World Action Models: A Survey

• Standard VLAs such as RT-2 [12], OpenVLA [78], and π0 [11] map observations and instructions to actions but do not explicitly model how interventions change the environment.The survey frames this limitation as constraining agents’ ability to reason about consequences before acting.Source: §1 · p. 2
• WAMs make a predicted future action-facing through rendered pixels, intermediate latents, or video-generation-free reasoning, with action decoded from or jointly trained with that future.Examples include UniPi [29], VPP [62], DUST [171], and Audio-WM.Source: §1 · p. 3
• WAM design choices trade predictive richness against compute, memory, and control-loop latency, and the field is moving toward generating less while retaining what control requires.Source: §1 · pp. 3–4
• The survey pairs a three-family taxonomy—Render-and-Decode, Latent-Only, and Video-Generation-Free—with a four-axis anatomy of substrate, backbone, action coupling, and deployment regime.It records each WAM as a 4-tuple and uses the two views to compare what a method generates and how it is built.Source: §1 · pp. 3–4
• The survey evaluates WAMs against five embodied requirements: interactability, causality, persistence, physical plausibility, and generalization.It also considers the compute, memory, data, and evaluation costs of methods addressing these requirements.Source: §1 · pp. 3–4
• The survey covers WAM training data from teleoperation, portable human demonstrations, simulation, and egocentric internet video, alongside emerging evaluation practices and stack-specific open challenges.Source: §1 · p. 4

2.1 Vision-Language-Action Models: Policies from Vision-Language Pretraining

• VLA policies bind vision, language, and action, but need not predict what follows an action; RT-2 [12], OpenVLA [78], and the π series [11] [72] exemplify this neighboring policy line.A direct VLA becomes a WAM only when a predicted future observation helps produce, choose, or check the action.Source: §2.1 · p. 4

2.2 World Models: Predictive Dynamics Beyond Video

• World models predict future observations under actions, which may be pixels, features, geometry, affordances, audio, symbols, or tokens; video generation is one instance, not the definition.Internet-scale video generators supplied transferable visual-dynamics priors, while other predictive substrates showed that useful futures need not be rendered video.Source: §2.2 · pp. 4–5

2.3 World Action Models: Making the Future Action-Facing

• A WAM uses a predicted future to produce, score, or train the action path, through cascades, action-scoring rollouts, or joint prediction—not merely an auxiliary or discarded future.Early cascades such as UniPi [29], VLP [30], and AVDC [81] used separate action modules; GR-1 and GR-2 [16] helped establish joint video-action prediction.Source: §2.3 · pp. 5–6, Figure 1 · p. 5

3 Three Design Philosophies of World Action Models

• WAMs divide into Render-and-Decode, Latent-Only, and Video-Generation-Free according to where action prediction is grounded, independently of whether prediction and action are cascaded or joint.The distinction concerns the inference path or action supervision: rendered pixels, an intermediate video-derived representation, or a path outside video generation.Source: §3 · p. 6, Figure 2 · p. 7
• The field’s progression moves from rendered-future control to latent interception and then video-generation-free prediction, while extending across varied domains and observation regimes.Recent applications include tabletop manipulation, dexterous hand control, tactile interaction, autonomous driving, and aerial manipulation.Source: §3 · pp. 6–7, Figure 3 · p. 8

3.1 Render-and-Decode World Action Models

• Render-and-Decode generates a rendered future for an action module, making appearance, motion, contact, and scene dynamics inspectable but placing visual synthesis in the control loop’s latency budget.Producing pixels incurs the full denoising or autoregression schedule, and video-quality metrics only weakly predict downstream task success.Source: §3.1 · pp. 7, 9

3.2 Latent-Only World Action Models

• Latent-Only retains video-derived dynamic structure but bypasses pixel decoding, reducing inference cost while sacrificing direct visual interpretability.Action can be decoded from latents, intermediate features, flow, masks, or value maps; several methods reach real-time inference.Source: §3.2 · pp. 10–11

3.3 Video-Generation-Free World Action Models

• Video-Generation-Free removes the video-generation backbone from prediction and instead uses compact representations to retain predictive supervision without pixel-level video generation.FLARE [204] predicts future-observation embeddings for action, while DDP [64] and Feedback-WM use latent imagination or transitions for out-of-distribution detection and online policy correction.Source: §3.3 · pp. 11–12

3.4 Where the three philosophies meet the formal anatomy

• The three philosophies locate action prediction in the inference path, while the four-axis anatomy separately describes substrate, backbone family, action coupling, and deployment regime.The taxonomy and component-level anatomy are complementary views for navigating the WAM census.Source: §3.4 · p. 12

• A paper-by-paper enumeration is insufficient because WAMs with the same label can differ in implementation, while related methods may use different terminology.The section instead seeks a common basis for comparing them.Source: §4 · p. 12
• The section treats WAMs as conditional joint distributions over future predictions and future actions, with design choices specifying how those futures are represented and used.Source: §4 · p. 12
• The section walks through the choices offered by four ingredients and notes the practical trade-offs associated with each.Those ingredients are predictive substrate, action coupling, architectural backbone, and deployment regime.Source: §4 · p. 12

4.1 A Unified Notation for World Action Models

• A WAM predicts a length-H future trajectory in substrate space together with a length-H future action chunk, conditioned on observation and action histories and a task instruction.Its parameters may belong to one shared backbone or to a retained world-model module paired with an action module.Source: §4.1 · pp. 12–13
• The substrate is the action-facing future representation: decoded frames or generator latents, hidden features, geometric primitives, or task-relevant affordance and score maps.The interface encoder defines the exposed representation, not necessarily the generator’s internal implementation; pixel decoding may be unnecessary at inference.Source: §4.1 · pp. 12–13
• Action coupling uses action-conditioned rollout, coupled joint generation, or substrate-first prediction followed by a smaller actor that may be trained separately.Rollout can fix an action chunk before prediction or update actions step by step; post-prediction actors include trackers, inverse-dynamics models, optimizers, planners, or policies.Source: §4.1 · p. 13
• The four-axis anatomy distinguishes substrate, backbone, action coupling, and deployment, with backbones including denoising, autoregressive, joint-embedding, hybrid, and language or vision-language models.Deployment may be open-loop, chunked or single-step closed-loop, or interactive with persistent state across calls.Source: §4.1 · p. 14
• Training usually fixes substrate and backbone, whereas deployment and some wrapper-level coupling choices can change at inference; changing coupling families often needs an added head, planner, or fine-tuning.The four questions organize the following subsections around what future is represented, how it is predicted, how actions enter or are recovered, and how the model is invoked in control.Source: §4.1 · p. 14

4.2.1 Pixel-grounded substrates

• Pixel-grounded futures include decoded observations and latents with fixed video decoders, preserving appearance and contact traces while potentially predicting detail control does not need.This category also includes sparse or multimodal observation futures and joint observation predictions; decoder-bound audio latents remain an edge case rather than a separate category.Source: §4.2.1 · pp. 14–15

4.2.2 Feature substrates

• Feature futures are learned states, teacher embeddings, or VLM tokens without a fixed observation decoder, offering compact semantic prediction but no direct visual-fidelity metric.Methods may use encoder-only states, intermediate denoising states, teacher targets, or future tokens; these are future carriers rather than merely discretized actions.Source: §4.2.2 · pp. 15–16

4.2.3 Geometric primitive substrates

• Geometric futures predict physical structure such as flow, tracks, depth, pose, motion vectors, or polylines, reducing representation size but losing appearance cues useful for semantic grounding.They are most reliable when motion, contact, or geometry carries task-relevant information, and less suitable when appearance or semantics is essential.Source: §4.2.3 · p. 16

4.2.4 Affordance map substrates

• Affordance futures are task-specific spatial maps or score fields, such as value, contact, mask, or progress maps, that can be compact and directly useful but require task-defined labels.They are distinct from observations, learned features, and physical-coordinate fields.Source: §4.2.4 · p. 16

4.2.5 Substrate and coupling are distinct axes

• Substrate and action coupling are separate axes: the substrate identifies where the future is formed, while coupling describes how that future reaches action.For example, Fast-WAM [193] is feature-substrate and encoder-only, whereas τ0-WM [209] is pixel-grounded and jointly denoises video latents with action velocity.Source: §4.2.5 · pp. 16–17

4.3.1 Action-conditioned rollout

• Action-conditioned rollout evaluates futures conditioned on externally supplied actions, with chunk-level candidates trading within-window reactivity for parallel evaluation.All WAMs in this family listed in the survey use chunk-level rollout; PointWorld, DreamAvoid [33], and EgoExo-WM [152] illustrate candidate scoring or selection. Step-wise rollout can update actions after each imagined state, but loses single-window parallelism.Source: §4.3.1 · pp. 18–19

4.3.2 Joint generation

• Joint generation produces substrate and action through one generative process, coupling their predictions but risking instability when their training losses conflict.Implementations include joint diffusion, joint autoregression, and mixture-of-experts hybrids; some pretrain the substrate head before adding the action head or use a careful loss-weight schedule.Source: §4.3.2 · pp. 19–20

4.3.3 Post-prediction action head

• Post-prediction heads decode actions from a generated substrate, often keeping the predictor frozen so a separate action expert can adapt across embodiments.The substrate may be a rendered video plan, structured visual or geometric representation, or latent feature; some methods bypass or shorten generation at inference, provided the retained state preserves action-relevant foresight.Source: §4.3.3 · pp. 20–21

4.3.4 Action representation and chunk size

• Actions may be continuous controls, discrete codebook tokens, or learned latent actions, while chunk size trades per-action latency against the ability to react during execution.One-action passes invoke the backbone at control frequency; longer chunks spread its cost but reduce reactivity, pushing real-hardware control toward smaller chunks or cheaper backbones.Source: §4.3.4 · p. 21

4.4.1 Iterative-denoising backbones (video diffusion)

• Diffusion backbones iteratively denoise a future substrate, with action-conditioned rollout, joint generation, and post-prediction heads as distinct coupling patterns.Joint generation can encourage action–future consistency, but competing generation and action losses may destabilize training; practitioners often pretrain the substrate head before adding the action head. Latent, feature, and geometric substrates reduce appearance burden, but their usefulness depends on retaining task-relevant information.Source: §4.4.1 · p. 22

4.4.2 Autoregressive backbones (next-frame, next-token)

• Autoregressive backbones serialize the chosen generative output into a causal stream, so autoregression itself does not determine whether actions are conditioned, jointly generated, or decoded later.Caching avoids recomputing the full history, but generation remains serial; latency depends on the stream’s contents and size, including frames, latent blocks, visual tokens, or actions.Source: §4.4.2 · pp. 23–24

4.4.3 Joint-embedding-predictive backbones

• Joint-embedding-predictive backbones deterministically predict future features without decoding them, enabling cheap inference while leaving actions to rollout planning or a separate post-prediction head.Because the embedding space lacks an intrinsic visual-quality measure, downstream-task accuracy is the test of the world model.Source: §4.4.3 · p. 24

4.4.4 Hybrid backbones (generative head plus action head)

• Hybrid backbones share a trunk between generative and action heads, supporting joint generation, post-prediction action decoding, or action-conditioned rollout.Joint generation can align predicted futures and actions but requires balancing competing objectives; post-prediction heads can be retrained per embodiment if the substrate is action-informative, while repeated candidate evaluation can make rollout costly.Source: §4.4.4 · pp. 25–26

4.4.5 LLM- and VLM-backbone WAMs

• LLM- and VLM-backbone WAMs qualify when a future substrate remains in the inference-time control pathway; direct observation-to-action policies without one fall outside this family.They inherit the cost of the VLM stack, prompting many designs to use action chunking and a frozen backbone at inference.Source: §4.4.5 · p. 26

4.4.6 The five families together

• The five backbone families are not mutually exclusive, and substrate choice limits which families are compatible; the remaining design freedom concerns where prediction sits relative to the control loop.Source: §4.4.6 · pp. 26–27

4.5.1 Open-loop rollout

• Open-loop rollout generates a long future once before execution, so WAM cost is paid per scenario rather than per control step, but the plan cannot react to real-world deviations.The actor consuming the rollout must absorb distribution shifts on its own.Source: §4.5.1 · p. 27

4.5.2 Chunked closed-loop control

• Chunked closed-loop control amortizes backbone cost across action chunks, but long chunks can become stale; task-aware variants replan when forecasts are likely to change the action.FFDC-WAM [161] triggers early replanning when predicted and observed latents drift, while NoiseGate [67], WALL-WM [92], and PALM [107] adapt denoising, chunk timing, or subtask progress.Source: §4.5.2 · pp. 27–28

4.5.3 Single-step closed-loop control

• Single-step closed-loop control predicts one substrate and action per call, maximizing reactivity but requiring Nfwd(1) < 1/fctrl and usually a sparse substrate.WorldVLA [14], VidMan, and other cited methods sit closest to this pattern; it is feasible when a forward pass fits inside the control period.Source: §4.5.3 · p. 28

4.5.4 Interactive simulator operation

• Interactive simulator operation continuously incorporates actor inputs and reuses cached or persistent state, enabling mid-stream steering while shifting cost pressure toward memory.For attention-based backbones, cached forward-pass cost is linear in M(t) per step and cumulative cost is quadratic in t.Source: §4.5.4 · p. 29

4.6 Putting the Four Ingredients Together

• WAMs are located by a 4-tuple of substrate Φ, action coupling F, backbone B, and deployment regime D, spanning four substrate categories, three coupling families, five backbone families, and four regimes.Substrates are pixel-grounded, feature, geometric, or affordance; coupling is action-conditioned rollout, joint generation, or a post-prediction head; backbones are diffusion, autoregressive, joint-embedding, hybrid, or LLM/VLM; regimes are open, chunked, single, or interactive.Source: §4.6 · p. 29
• Example placements distinguish methods across the tuple: F1 [111] and UWM share a chunked hybrid joint-generation route, while WorldVLA [14] is autoregressive/single and FLARE [204] post-prediction/joint-embedding.F1, WorldVLA, and UWM use pixel-decodable latent grids; FLARE uses a latent teacher-target substrate.Source: §4.6 · p. 29
• Wan [156] and CogVideoX [185] alone are not WAMs because they fix only substrate and backbone; adding coupling and deployment yields wrappers such as CosmosPolicy [79], VidMan, and VideoVLA [140].Source: §4.6 · p. 29
• WAM design attention centers on training-time substrate and backbone choices, while inference-time coupling and deployment are more flexible and host sharper mechanism variation.The expensive corner combines decoded pixel-grounded representation, joint generation, diffusion, and single-step closed-loop deployment; few methods use it directly in production.Source: §4.6 · p. 29
• The census applies the tuple to Render-and-Decode and Latent-Only methods in Table 1 and Video-Generation-Free methods in Table 2.Source: §4.6 · p. 29, Table 1 · p. 30, Table 2 · p. 31

5.1 Interactability

• Action becomes interactable when it enters prediction: post-generation decoding is modular but pays full generation cost, while earlier coupling shapes futures more directly and bottlenecks trade context for cheap inference.The design spectrum runs from decoding control after a generated future to shared action-and-future prediction and narrow future channels.Source: §5.1 · pp. 32–33

5.2 Causality

• Causal WAMs prevent future-frame leakage while exposing next actions quickly, using cached token streams, leakage-free chunk masks, early denoising exits, and overlapping prediction with execution.The controller often needs the next action before a fully refined future video; causal structure determines both what information is available and where computation is spent.Source: §5.2 · p. 33

5.3 Persistence

• Persistence must balance compounding drift, history-cost growth, and forgotten scene evidence; observation replacement re-grounds cheaply, while bounded-memory strategies constrain long-horizon cost.DreamZero uses key-value-cache observation replacement for 7 Hz closed-loop control with a 14B autoregressive video-diffusion model [188]; DexWM operates over 900 hours of interaction with bounded test-time memory [44].Source: §5.3 · p. 34

5.4 Physical Plausibility

• Physical plausibility favors the coarsest embodiment-grounded future that still constrains action, from geometry, flow, and masks to latent forecasts, rather than photorealism alone.The abstraction ladder trades inspectable appearance for capacity focused on geometry, contact, and dynamics; how useful a future is for action predicts task success better than how good it looks.Source: §5.4 · pp. 34–35
• Contact-rich control also needs tactile, force, and proprioceptive or kinematic constraints, which make predictions action-relevant and feasible for the executing body.These low-dimensional signals carry dense information where contact determines success, while embodiment-aware constraints shrink predictions to the robot’s feasible manifold.Source: §5.4 · pp. 34–35

5.5 Generalization

• Generalization draws on video priors, transferable flow or mask and latent substrates, and action abstractions that separate transferable dynamics from embodiment-specific execution.The transfer sources include video pretraining [16], flow [207], semantic masks [109], feature forecasts [112] [200], and abstract action/data recipes [112] [149] [171].Source: §5.5 · pp. 35–36
• No design generalizes across every shift: methods should predict at the invariant level that still constrains control and state the target shift, transferable axis, and robot-specific adapter.Pixels carry semantic and physical priors but appearance costs; flow and masks need reliable executors; latent actions and features need grounding for precise control.Source: §5.5 · p. 36
• Across the five properties, gains in action coupling, bounded memory, or transferable abstraction can impose costs on leakage control, long-horizon plausibility, or evaluation.The section frames WAM design as competing pressures whose balance should be made measurable rather than asserted.Source: §5.5 · p. 36

6.1 Data sources

• WAM training draws on five source groups—teleoperation, portable human demonstrations, internet video, simulation, and synthetic trajectories—to balance scale, action-label quality, physical grounding, and access.In practice, sources are mixed: internet video supplies priors, teleoperation trusted labels, portable data scale, simulation controllable coverage, and synthetic trajectories help fill costly collection gaps.Source: §6.1 · pp. 36–37
• Teleoperation offers clean action-conditioned trajectories but consumes robot and operator time; portable human demonstrations scale collection but leave an embodiment gap to close.Teleoperation records commands with visual observations, while portable demonstrations capture human video and hand tracks without a robot in the loop.Source: §6.1 · pp. 36–37
• Internet video supplies broad visual variation, object dynamics, and manipulation patterns, but usually lacks robot actions, so its use requires explicit action recovery or proxy learning.It can also pretrain video backbones or strengthen latent-future prediction before action decoding.Source: §6.1 · p. 37
• Simulation provides exact labels, controlled curricula, and low marginal cost after authoring, but its sim-to-real gap often requires internet-pretrained priors or abstract substrates.Simulation spans classical physics platforms, general backends, household and digital-twin datasets, and aggregators.Source: §6.1 · p. 37
• Synthetic WAM trajectories can fill expensive physical-data gaps, but inherit generator failure modes; limited access to some generators and datasets also constrains independent verification.Generated videos may be converted into pseudo-actions or recovered actions for policy training.Source: §6.1 · p. 37

6.2 Evaluation

• Evaluation must pair visual prediction quality with closed-loop action success, while also testing physical realizability, long-horizon coherence, and compute, memory, and latency requirements.Visual metrics are cheap and familiar but reward realism rather than action utility; a visually sparse representation may still support successful control.Source: §6.2 · p. 38
• Closed-loop benchmarks measure task completion across simulation, real hardware, and learned-model rollouts, trading throughput against physical validity and evaluation cost.Cascaded methods also require separate substrate and execution checks because the generated future and action translator can fail independently.Source: §6.2 · p. 38
• Physical-plausibility and long-horizon metrics remain underdeveloped, motivating inexpensive visual and representation screens followed by selective closed-loop tests under explicit resource budgets.Current evaluation lacks hour-scale rollout at control frequency, and the conversion from cheap screens to expensive tests remains an open question.Source: §6.2 · p. 38

7 Open Challenges

• The challenges are analytically distinct but coupled: data shapes generalization, action abstraction affects physics, and inference cost influences nearly every design choice.The discussion first considers how much future computation WAMs should build in or spend at runtime, then turns to which data should train each stage.Source: §7 · p. 39

7.1 Dream More or Act More?

• WAMs trade detailed future forecasts for faster action: S-VAM distills video diffusion into foresight features, while Fast-WAM, GigaWorld-Policy, and X-WAM reduce or overlap inference-time video computation [51] [181] [186] [193].These approaches suggest the video-training objective can matter more than producing a fully rendered forecast at test time.Source: §7.1 · p. 39
• DreamZero, CosmosPolicy, and NovaPlan retain stronger imagination in control at a latency cost; the challenge is a controllable fidelity–latency curve, not simply smaller models [38] [79] [188].Runtime budgets could adapt to expected action value: use more computation near uncertainty, contact, or irreversible error, and less for routine actions. Existing replanning triggers can respond to task phase, denoising state, or process cues, but binary decisions do not determine how much horizon to regenerate [55] [140] [161] [33] [36] [67] [146].Source: §7.1 · p. 39
• A stronger design would train a shared backbone with video objectives and let its action path choose inference-time computation, balancing control-rate latency against the loss of rendered forecasts when video tokens are skipped [129] [166] [198] [186].Video generators add action conditioning, while WAMs also borrow latent or joint-embedding substrates; video-scale models need substantial optimization for control rates [188] [193].Source: §7.1 · p. 39

7.2 What Data Should Each Stage Learn From?

• Training should assign each data source to the stage, objective, and model component it supervises: internet video teaches visual regularities, demonstrations align motion with intent, robot teleoperation grounds action, and rollouts train gates or budget controllers.Pooling sources can obscure which variable each one teaches.Source: §7.2 · p. 39
• Broad video pretraining can support downstream action when robot data are too narrow, but robot-domain video helps where mismatched egocentric video can hurt, and human-video gains vary by evaluation split [16] [41] [62].Source: §7.2 · p. 40
• Alignment data formats have distinct biases: EgoScale, EgoMimic, and EgoVerse support staged or diverse human-video training, while FRAPPE distinguishes web-scale egocentric, task-specific egocentric, and robot data roles [77] [135] [205] [200].The minimum capture hardware, viewpoint match, and data quality needed for each embodiment remain unresolved.Source: §7.2 · p. 40
• Action-free methods reduce labeled robot-control needs but still require executable-action grounding; synthetic trajectories likewise inherit generator errors, and contact-rich demonstrations remain difficult to capture [20] [21] [43] [59] [175] [112] [149] [171] [187] [73] [117].Abstract actions need a decoder, inverse-dynamics model, or robot-specific calibration; generated trajectories need grounding in robot dynamics. A joint scaling law across source, stage, quality, model size, and action grounding is still missing.Source: §7.2 · p. 40

7.3 Can Memory Keep Up?

• Useful WAM memory must maintain task state during action production, with costs growing with scene complexity rather than episode length.The open challenge is not simply extending the prediction horizon.Source: §7.3 · p. 40
• IRASim accumulates temporal drift despite overlapping context, PhysGen is bounded by fixed physical-package context, and sparse-memory chunking degrades in dynamic scenes [66] [144] [214].These are memory failures, not only prediction-quality failures.Source: §7.3 · p. 40
• Test-time memory bounds per-step recall but risks forgetting earlier state [27], while Dreamer 4 enables continual closed-loop imagination but leaves long-term object-level spatial detail unresolved [58].The challenge is to combine bounded memory, spatial indexing, and observation replacement in a method that remains reactive over long tasks.Source: §7.3 · p. 40

7.4 How Can WAMs Generalize?

• Domain shifts stress different model components, and Act2Goal drops steeply while FRAPPE remains far below a practical rate under harder randomized conditions [200, 208].Appearance, object dynamics, morphology, action space, camera placement, and contact regime should be treated as distinct shifts; scale alone does not ensure generalization.Source: §7.4 · p. 40
• A video-conditioned policy improves over baselines, but the hardest motion and object splits retain a large gap, so prediction helps only when its substrate matches the shift.Source: §7.4 · p. 41
• Generalization requires matching data, substrate, and adaptation to a declared shift: DreamZero transfers after short video-only adaptation [188], while EgoScale reports a dexterous-pretraining scaling law [205].Removing video pretraining from a strong policy sharply reduces task completion [62], but does not establish which embodiment changes the prior can absorb; the challenge is testing the chosen substrate and action decoder against the target shift.Source: §7.4 · p. 41

7.5 What Grounds Abstract Action?

• Abstract targets such as latent codes, flow fields, and point tracks enable learning from action-free or weakly labeled video, but 2D flow misses depth, out-of-plane rotation, and contact force [81].Rigid-transform decoders also fail for articulated or deformable objects.Source: §7.5 · p. 41
• Point-track and flow-based adapters make action interfaces inspectable, but inherit errors from trackers, depth estimators, and local-linearity assumptions [49, 70, 94].Source: §7.5 · p. 41
• Latent actions can remain opaque: LAPA struggles with fine grasping when its codebook is coarse [187], ALAM's consistency is not a force or contact explanation [149], and LDA-1B execution relies on an embodiment-specific decoder [112].The open challenge is retaining abstract actions' data advantage while adding physical grounding and an inspection handle for failure analysis.Source: §7.5 · p. 41

7.6 When Is a Future Physical?

• Visual plausibility does not ensure embodied execution: frame-conditioned flow misses momentum, while tactile proxies, point clouds, and smoothness rewards can omit contact forces or material properties.A flow-based model conditioned on one frame fails push tasks when objects continue moving after contact ends [50]; tactile-flow divergence gives an uncalibrated force proxy, point clouds omit force magnitude, friction, and compliance [192], and smoothness and joint-limit scores can miss contact [162].Source: §7.6 · p. 41
• Physics-aware substrates and losses offer partial routes toward futures judged by embodiment-level executability, not visual convincingness.RoboScape and RoboScape-R add geometric and consistency constraints; OmniVTA and DexWM incorporate tactile or dexterous state; AdaWorldPolicy trains with action-conditioned future structure [44] [139] [148] [191] [206].Source: §7.6 · p. 41

7.7 What Should Evaluation Report?

• World-model scores may poorly predict policy utility, while averaging task phases can conceal failures in precise interaction.MotuBrain finds only weak correlation between its world-model score and downstream success [121]; HarmoWAM’s transit-useful schedule can fail at precise interaction [36].Source: §7.7 · p. 41
• Evaluation needs an accuracy-at-budget protocol that reports success, latency, sustained horizon, peak memory, and contact-sensitive failure tags together.Latency is often reported apart from accuracy and across different hardware or task settings [30] [69] [90] [113] [183].Source: §7.7 · p. 42
• Fast-WAM isolates test-time video generation after video co-training, while SANTS and NoiseGate pair success with learned reductions in denoising cost.These are emerging examples of more informative evaluation reporting [67] [146] [193].Source: §7.7 · p. 42

The challenges are coupled

• WAM design trade-offs couple memory, compute, substrate richness, action abstraction, contact data, and transfer errors rather than permitting isolated fixes.Deeper search raises compute; richer substrates improve contact realism but increase evaluation and inference costs; latent actions reduce labeled-control needs but weaken physical grounding without calibrated decoding. Contact-aware prediction needs force, tactile, or failure data unavailable in cheap visual screens, while upstream estimator errors can affect grounding and transfer.Source: The challenges are coupled · p. 42

• WAMs must use past observations, past actions, and task context to anticipate an actionable future while producing the action that should follow.This framing explains why differently named works can belong to the same family.Source: §8 · p. 42
• The survey organizes methods by what they generate before action decoding and by their predictive substrate, backbone, action coupling, and deployment regime.Together, these views support analysis of interactability, causality, persistence, physical plausibility, and generalization.Source: §8 · p. 42
• Datasets and benchmarks matter because they teach or validate only the action variables and properties they record, infer, or expose.The survey argues that data and evaluation are not secondary details.Source: §8 · p. 42
• Across the surveyed works, the trajectory is toward selectively imagining the future needed for action rather than simply increasing video prediction.The remaining challenges include choosing a substrate, coupling action without leakage, preserving long-horizon state, grounding predictions in physical constraints, and reporting capability alongside compute, memory, and latency.Source: §8 · p. 42