The World Action Model Atlas

World ActionModels

Dream Less, Act More

A World Action Model predicts a future and keeps that forecast in the path that produces the action. That contract, not any particular backbone, is what makes it a WAM. This is a full walkthrough of the survey that named the family: the boundaries it draws, its two viewpoints, its four-axis anatomy, the properties embodiment demands, and its thesis that the strongest models generate less of the future while keeping what control needs.

Following Shen, Zhang, Liao, Li, Tan, Wang, Yan & Wang (NUS), World Action Models: A Survey (arXiv 2606.20781v1, Jun 2026). A study map of one survey's framework, not a settled consensus. Method names are the survey's; any census-cell assignment is a reading of its prose.

prediction / future (dream) action (act) the action-facing coupling
observation + task o≤t , l VLA: skip the future, act from the present predict future substrate s DREAM make it action-facing action chunk a_(t:t+H-1) ACT robot
Section 01 · the definition

01What separates a WAM from its neighbors

The label spread across three communities faster than the idea did, so the survey opens by drawing hard lines. Five nearby objects, told apart by what each is required to model. The point to hold: a WAM is defined by a contract on where the future goes, not by which backbone produces it.

POLICY

Vision-Language-Action

Maps an observation and instruction straight to an action. It inherits semantic and spatial priors from a pretrained VLM, but it never models what will be observed after it acts, so it cannot reason about the consequence of its own intervention.

p( a | o, l )
PREDICTOR

World model

Predicts the future observation that follows from the current state and an intervention. The future may be pixels, a hidden feature, a geometric state, an affordance, audio, or a symbol. It imagines, but does not necessarily choose the action.

p( o' | o, a, l )
GENERATOR

Video generation → video world model

Generates a future visual from conditioning. When the conditioning includes the agent's action and the output is a future observation, the same family becomes a video world model: a high-capacity, internet-pretrained source of visual dynamics priors (Wan, CogVideoX, Cosmos, Sora).

p( o' | r )  →  p( o' | o, a, l )
THE WAM CONTRACT

World Action Model

Links the two. The predicted future is kept in the action path: it helps produce the action, score a candidate action, or train the action head inside one model. A direct VLA with an auxiliary future loss, a simulator used only as an RL environment, or a future head discarded before action use does not qualify.

p( o', a | o, l )

1.1Three ways the future reaches the action

The coupling is a factorization of the joint. All three are WAMs; they differ in whether prediction leads, follows, or fuses with the action.

Predict-then-act (cascade)

Imagine a future, then decode the action from it with a separate module: inverse dynamics, a pose tracker, a trajectory optimizer, a planner, or a policy. The historically early form.

p(o',a|o,l) = p(o'|o,l) · q(a|o,o',l)UniPi, VLP, AVDC
Act-then-predict (scoring)

Propose an action, predict its consequence, and let the predicted consequence decide which action executes. Covers candidate scoring and planning.

p(o',a|o,l) = q(a|o,l) · p(o'|o,a,l)3D-ALP, PointWorld
Joint generation

One backbone or tightly coupled experts emit the future and the action together in a shared state, through diffusion, autoregression, or a hybrid.

L = E[ -log p(o',a|o,l) ]GR-1/2, PAD, UWM, WorldVLA
Section 02 · view one · design philosophy

02How far do you carry the dream?

The first viewpoint sorts every WAM by where action is decoded along the inference path: at rendered pixels, at an intermediate latent, or outside the video path entirely. It is a spectrum of how much future you generate before you act, it is exhaustive over the census, and the field is sliding down it. Three mutually exclusive philosophies, each with its internal branches.

P1

Render-and-Decode

carry generation to pixels

Run the video backbone all the way to pixel output, then decode the action from the rendered future. The premise is that a full visual future is worth producing inside the control loop because it preserves the entire visual prior and exposes appearance, motion, and contact in an inspectable form. The standing cost: synthesis is in the latency budget every step, and video-quality metrics only weakly predict task success.

o, lvideo gento pixelsframesaction decoder
Founding · video as the plan

Text-conditioned video, then recover action. UniPi (inverse dynamics), VLP (search over language plans), AVDC (dense correspondences replace action labels).

Demonstration sources

Move the interface onto human and task video. Dreamitate (track stereo video), Gen2Act (generated human motion), This&That (gesture coords), GR-MG (progress-aware goal image).

Modular recovery

Keep world model and actor separate and reusable across embodiments. DreamGen (offline pseudo-actions), RIGVid, VERA (action-free planner plus Jacobian inverse-dynamics translator).

Geometry and search between future and action

NovaFlow, Dream2Flow (distill video into 3D object flow), 3D-ALP (3D world model as a rollout oracle in test-time search).

Joint (action inside the backbone)

GR-1 anchors it: future image tokens and action tokens in one autoregressive stream. GR-2 scales it, WorldVLA masks to cut action leakage, PAD denoises images and actions in one DiT, Motus denoises sparse video and action together.

Richer rendered state

Render more than RGB so geometry is easy to extract. TesserAct (RGB-D-Normal 4D), X-WAM (multi-view RGB-D plus 3D reconstruction), Dex4D (frames to 3D point tracks).

Sparse future

Render only the milestones the actor needs. CoT-VLA (visual chain-of-thought before action), pi0.7 (async subgoal images to a PaliGemma action expert), BagelVLA.

Foundation-scale and selective

Make a large video prior responsive. DreamZero (Wan-class online video-action policy, chunked KV-cache replacement), Vidar, Veo-Act; and spend it only when it helps: DreamAvoid (imagine on critical phases), CKT-WAM.

dreams → the most. Full visual prior, full pixel cost every step, only weakly measured by video fidelity.
P2

Latent-Only

keep the video prior, stop before pixels

Keep the video-world-model lineage but remove pixel decoding from the control path. Action is read from an intermediate latent, a partially denoised feature, a flow field, a mask, or a value map, each cheaper than a full render. The recurring move: a video trunk whose VAE decoder is simply bypassed at inference. You keep the temporal and physical structure learned from video and drop the renderer, which is what lets several of these reach real time, at the cost of the pixel-space supervision that makes rendered futures inspectable.

o, lvideo trunklatentlatentdecoderaction head
Interception point (stop at the latent)

VPP (inverse dynamics on Stable Video Diffusion features), VILP, UVA, and Genie Envisioner (GE-Base video DiT; GE-Act decodes action chunks from a one-step denoised latent via flow matching).

Tap the denoising trajectory

mimic-video (action decoder on an intermediate ODE checkpoint of Cosmos-Predict2), DiT4DiT, S-VAM (self-distill multi-step denoising to one pass), Fast-WAM (make full generation optional at inference).

Structured motion instead of texture

Im2Flow2Act, 3DFlowAction (object flow), TraceGen (3D traces to joint commands via IK), 3PoinTr, RoboFlow4D.

Structured substrate, contact and task state

MWM (semantic-mask evolution), OmniVTA (tactile latents), EgoExo-WM (DINOv3 latent regressor under MPC).

Joint backbone, latent action path

UWM is the reference: independent diffusion timesteps let it collapse the visual branch at inference while keeping the video-trained representation. CosmosPolicy, MotuBrain, AIM (action through a spatial value map).

Adaptive imagination and memory

DAWN (refine V-JEPA latent world tokens against DiT action hypotheses), CLWM (forecast DINOv3 features, Dual-State test-time-training memory), FFDC-WAM and tau0-WM (invoke the heavy future model only when confidence drops).

dreams → in latent space. Video prior kept, renderer skipped, often real-time, less interpretable.
P3

Video-Generation-Free

drop the video backbone entirely

Remove the pixel-level video generator from the predictive path (the LLM, VLM, JEPA, and flow-matching parts are still generative in the broad sense). The future is predicted in a compact space: LLM or VLM tokens, a joint-embedding-predictive encoder, a deterministic regressor over frozen vision features, or a non-video diffusion head over flow or affordances. The predictive signal survives, judged by action success rather than photorealism. The newest of the three families, and the sharp end of "dream less."

o, lVLM / JEPAno video genfuture tokensaction expert
Feature prediction (teacher / foundation targets)

FLARE aligns policy tokens to future-observation embeddings from a frozen teacher; FRAPPE uses multiple vision foundation models; DexWM regresses future features; LDA-1B learns dynamics, policy, and forecasting in a structured DINO latent on a Qwen3-VL stack.

Learned token or latent transition

DUST augments an Eagle-2 VLM with a dual-stream diffusion head predicting future embedding tokens and action together; ALAM learns latent transitions from action-free video with composition and reversal consistency.

Structured non-pixel futures

PointWorld (action-conditioned 3D point flow, MPPI planning), PALM (affordance maps plus progress), HiF-VLA (motion-vector foresight tokens), ICLR-VR (gripper-keypoint polyline).

Other modalities and feedback

Audio-WM (future Mel latents when contact sound carries task evidence), DDP (latent imagination on out-of-distribution events), Feedback-WM (latent transition model as an observer correcting diffusion-policy denoising online).

dreams → the least. No video backbone, the future lives as tokens, features, geometry, or audio.

Two clarifications the survey stresses. This where-is-action-decoded split is separate from cascaded-versus-joint: a Latent-Only WAM can be either, and so can a Render-and-Decode one. And the split is chronological: Render-and-Decode came first, Latent-Only emerged as the field dropped pixel decoding from the control path, and Video-Generation-Free arrived as LLM, VLM, and JEPA backbones learned to carry the future without a video core. The application domain broadened over the same period to tabletop manipulation, dexterous hands, tactile contact, driving, and aerial work.

Section 03 · view two · component anatomy

03Every WAM as a 4-tuple

The second viewpoint treats a WAM as a conditional joint over a future trajectory and an action chunk, then reads off four separable axes. The philosophy fixes what must be generated; the 4-tuple fixes how it is built. Two methods with the same name but different choices on any one axis behave nothing alike.

pΘ( st+1:t+H , at:t+H-1 | o≤t , a<t , l )  ·  predict a future substrate s and an action chunk a of horizon H from the context

Axis 1Predictive substrate · where it dreams

The space of the future the WAM exposes to the actor. Richer substrates carry more and cost more.

Pixel-grounded

Decoded future frames, or the VAE / VQ latent grid that decodes to a clip. Full appearance and motion.

GR-2, TesserAct, DreamZero
Feature

A learned hidden state or token block with no decodable image: a teacher target, a generative-trunk feature tap, or a VLM future-token block.

VPP, FLARE, DAWN
Geometric primitive

A structured object in physical coordinates: flow, point clouds, depth, pose, or motion vectors.

PointWorld, TraceGen, 3DFlowAction
Affordance map

A task-relevant label or score map: a value map, affordance, or end-effector heatmap.

AIM, PALM

Axis 2Action coupling · how action enters and leaves

A factorization of the joint. Rollout can be chunk-level (predict a whole action chunk, then the future) or step-wise (interleave one action and one future step at a time).

Action-conditioned rollout

The action drives the prediction. Chunk-level or step-wise, the latter giving finer causal control at higher cost.

q(a|c) p(s|c,a)
Joint generation

Future and action emerge together in one generative state, coupled tightly by construction.

p(s,a|c)
Post-prediction head

Predict the future first, then a separate head reads it to produce the action. The cascade form.

p(s|c) q(a|s,c)

A fourth choice sits underneath all three: action representation and chunk size, discrete tokens versus continuous flow, and how many steps H are committed before replanning. Larger chunks are cheaper but stale faster.

Axis 3Architectural backbone · how the prediction is produced

Five function families realize the future-prediction factor. The first two dominate Render-and-Decode; JEPA and LLM/VLM backbones power most Video-Generation-Free work.

Iterative-denoising (video diffusion)

Denoise future frames or latents over multiple steps. Highest fidelity, highest per-step cost.

PAD, UWM, Genie Envisioner
Autoregressive (next-frame, next-token)

Predict the next frame or token, interleaving image and action tokens in one causal stream.

GR-1, GR-2, WorldVLA
Joint-embedding-predictive (JEPA)

Predict in a learned embedding space, no pixel decode. Cheap, abstract, hard to inspect.

V-JEPA, DAWN
Hybrid (generative head + action head)

A shared trunk feeding a generation head and an action head, tuned to route work between them.

DUST, AIM
LLM- and VLM-backbone

Attach a lightweight action expert to a language or vision-language model; the future is a token block.

FLARE, LDA-1B

Axis 4Deployment regime · how it runs in a loop

Where latency meets the three choices above. This axis decides whether a design is usable in real control.

Open-loop rollout

Predict a future and execute the derived actions without re-observing mid-rollout. Simplest, most drift-prone.

Chunked closed-loop

Predict a chunk, execute it, re-observe, replan. The common practical middle ground.

Single-step closed-loop

Re-observe and re-predict every step. Tightest control, highest compute per action.

Interactive simulator

Run the model as a steerable world you feed actions into, for planning, evaluation, or data generation.

Section 04 · the census

04Representative WAMs on one row each

The two viewpoints on a grid: philosophy on the left, then the four-tuple. Read a row and you have placed a method. Assignments follow the survey's prose; treat any single cell as a reading, not a quoted table.

MethodPhilosophySubstrateBackboneCouplingRegime
UniPiRender-and-Decodepixeldenoisingpost-head (inv. dyn.)open-loop
AVDCRender-and-Decodepixeldenoisingpost-head (correspondence)open-loop
GR-1 / GR-2Render-and-Decodepixel (VQ tokens)autoregressivejointchunked closed-loop
WorldVLARender-and-Decodepixel (VQ)autoregressivejoint (masked)chunked closed-loop
PADRender-and-Decodepixeldenoising (DiT)jointchunked
MotusRender-and-Decodepixel (sparse)denoisingjointchunked
TesserActRender-and-Decodepixel (RGB-D-Normal 4D)denoisingpost-headopen-loop
CoT-VLARender-and-Decodepixel (sparse CoT)autoregressive / VLMpost-headchunked
DreamZeroRender-and-DecodepixelAR video diffusionjoint (online)chunked closed-loop
VPPLatent-Onlyfeature (SVD latent)denoisingpost-headclosed-loop
Genie EnvisionerLatent-Onlyfeature (1-step latent)denoising (DiT)post-head (flow)chunked closed-loop
mimic-videoLatent-Onlyfeature (ODE checkpoint)denoisingpost-headclosed-loop
TraceGenLatent-Onlygeometric (3D trace)denoisingpost-head (IK)open-loop
UWMLatent-Onlypixel-latentdenoisingjointclosed-loop
AIMLatent-Onlyaffordance (value map)hybridpost-headclosed-loop
DAWNLatent-Onlyfeature (V-JEPA)JEPAjoint refineclosed-loop
FLAREVideo-Gen-Freefeature (teacher target)LLM / VLMpost-headclosed-loop
DUSTVideo-Gen-Freefeature / tokenVLM + diffusion headjointclosed-loop
LDA-1BVideo-Gen-Freefeature (DINO latent)LLM / VLMjointclosed-loop
PointWorldVideo-Gen-Freegeometric (3D point flow)deterministic regressoract-then-predict (MPPI)closed-loop
Audio-WMVideo-Gen-Freefeature (Mel latent)non-videopost-headclosed-loop
Section 05 · the embodiment test

05Five properties control actually demands

Placing a WAM is half the survey. The other half is a critical checklist: what a predictive-action model must satisfy once a robot depends on it. Every design choice above is ultimately judged against these, and against what each solution costs in compute, memory, data, and evaluation effort.

01

Interactability

Accept control signals during generation, not only at the start of a rollout, so the future can be steered as intent changes.

02

Causality

Never let a predicted future leak into the action executed now. The forecast informs; it must not encode the answer it is meant to help find.

03

Persistence

Hold long-horizon predictions together as the robot acts and replans, without drift or forgetting across the episode.

04

Physical plausibility

Predict futures the embodiment can actually realize, respecting contact and dynamics, not merely plausible-looking video.

05

Generalization

Keep the predictive-action contract useful across tasks, objects, scenes, cameras, and embodiments, not just the training distribution.

Section 06 · data and evaluation

06What trains a WAM, and how you judge it

The survey organizes the fuel and the yardsticks. Both are unsettled, and the mismatch between them is itself a finding: the metrics inherited from video generation do not measure what control cares about.

6.1Data sources

Robot teleoperation

Action-labeled demonstrations collected on the target robot. Highest fidelity, lowest scale, most expensive.

Portable human demonstrations

Handheld or wearable capture of people doing tasks. Cheaper and broader, but the embodiment gap must be crossed.

Internet-scale video

Egocentric and instructional video at web scale. Rich dynamics priors, but action-free, so the action must be recovered or inferred.

Simulation

Cheap, labeled, resettable, and safe. The sim-to-real gap in appearance and contact is the standing tax.

Synthetic from WAMs

A WAM generates its own training futures, a data engine that risks compounding its own errors if unfiltered.

6.2Evaluation practices

Visual fidelity

Video-quality metrics on the predicted future. Easy to compute, but they only weakly predict task success.

Closed-loop benchmarks

Task success in the loop, in sim or on hardware. What actually matters, and the hardest to standardize.

Physical plausibility & coherence

Whether long-horizon futures stay dynamically consistent and realizable, not just sharp per frame.

What current evaluation misses

The gap between fidelity and success, causality leaks, and cost inside the loop go largely unmeasured.

Section 07 · open challenges

07What the four-axis view brings into focus

The survey closes on seven open problems, grouped by where they arise in the WAM stack. Each is a live research direction, not a solved trade-off.

7.1

Dream more or act more?

When is a richer imagined future worth its latency, and when is a compact substrate, or no rendered future at all, the better bet? The field's central open tension.

7.2

What data should each stage learn from?

Pretraining, alignment, and action learning may each want different sources. Matching data to stage is unresolved.

7.3

Can memory keep up?

Long-horizon control needs persistent state, but caches grow and attention costs rise. What memory mechanism scales?

7.4

How can WAMs generalize?

Holding the predictive-action contract useful across tasks, embodiments, cameras, and scenes remains hard.

7.5

What grounds abstract action?

Bridging a compact or latent future to an executable, embodiment-specific action is an open interface problem.

7.6

When is a future physical?

Distinguishing a dynamically realizable prediction from a merely photoreal one needs better notions and tests.

7.7

What should evaluation report?

Given that fidelity weakly predicts success, the field still lacks agreed measures that capture control-relevant quality and cost.

Section 08 · the thesis

08Dream less, act more

Across every axis the same pattern emerges: predictive richness is bought with compute, memory, latency, and action-label cost, all of which land inside a control loop. So the strongest WAMs tend to generate less of the future while still acting on what they need. The chronological arc is that trajectory made visible.

2023
Render the plan

Generate a full video future, then decode action. UniPi, VLP, AVDC.

2024
Join future and action

Shared video-action streams and diffusion. GR-1/2, PAD.

2025
Drop the renderer

Stop at latents and features, reach real time. VPP, Genie Envisioner, UWM.

2026 →
Drop the video model

Predict in VLM, JEPA, or geometric space. FLARE, DUST, PointWorld.

The interesting question, the survey insists, is never how much a WAM costs on its own. It is how that cost trades against the five properties embodiment imposes. That trade is the whole design space, and it is where your own work gets positioned.

Section 09 · the positioning lens

09Placing a model, and your own work

Five questions that force the framework into working memory, phrased the way they get asked.

What makes a WAM, versus a VLA or a world model?
A VLA is p(a | o, l), action from the present, no future. A world model is p(o' | o, a, l), a future with no committed action. A WAM keeps the predicted future action-facing: it produces, scores, or trains the action. A future generated and discarded before the action does not count, and neither does a sim used only as an RL environment.
Give me the three design philosophies in one breath.
Render-and-Decode goes to pixels then decodes action. Latent-Only keeps the video prior but stops before pixel decoding, reading action from a latent or feature. Video-Generation-Free drops the video backbone and predicts in VLM, JEPA, or geometric space. A "dream less" spectrum the field is moving down.
How would you specify a WAM precisely?
As a 4-tuple: substrate (pixel, feature, geometric, affordance), coupling (rollout, joint, post-head, plus chunk size), backbone (denoising, autoregressive, JEPA, hybrid, LLM/VLM), and deployment regime (open-loop, chunked or single-step closed-loop, interactive). Same name, different axis, different behavior.
Why "dream less, act more"?
Because predictive richness costs compute, memory, latency, and action labels inside the loop, and rendering a full future rarely pays off relative to a compact substrate carrying what the action needs. The strongest recent WAMs generate less future, judged by the five embodiment properties rather than video fidelity.
What are the five properties, and why do they bite?
Interactability (steer during generation), causality (no future leakage into the current action), persistence (hold long horizons), physical plausibility (realizable, not just photoreal), and generalization (across tasks, cameras, embodiments). They bite because visual-quality metrics do not measure any of them, which is why evaluation is an open problem.