the instruction/task block despite its new-observations label. Equations (8)–(9) instead place feedback in observation history, separately from task context. Preserve this schematic ambiguity; it does not specify an algorithm that rewrites instructions after every action.

5.3 Evaluate imagination at the controller's deadline

Reader analysis

The deployment equations explain why a representation cannot be judged independently of its schedule. In the paper's chunked abstraction, the controller predicts K actions and refreshes after that window, amortizing one forecast over several control ticks. A larger window can reduce invocation overhead while leaving the action sequence stale when contact or the scene changes. Observation replacement helps re-ground subsequent predictions, but it does not remove every memory or executability problem. Reader interpretation: compare success under the same control deadline and declared task shift before interpreting cheaper imagination as an improvement. Separate visual prediction scores from physical or simulated execution, and report latency, sustained horizon and peak memory together. The survey recommends this comparison; its own census does not supply the controlled experiments needed to decide which representation wins. E13E16E18E20

Figure 6. Distinguish evaluating proposed controls from producing controls. Original paper, p. 18 ↗

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

How to read it. The left panel begins with an action source as well as observations. Its candidates condition the predictor, and an outer procedure uses their predicted consequences to choose controls. The dashed return curve is a schematic search/refinement connection; Equation (17) supplies the actual chunk-level and step-wise factorizations. The prose says the current census uses the chunk-level submode, so the curve alone is not evidence of reaction inside an imagined window. The middle panel generates future and action in a coupled process. The right panel predicts first, then lets a smaller actor consume that future; UniPi's inverse-dynamics module is one instance of this arrangement. E08E09E10E21

What it supports. The three families differ in where action becomes bound to prediction. Candidate rollout enables counterfactual evaluation, joint generation couples outputs, and post-prediction decoding supports a modular actor. Those distinctions locate where repeated prediction, shared objectives or downstream translation may become expensive or fail; the drawing does not measure those costs.

Where the evidence stops. The middle label Joint Denoising is narrower than the verified family. WorldVLA appears there, but §4.3.2 and Table 1 call it autoregressive with joint-generation coupling. Read this panel as joint generation, without changing the original graphic or attributing diffusion to every listed method.

5.4 Training and inference

During training

Source description

These are surveyed recipes, not training runs performed for this paper. Joint methods commonly combine substrate and action losses, potentially after video pretraining; the losses can compete. Post-prediction methods often freeze the predictor while training an embodiment-specific actor. The survey supplies no universal optimizer, mixture ratio, loss weight or hardware recipe. E09E10E19

Source description

Joint-embedding prediction matches context-derived features to future target embeddings with stop-gradient on the target. The generic exposition uses a moving-average target encoder; the FLARE example instead uses a frozen teacher. Internet video, teleoperation, portable human demonstrations, simulation and synthetic neural trajectories offer different action-label quality and embodiment alignment. E11E17

During inference

Source description

The survey describes UniPi as video prediction followed by inverse dynamics, PointWorld as external candidate search over action-conditioned point-flow forecasts, and FLARE as future teacher-feature prediction consumed by an action expert. These are descriptions from the survey; their original implementations were not independently verified here. E08E10E11

Source description

Prediction need not be fully rendered at test time: UWM can collapse its visual branch, while Fast-WAM acts through a future-trained hidden state. Feedback and bounded memory remain separate requirements. Observation replacement can re-ground predictions, whereas long chunks risk stale actions and full-history storage raises memory cost. E05E07E13E16

5.5 Implementation flow

  1. Locate the future-to-action connection

    Trace whether a predictor supplies a future to an action decoder, evaluates proposed actions, or predicts future and action together. An inverse-dynamics module is one possible decoder, not the definition of WAM. The joint notation also permits separate predictor and actor parameters. E03E06

  2. Identify the generation obligation

    Render-and-Decode requires rendered futures in its prediction/action-supervision route. Latent-Only retains a video-derived predictive path while bypassing pixel decoding during control. Video-Generation-Free removes the video-generation core; it can still use diffusion, language generation or video-derived training data. E04E05E15

  3. Identify the representation separately

    A VAE or VQ grid with a fixed observation decoder remains pixel-grounded even if that decoder is bypassed. Feature futures have no fixed observation decoder; geometric futures encode coordinates or motion; affordance futures encode task maps or scores. Thus Latent-Only and feature-substrate are not synonyms. E07

  4. Separate coupling from the predictor family

    Action-conditioned rollout supplies candidate actions before consequence prediction. Joint generation couples future and action; a post-prediction head decodes actions after forecasting. Diffusion, autoregression, joint-embedding prediction, hybrid and LLM/VLM backbones parameterize these interfaces in different ways. Autoregression alone does not imply joint action prediction. E08E09E10E11E12

  5. Specify when feedback reaches the predictor

    Open-loop use fixes the forecast before execution; chunked control refreshes it between action windows; single-step control invokes it each tick; interactive operation retains state across continuing inputs. These describe predictor invocation, not proof that the robot can execute a predicted trajectory. E13E14E18

6. Experiments & results

This survey organizes embodied models by how future prediction becomes useful to action. Its two complementary views separate the obligation to render a future from the representation, predictor, action interface and control schedule implementing it. The authors argue for selective imagination: retain information that constrains executable control while accounting for latency, memory and action-label cost. This is a conceptual synthesis of cited systems, not a newly trained policy or a controlled performance comparison (E02, E03, E20).

Source and visual limitations
Reader analysis

This source is a survey with six conceptual or taxonomy figures and two classification census tables. It contains no original quantitative benchmark-results table, experimental ablation table or ablation plot. Section 6 discusses evaluation practices and Section 7 reports lessons from cited work without a harmonized experimental protocol. Accordingly, this edition uses four mechanism diagrams and an original census-table excerpt; its results-section visual is qualitative taxonomy evidence. No numerical result or diagnostic experiment has been invented to fill the missing experiment types. E02E15E18E20

6.1 Read the original evidence

Table 2 (continuation; DUST–Feedback-WM rows). A generation-free philosophy still permits several backbones and action interfaces. Original paper, p. 32 ↗

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

How to read it. Read across one row before comparing columns. DUST is listed with a VLM-token feature substrate, hybrid backbone, joint generation and chunked deployment. PALM instead uses affordances, an LLM/VLM backbone and a post-prediction head. Audio-WM supplies the useful counterexample to equating generation-free with non-diffusion: its backbone column says Diffusion while its substrate is audio latent. The dates are first-submission dates, not training budgets or performance values. This crop preserves the continuation's complete header and rows; FLARE begins the table on the preceding page and is outside this excerpt. All entries here are the survey's classifications rather than independent reviews of those algorithms. E13E14E15

What it supports. The census illustrates recombination: the absence of a video-generation core does not fix substrate or coupling. DUST and PALM share the generation-free family while differing in both. This is qualitative evidence for the usefulness of multiple axes, not an experimental result showing one design outperforms another.

Where the evidence stops. PointWorld's Open / Chunked cell must be read with §4.5.1: the survey says its reported experiments do not replan during execution. An intended or possible regime should not be promoted to demonstrated closed-loop performance. No success rates, shared splits or uncertainty estimates appear here.

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 census is a design map, not a leaderboard. Section 6 discusses disparate benchmarks and metrics but supplies no harmonized task/split/baseline table or original ablation. Its selective-imagination thesis is a synthesis of cited evidence, not a pooled causal estimate. E15E18E20

Reader analysis

Some source labels conflict. Figure 6 names its middle family Joint Denoising while including autoregressive WorldVLA. Figure 5 labels JOPAT Feature ∧ Geometric, whereas the substrate text and Table 1 specify a pixel-decodable visual future plus geometry. These discrepancies should remain visible rather than become authoritative algorithm labels. E21E22

Reader analysis

The membership boundary needs care: §2.3 rejects an auxiliary future loss alone yet admits a future-trained retained action representation. The distinction requires inspecting information flow and supervision, not merely counting losses or assuming online rendered imagination. E03E05E07

Source description

The authors identify missing standardized physical and long-horizon evaluation, uncertain mapping from perceptual scores to execution, and unresolved data-stage scaling laws. Compact flow or latent actions still need embodiment grounding; reduced visual detail can discard contact or semantic information needed for control. E18E19E20

7.2 Questions for discussion

  1. How should the WAM boundary operationally distinguish retained future-trained action features from an unused auxiliary future head? (E03, E07)
  2. Which representation preserves contact information while allowing the controller to meet its deadline? (E13, E16, E20)
  3. What evidence is needed before an intended closed-loop design can be labeled a demonstrated closed-loop result? (E13, E15, E18)

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproducing this survey means re-establishing versioned classifications and evidence, not training a single model. Preserve whether each reported regime was actually evaluated: Table 2 gives PointWorld Open / Chunked, while §4.5.1 explicitly says its reported experiments do not replan during execution. E13E15

Reader analysis

Reader-proposed checks: audit the same methods across figures, equations and census rows; then test optional future generation while holding future co-training, task split, action cadence and hardware constant. Record success together with latency, horizon, peak memory and contact failures. These checks are proposals, not reported or completed experiments. E05E13E20E21E22

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: Reconstruct the census with a versioned consistency audit

Reader-proposed, not performed: independently annotate UniPi, UWM, Fast-WAM, FLARE, PointWorld, WorldVLA and JOPAT from the supplied survey's definitions, prose and table rows. Record generation requirement, fixed decoder, action input, coupling and evaluated deployment separately, then compare annotations to the original census. Use the explicit factorizations as the control against classifications inferred from diagram titles alone. The audit should surface the WorldVLA denoising label, the JOPAT substrate mismatch and PointWorld's evaluated open-loop qualification. An unresolved disagreement after separating these variables would falsify the claim that the published labels alone are an unambiguous placement procedure. Extending the audit to original implementations would require separately supplied, verified sources. E03E05E06E07E13E15E21E22

Check 2: Separate future co-training from test-time generation

Reader-proposed, not performed: use an implementation that explicitly supports optional future-video generation. Compare one future-co-trained checkpoint with generation enabled versus disabled, holding action weights, task split, observations, action chunk, hardware and deadline fixed. Add a matched training run without future co-training to distinguish representation learning from runtime imagination. Evaluate identical task instances, separate transit and contact failures, and report paired success differences with uncertainty, end-to-end latency, missed deadlines and peak memory. If disabling generation preserves success while reducing latency only for the co-trained model, that supports the survey's distinction. If contact success deteriorates despite equal deadlines, it identifies a boundary where retained features fail to replace explicit prediction. E05E07E13E18E20

8.3 Reading coverage

Visual audit: All body pages 1–42 were rendered and visually inspected, including the title/author page, Figures 1–6, complete Tables 1–2, formal method/training/deployment equations and evaluation/challenge sections. All five final crops were individually viewed. Figure 6's joint-denoising label and Figure 4's feedback-arrow/coupling-label ambiguities are disclosed beside their crops; Figure 5's JOPAT mismatch is documented in the base report. Reference-only pages 43–57 were read as text. There is no appendix in this PDF; separate supplements remain unverified. No referenced implementation was inspected or executed.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42. 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, abstract and contents (pp. 1–2)
  • 1 Introduction (pp. 2–4)
  • 2 The Emergence of World Action Models, §§2.1–2.3 (pp. 4–6)
  • 3 Three Design Philosophies, §§3.1–3.4 (pp. 6–12)
  • 4 What Makes a World Action Model, §§4.1–4.6, including all backbone objectives, coupling factorizations, deployment equations and census rows (pp. 12–32)
  • 5 Core Properties, §§5.1–5.5 (pp. 32–36)
  • 6 Data and Evaluation, §§6.1–6.2 (pp. 36–38)
  • 7 Open Challenges, §§7.1–7.7 (pp. 39–42)
  • 8 Conclusion (p. 42)
  • References [1]–[215] (pp. 42–57)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • The extraction limitation above was addressed by visually inspecting all body pages, 1–42, all six figures and both census tables; reference-only pages 43–57 were read as text.
  • Separate supplemental material availability has not been fully verified.
  • No separate appendix is present in the supplied PDF. Referenced papers, linked code, datasets and the survey homepage were not inspected; no experiments were reproduced.
  • Identity/version note: the supplied title page identifies arXiv:2606.20781v1, 18 June 2026. The main title and all eight authors match the catalog; the title page additionally displays the subtitle Dream Less, Act More. No other revision was supplied or compared.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

E01PDF p. 1, title/author/affiliation block and arXiv marginInspect

The main title is World Action Models: A Survey; the subtitle is Dream Less, Act More. Authors are Qiuhong Shen, Shihua Zhang, Yue Liao, Qi Li, Zhenxiong Tan, Shizun Wang, Shuicheng Yan and Xinchao Wang. Affiliation: National University of Singapore. The margin identifies arXiv:2606.20781v1, 18 Jun 2026.

Go to primary source ↓
E02PDF pp. 1–4, abstract, contents and §1Inspect

The paper introduces a survey organized by three generation philosophies and four implementation axes, followed by core properties, data, evaluation and challenges.

Go to primary source ↓
E03PDF pp. 4–6, §§2.1–2.3, Figure 1 and Eqs. (1)–(7)Inspect

The source distinguishes direct action prediction, world prediction and future/action coupling; it admits cascades, action scoring and joint prediction, discusses future-trained retained features, and excludes auxiliary future losses or discarded future heads without action use.

Go to primary source ↓
E04PDF pp. 6–7, §3 and Figure 2; p. 11, §3.3 openingInspect

Render-and-Decode, Latent-Only and Video-Generation-Free concern the generation requirement and video-predictor lineage, separately from cascaded versus joint coupling. Generation-free does not exclude other generative components.

Go to primary source ↓
E05PDF p. 8, §3.1 modular Render-and-Decode paragraph; pp. 10–11, §3.2Inspect

DreamGen uses generated futures for offline pseudo-action supervision. UWM can collapse visual inference; Fast-WAM and other latent methods retain video-trained representations without obligatory full video generation.

Go to primary source ↓
E06PDF pp. 12–14, §4.1, Figure 4 and Eqs. (8)–(12)Inspect

The horizon-level joint conditions on observation history, past actions and task context. Θ may comprise separate modules. The three action factorizations distinguish candidate-conditioned rollout, joint prediction and a post-prediction action head.

Go to primary source ↓
E07PDF pp. 14–17, §§4.2.1–4.2.5, Eqs. (13)–(16) and Figure 5Inspect

Fixed observation decoders distinguish pixel-grounded latent futures from features. Other substrates include geometry and task maps. Fast-WAM is described as using a hidden state with its future branch masked during inference.

Go to primary source ↓
E08PDF p. 18, Figure 6 and §4.3.1, Eq. (17); p. 19, §4.3.1 continuationInspect

Candidate-conditioned rollout combines an external action source with consequence prediction. PointWorld uses MPPI proposals and scene-flow scoring. The surveyed census uses chunk-level rollout; step-wise rollout is presented as background/generalization.

Go to primary source ↓
E09PDF pp. 19–20, §4.3.2, Eqs. (18)–(19)Inspect

Joint generation includes diffusion and autoregressive realizations. A weighted substrate/action loss is a common formulation; shared objectives can compete and video pretraining may precede action training.

Go to primary source ↓
E10PDF pp. 20–21, §4.3.3, Eq. (20), and §4.3.4Inspect

Post-prediction actors consume generated futures; predictor weights are frozen in many implementations. UniPi uses generated image trajectories and inverse dynamics. Controls can be continuous, discrete or learned latent actions.

Go to primary source ↓
E11PDF pp. 21–24, §§4.4.1–4.4.3, Eqs. (21)–(29)Inspect

Diffusion uses iterative denoising, autoregression predicts a causal stream, and JEPA predicts target features with stop-gradient. FLARE uses future embeddings from a frozen teacher and a downstream action expert. Backbone type does not by itself determine action coupling.

Go to primary source ↓
E12PDF pp. 25–26, §§4.4.4–4.4.6, Eqs. (30)–(32)Inspect

Hybrid trunks expose generative and action heads; LLM/VLM WAMs retain a future substrate in their action pathway. The five backbone families are not mutually exclusive.

Go to primary source ↓
E13PDF pp. 27–28, §§4.5.1–4.5.3, Eqs. (33)–(35)Inspect

Open-loop, chunked and single-step deployment allocate predictor calls differently. Chunk cost is ceil(T/K) times Nfwd(K), with a timing constraint and stale-chunk tradeoff. PointWorld is explicitly described as evaluated without replanning during execution.

Go to primary source ↓
E14PDF p. 29, §§4.5.4–4.6, Eqs. (36)–(37)Inspect

Interactive operation retains cached or latent state. The tuple orders substrate Φ, action coupling F, backbone B and deployment D; it describes WAM design without requiring a monolithic network.

Go to primary source ↓
E15PDF pp. 30–32, Tables 1–2, headers and all rowsInspect

Both tables are literature censuses with method, first-submission date, substrate, backbone, coupling and deployment columns. Table 2 starts with FLARE on p. 31; its p. 32 continuation includes DUST, Audio-WM, HiF-VLA, PointWorld, PALM and others. PointWorld is labeled Open / Chunked.

Go to primary source ↓
E16PDF pp. 32–36, §§5.1–5.5Inspect

The survey relates action interactability, leakage control, persistence, embodiment realizability and specified distribution shifts. Observation replacement addresses drift; memory and abstraction impose additional tradeoffs.

Go to primary source ↓
E17PDF pp. 36–37, §6.1, five data-source groupsInspect

Teleoperation, portable human capture, internet video, simulation and synthetic neural trajectories differ in scale, trusted action labels, embodiment gap and generator error.

Go to primary source ↓
E18PDF p. 38, §6.2, all four evaluation paragraphsInspect

Visual metrics are separated from closed-loop simulation, hardware and learned-model evaluation. Cascades require substrate and execution checks. Standardized physical and hour-scale control evaluation remain missing; the survey offers no common quantitative comparison table.

Go to primary source ↓
E19PDF pp. 39–41, §§7.2–7.6Inspect

Data must be assigned to components and stages; action-free learning still requires robot grounding. The survey identifies unresolved source/stage scaling, bounded-memory integration, targeted transfer and physical interpretation of abstract actions.

Go to primary source ↓
E20PDF p. 39, §7.1; pp. 41–42, §7.7 and §8Inspect

The authors advocate fidelity-latency tradeoffs and reporting success, latency, sustained horizon, peak memory and contact-sensitive failures together. The conclusion synthesizes selective imagination rather than presenting a new trained model.

Go to primary source ↓
E21PDF p. 18, Figure 6 middle panel; p. 19, §4.3.2 joint autoregression; p. 30, Table 1 WorldVLA rowInspect

Figure 6 places WorldVLA under Joint Denoising; the prose and census identify its autoregressive backbone and joint-generation coupling. Joint generation is the verifiable broader family.

Go to primary source ↓
E22PDF p. 17, Figure 5 JOPAT joint cell; p. 16, §4.2.3 structured physical futures; p. 31, Table 1 JOPAT rowInspect

Figure 5 labels JOPAT Feature ∧ Geometric. The text describes a pixel-decodable visual future with tracks/visibility, and the census labels Pixel (latent) ∧ Geometric.

Go to primary source ↓
E23PDF p. 13, Figure 4 return arrow and action-coupling block; pp. 12–14, Eqs. (8)–(12) and §4.1Inspect

The return arrow is labeled as new observations but visually points toward the instruction/task block. The context equation separates observations from task context. The panel lists pre, mid, joint, head, whereas the formal text defines three coupling factorizations.

Go to primary source ↓

8.5 Primary sources

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