Awesome-WAMs
GitHub
PAPER REPORTENAll readings ↗

DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks

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

Authors: Yueci Deng; Guiliang Liu; Kui Jia

Affiliations: DexForce AI

Source: 2604.16484 ↗ · Catalog record

Reading: 216 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: CLWM uses predicted semantic features to guide actions and speculative computation, while its strong manipulation results leave memory fidelity, latency–accuracy tradeoffs, and several implementation details unresolved. e04e05e07e09e10e13e15e16e17

At a glanceWhat to know
Research problem
Reader analysis

Pixel reconstruction, growing historical KV caches, and sequential sensing/computation can make generative manipulation policies expensive. The authors target semantic dynamics, constant-size historical memory, and lower blocking latency; their stronger claims of physical causal reasoning are motivations rather than separately established causal-identification results. e02e03

Core mechanism
Source description

CLWM replaces pixel or VAE prediction targets with frozen DINOv3 features, then decodes actions through a shared-backbone, two-stage flow model. e04e05

A key reported resultRoboTwin bimanual manipulation: 94.00%

Reported average task success (%). Table 1; common fine-tuning budget of 25,000 synthetic trajectories and 40,000 iterations. Exact test split, rollout counts, and seeds are unstated.

LingBot-VA 91.55%; Motus 87.02%; π0.5 76.76%; X-VLA 72.84%. The reported average exceeds LingBot-VA by 2.45 percentage points. Success is uneven: CLWM reaches only 40% on Hanging Mug and 65% on Turn Switch, versus Motus at 78% on Turn Switch. e13

Reading caution
Reader analysis

No isolated DINOv3-versus-VAE, shared-backbone, TTT-quality, or SAI-accuracy ablation is provided. Data-pipeline ablations cannot establish which architectural component causes the main success gains. The claimed history-augmentation stability guarantee lacks a proof or measured perturbation analysis. e04e09e14e15e16

Core contributions

  • Source description

    CLWM replaces pixel or VAE prediction targets with frozen DINOv3 features, then decodes actions through a shared-backbone, two-stage flow model. e04e05

  • Source description

    Dual-state TTT separates observed history from imagined futures; SAI uses the latter to start the next inference while the robot moves. e07e09

  • Source description

    EmbodiChain combines generated simulation assets, domain expansion, recovery trajectories, and online supervised data streaming. e10e11

Figure 1. Predict a semantic future, then condition the action generator on it. Original paper, p. 4 ↗

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

How to read it. Begin at the bottom: the task prompt, historical latent features, and historical actions supply context. The upper-left DINOv3 path shows how camera images enter the feature space, and snowflakes mark the frozen DINOv3 and language components. Follow the feature branch to predicted future features, then read the action branch as the next stage described in the caption. The broad TTT bar provides historical context to both. Although the drawing separates video and action blocks, Equation (7) on page 5 specifies a shared transformer core with distinct modality projections and flow-time embeddings. The colored feature maps illustrate the representation, without providing a numerical prediction-quality measure. e03e04e05e08e09

What it supports. The architecture places future prediction directly in the deployed control path: action generation conditions on predicted semantics. This supports a unified shared-core model with sequential video-feature and action generation. It also explains why the predicted observation can later serve as a temporary input while physical execution is still underway.

Where the evidence stops. The update label says s = 0, but the flow interpolation on pages 4–5 reaches data at s = 1. Section 3.3 also finishes at 1. The original label is preserved; its timing convention remains unresolved.

2. Motivation

2.1 The problem and the proposed response

Reader analysis

Pixel reconstruction, growing historical KV caches, and sequential sensing/computation can make generative manipulation policies expensive. The authors target semantic dynamics, constant-size historical memory, and lower blocking latency; their stronger claims of physical causal reasoning are motivations rather than separately established causal-identification results. e02e03

2.2 What this reading follows

A robot can start computing its next movement before its current movement ends if it can anticipate the next observation. DexWorldModel builds this idea around DINOv3 features: predict a semantic future, then infer the actions that reach it. A shared transformer backbone performs both stages, while separate persistent and working memories keep real experience apart from imagined context. EmbodiChain supplies synthetic adaptation trajectories through domain expansion and online streaming. Read the architecture alongside the data ablations: the paper reports strong execution success, but those experiments do not separately establish every architectural explanation offered for it. e04e05e07e09e10e13e15e16e17

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 categoryNot assigned
ArchitectureNot assigned
Prediction paradigmNot assigned
QuadrantNot assigned

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

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

The supplied catalog fields are unassigned, so no existing quadrant can be confirmed or contradicted. Architecture evidence supports a candidate One Model assessment: video and action branches share core transformer blocks. The prediction paradigm is sequential future-feature prediction followed by inverse-dynamics-style action generation, rather than simultaneous joint sampling or a training-only auxiliary world loss. e03e04e05

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
  • Language instruction; current and historical visual observations encoded by frozen DINOv3
  • Historical executed actions summarized with visual features in TTT memory
  • Predicted future semantic feature maps
  • Action chunks of 16 steps; the standardized dual-arm action representation has 30 continuous dimensions

4.2 Equations and their role

ϕvid=ϕvidout∘ϕshare∘ϕvidin,ϕact=ϕactout∘ϕshare∘ϕactin\phi_{\mathrm{vid}}=\phi_{\mathrm{vid}}^{\mathrm{out}}\circ\phi_{\mathrm{share}}\circ\phi_{\mathrm{vid}}^{\mathrm{in}},\qquad \phi_{\mathrm{act}}=\phi_{\mathrm{act}}^{\mathrm{out}}\circ\phi_{\mathrm{share}}\circ\phi_{\mathrm{act}}^{\mathrm{in}}
Equation (7): video and action mappings use their own input/output projections around the same transformer core. Shared parameters establish architectural integration despite sequential generation. e04
x(s)=(1−s)ϵ+sx,LCFM=Es,ϵ,x,c[∥vϕ(x(s),s∣c)−(x−ϵ)∥2]x^{(s)}=(1-s)\epsilon+sx,\qquad \mathcal{L}_{\mathrm{CFM}}=\mathbb{E}_{s,\epsilon,x,c}\left[\left\|v_{\phi}(x^{(s)},s\mid c)-(x-\epsilon)\right\|^2\right]
Equations (4)–(5): s is flow time, epsilon is Gaussian noise, x the data target, c the conditioning context, and v the learned velocity. Thus s=0 is noise and s=1 is data. Equations (8) and (10) specialize this objective to future features and actions. e05
Wtlong=Wt−1long−η∇Wℓself(Wt−1long;ht),Wtwork←Wtlong\mathcal{W}^{\mathrm{long}}_t=\mathcal{W}^{\mathrm{long}}_{t-1}-\eta\nabla_{\mathcal{W}}\ell_{\mathrm{self}}(\mathcal{W}^{\mathrm{long}}_{t-1};h_t),\qquad \mathcal{W}^{\mathrm{work}}_t\leftarrow\mathcal{W}^{\mathrm{long}}_t
Equations (14)–(15): W denotes memory weights, eta the test-time learning rate, and h_t=[f_t,a_{t-1}] the new observed feature and executed action. The self-supervised reconstruction loss updates persistent history; cloning creates transient context. Equation (16) updates only the clone with a predicted feature. e07

5. Method in detail

5.1 A predicted feature becomes an action condition

Source description

At an observation step, CLWM encodes the camera image with frozen DINOv3 and uses language plus historical context to generate the next semantic feature. This is a learned future under the task condition; it is not a measurement returned by the environment. A second flow process generates the action chunk conditioned on that predicted feature. The two processes share transformer blocks but have different projections and flow-time embeddings. This makes the feature forecast operationally important: errors can influence the action condition. Training regresses the velocity of a linear noise-to-data interpolation for each modality and perturbs historical features to expose the action model to imperfect context. Equation (9) specifies the historical perturbation; the separately written augmented future-feature condition in Equation (10) is less explicitly defined. e04e05e06

Figure 2. Persistent memory receives observed experience; a temporary copy receives predicted experience. Original paper, p. 6 ↗

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

How to read it. Read panels (a) and (b) as a change in how historical context is retained. Panel (b) still contains an attention block; the proposed replacement concerns the accumulating KV cache. In panel (c), observed features and actions feed long-term memory, and the dashed fork arrow creates working memory. A generated feature updates that working branch before action generation. Equations (14)–(16) on page 7 specify observed history as the new feature together with the previously executed action, then a clone and predicted-feature update. Thus the diagram's coarse time labels should not be read as permission to put an unexecuted action into persistent history. e07e08e09e14

What it supports. The intended boundary is between observed and speculative information. Both memories have fixed-size weights, which removes growth in stored historical tokens as an episode lengthens. The source reports a flat memory footprint over a 2,000-step episode, but this figure itself depicts the design rather than a measured memory curve.

Where the evidence stops. The figure and caption mark completed samples s = 0, conflicting with the equations and SAI's s = 1 endpoint. Constant storage also does not demonstrate faithful recall over arbitrary horizons; the reported efficiency test does not measure forgetting.

5.2 Speculate in a copy, then return to observed history

Reader analysis

The useful distinction in TTT is which information is allowed to persist. The long-term weights are updated from an actual observation and executed action; the working copy can absorb a predicted future without making it permanent. SAI extends this separation across computation and physical time. While a chunk executes, predicted features stand in for the unavailable next observation and support partial denoising. Once sensing arrives, calibrated context guides the remaining integration. The source's s = 0 completion labels conflict with its noise-to-data equations, so reproducing the exact handoff needs clarification. As reader analysis, the expected speed advantage depends on both overlap duration and forecast accuracy: a cheap handoff can still produce a poor action when the speculative state becomes wrong. The paper gives no matched latency–accuracy sweep. e07e08e09e14

5.3 Account for fresh trajectories, not only optimizer steps

Reader analysis

EmbodiChain first makes simulation experience more varied through assets, reachable interactions, recovery sequences, and temporally coherent appearance changes. ODS then controls how quickly that experience replaces previously consumed trajectories. These are different mechanisms: Table 2 changes the data-generation recipe at a fixed trajectory count, whereas Table 3 changes maximum replay at a fixed optimizer budget. The latter comparison strongly favors lower reuse, but the paper does not report the total unique trajectories or generation compute for each row. My interpretation is that fresh experience helps this system, while online delivery itself remains insufficiently isolated. The method still applies supervised losses to generated demonstrations; its analogy to online and offline reinforcement learning does not introduce a specified reward-optimization algorithm. A useful reproduction must log both data diversity and consumption cost. e10e11e15e16

5.4 Training and inference

During training

Source description

Pretraining combines RoboMind, Agibot World Beta, and InternData-A1. AdamW uses learning rate 1e-4, global batch 128, and about 20 epochs. The paper reports 64 NVIDIA H100 GPUs and approximately 20 days for the entire training. Dataset mixture sizes and detailed preprocessing rules are absent. e12

Source description

Video and action flow losses regress interpolation velocities. Historical features are noise-mixed with probability 0.5 and augmentation interpolation values in [0.5,1]. Task adaptation uses synthetic data; RoboTwin comparisons fine-tune all models on 25,000 trajectories for 40,000 iterations at learning rate 1e-5. e05e06e12e13

During inference

Reader analysis

SAI uses predicted features as surrogate observations to integrate from noise at s=0 to an unspecified intermediate s_mid during action execution. When true sensing returns, it updates long-term memory, replaces speculative conditioning, and finishes integration to s=1. This reduces waiting only when useful precomputation overlaps physical motion. e09

Reader analysis

Figures 1–2 and Section 3.2 instead label a completed prediction and working-memory update s=0. This contradicts the interpolation equations and Section 3.3. The verified conceptual order is prediction, working-memory update, then action generation; the implementation convention remains unresolved. e08

5.5 Implementation flow

  1. Encode and share

    DINOv3 base extracts patch features with patch size 16. Video and action branches share Wan2.2-5B-initialized transformer blocks; flow-time embeddings and input/output projections are modality-specific. Figure 1 also marks the language model as frozen without identifying it. e04e12

  2. Predict before acting

    The video branch denoises a future feature conditioned on history and language. The action branch then conditions on that future to generate the corresponding motor chunk: an inverse-dynamics-style factorization used during inference. Physical execution and the next observation close the loop. e03e05e09

  3. Maintain two histories

    A residual TTT-MLP reconstructs projected tokens. Its key/value/query projections train in the outer loop; memory weights update online. New observations and executed actions update long-term weights. A cloned working memory incorporates the predicted future between video and action generation, while remaining fixed within integration stages. e07

  4. Generate adaptation experience

    EmbodiChain refines assets and feasible layouts, samples diverse reachable interactions, relabels recovery motions, and varies appearance with temporal consistency. Asynchronous workers feed a shared circular buffer with bounded reuse to a supervised learner. e10e11

6. Experiments & results

DexWorldModel introduces CLWM, which predicts future DINOv3 features and then generates actions conditioned on that prediction. Shared transformer blocks, separate persistent and speculative memories, and asynchronous denoising connect world prediction to robot execution. EmbodiChain supplies synthetic adaptation data. Reported manipulation results are strong, but protocol omissions and a contradictory flow-time convention limit reproducibility.

6.1 Read the original evidence

Table 1. CLWM leads the reported average, with substantial task-level variation. Original paper, p. 13 ↗

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

How to read it. Start with the bottom Average row, then compare individual task rows rather than treating the aggregate as uniform reliability. Ours is CLWM, and every cell is a success percentage. Page 14 states that the comparison fine-tunes all models on 25,000 synthetic trajectories for 40,000 iterations at learning rate 1e-5. Hanging Mug and Turn Switch are useful counterweights to the many near-ceiling rows: they remain difficult for CLWM. The full table is retained so readers can inspect both wins and exceptions. Its boldface is source formatting, not evidence of statistical significance; in Turn Switch, Motus's numerical score exceeds Ours despite Ours being bold. e13e15e16

What it supports. The reported averages are 94.00% for CLWM and 91.55% for LingBot-VA, a difference of 2.45 percentage points. CLWM nevertheless scores only 40% on Hanging Mug and 65% on Turn Switch; Motus scores 78% on the latter. The evidence supports a higher reported aggregate rather than universal task superiority.

Where the evidence stops. The paper does not specify rollout counts, exact evaluation splits, or uncertainty for this table. The separate three-task ablations use different data budgets, so their success rates should not be substituted into this benchmark comparison.

Table 4. Reported physical execution succeeds without target-task real-data adaptation. Original paper, p. 15 ↗

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

How to read it. Read across the CLWM row in the task order shown: water pouring, table rearrangement, hand-over and place, then pan open and place. Read vertically against Sim2Real-VLA first, because the source assigns those two methods identical synthetic budgets and the same EmbodiChain domain-expansion and streaming pipeline. The π0 and GR00T N1.5 rows instead use 50 real expert demonstrations per task, so the whole table does not share one data protocol. Also read Section 5.1 on page 12 before interpreting zero-shot: it describes public robot-data pretraining. The defensible distinction is simulation-only target adaptation, rather than an absence of robot-data pretraining. e12e17

What it supports. CLWM reports 95%, 90%, 80%, and 65% success, compared with Sim2Real-VLA's 80%, 80%, 40%, and 35%. The taskwise differences are 15, 10, 40, and 30 percentage points. These values concern execution on the Agilex CobotMagic platform, providing a different kind of evidence from generated feature visualizations.

Where the evidence stops. The caption claims success fractions with standard errors, yet the actual cells contain only percentages. Trial counts and the numerical synthetic budget are absent. Its blanket simulation-only caption also requires the training-protocol qualifications in the surrounding text.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
RoboTwin bimanual manipulation

Table 1; common fine-tuning budget of 25,000 synthetic trajectories and 40,000 iterations. Exact test split, rollout counts, and seeds are unstated.

94.00%

Reported average task success (%)

LingBot-VA 91.55%; Motus 87.02%; π0.5 76.76%; X-VLA 72.84%.

The reported average exceeds LingBot-VA by 2.45 percentage points. Success is uneven: CLWM reaches only 40% on Hanging Mug and 65% on Turn Switch, versus Motus at 78% on Turn Switch. e13

Long-horizon deployment efficiency

Section 5.3.2 describes a 2,000-step memory comparison and blocking-latency profiling against a sequential pipeline.

Authors report constant O(1) memory and approximately 50% less blocking latency.

Historical-memory scaling and blocking-latency reduction

KV-cache history scales as O(T); sequential inference waits for the next real observation.

These are prose-reported findings, without memory measurements, absolute milliseconds, profiler configuration, or numerical curves. Fixed memory capacity does not establish unlimited reliable recall. e14e07

EmbodiChain domain expansion

Hanging Mug, Turn Switch, Stack Bowls; 2,000 synthetic trajectories per task. OOD changes objects, textures, lighting, and layouts.

Spatial baseline 64 / 25; +visual augmentation 75 / 42; +physics 81 / 56; +reachability 95 / 82.

Average ID / OOD success (%)

Full pipeline improves OOD success by 57 percentage points over spatial randomization.

Cumulative additions support the full combination under this budget, without isolating interactions or proving representation-level causality. e15

Online data streaming replay bound

Same three tasks; 5,000 iterations, batch 64. Static baseline has 1,500 demonstrations; streamed trajectories are evicted after 213, 50, or 10 uses.

Static 62 / 85 / 88; ODS213 60 / 84 / 85; ODS50 92 / 92 / 96; ODS10 96 / 98 / 98.

Hanging Mug / Turn Switch / Stack Bowls success (%)

ODS10 gains 34 / 13 / 10 percentage points over static training.

Lower replay limits help here, but also change the amount of unique data consumed. Equal optimizer iterations do not establish equal generation cost or a universal Efficiency Law. e16e11

Four-task physical manipulation on CobotMagic

Agilex CobotMagic; CLWM and Sim2Real-VLA use equal, numerically unspecified synthetic adaptation budgets. π0 and GR00T N1.5 receive 50 real demonstrations per task.

CLWM 95 / 90 / 80 / 65.

Water pouring / table rearrangement / hand-over and place / pan open and place success (%)

Sim2Real-VLA 80 / 80 / 40 / 35; π0 25 / 20 / 20 / 5; GR00T N1.5 35 / 20 / 15 / 5.

This is reported physical execution, not video plausibility. Zero-shot transfer means no target-task real-data adaptation; public robot pretraining remains. Table 4 promises standard errors but displays none, and supplies no trial counts. e17e12

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 2. Success improves as visual, physical, and reachable-motion diversity are added. Original paper, p. 14 ↗

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

How to read it. Read downward as cumulative additions, beginning with object-position and orientation randomization. The next rows add visual augmentation, physics-grounded generation, and reachability-aware sampling. Compare ID with OOD within each row, then follow each column down. The surrounding protocol fixes 2,000 synthesized trajectories per task across Hanging Mug, Turn Switch, and Stack Bowls. OOD covers unseen objects, novel textures, extreme lighting, and unfamiliar layouts. Section 4 explains the last addition in task-space terms: feasible interactions are selected for diversity in approach direction, contact geometry, and outcomes. The table summarizes three tasks and does not show separate scores for each. e10e15

What it supports. The full pipeline reports 95% ID and 82% OOD success, compared with 64% and 25% for spatial randomization alone. The final reachability addition raises the preceding configuration from 81% to 95% ID and from 56% to 82% OOD. These are sizeable improvements under the stated fixed trajectory budget.

Where the evidence stops. This cumulative design does not independently remove each module or estimate interactions. Aggregate OOD success also does not separate sensitivity to lighting, object identity, and layout, and the table supplies no evaluation counts or error bars.

Table 3. Fresh-data turnover matters more here than merely using a streaming buffer. Original paper, p. 15 ↗

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

How to read it. Each ODS subscript is the maximum number of times a trajectory may be sampled before eviction, not a dataset size or task count. All rows use 5,000 optimizer iterations with batch size 64. The static baseline repeatedly samples 1,500 demonstrations, averaging approximately 213 uses per trajectory under the paper's accounting. ODS213 therefore tests streaming with slow turnover; it performs similarly to the static row. Read down each task column as the cap falls to 50 and then 10. Section 4.3 describes asynchronous generation and bounded reuse, but the table does not report how many distinct trajectories were actually generated or consumed. e11e16

What it supports. At the cap of 10, Hanging Mug rises from the static baseline's 62% to 96%, while Turn Switch rises from 85% to 98% and Stack Bowls from 88% to 98%. The high-reuse streaming condition does not improve over static training. The supported pattern concerns turnover under this optimization budget.

Where the evidence stops. More turnover also means access to more unique data. Without an offline comparison matched for unique trajectories and generation compute, the table cannot isolate online timing from dataset diversity or establish the schematic Efficiency Law as a measured scaling relation.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

No isolated DINOv3-versus-VAE, shared-backbone, TTT-quality, or SAI-accuracy ablation is provided. Data-pipeline ablations cannot establish which architectural component causes the main success gains. The claimed history-augmentation stability guarantee lacks a proof or measured perturbation analysis. e04e09e14e15e16

Reader analysis

The real-world section describes training as exclusively simulated, although Section 5.1 describes public robot pretraining. Preserve the narrower adaptation claim. Missing evaluation definitions, splits, counts, and uncertainty prevent statistical assessment of the reported improvements. e12e13e15e16e17

7.2 Questions for discussion

  1. How much action information is lost when DINOv3 features replace reconstruction-oriented latents?
  2. Does SAI retain accuracy when an unexpected contact invalidates its speculative observation?
  3. Would the ODS advantage persist with equal unique data and total generation compute?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Required components include the stated pretrained encoders/backbone, the stated action representation, simulation generator, and bimanual interface. TTT uses a two-layer MLP with 4× expansion, GELU, and gate initialization 0.1. Missing settings include inner-loop learning rate, projection ranks, language-model identity, loss weighting, ODE step count, s_mid, camera/frame configuration, buffer capacity, and actual unique-trajectory throughput. e04e07e09e11e12e17

Reader analysis

Reader-proposed checks should first compare sequential versus speculative inference under matched solver budgets and controlled perturbations, then compare streaming with offline data at matched unique-trajectory budgets. Both target unseparated mechanisms rather than attempting the entire reported training run. e09e14e16

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: Measure SAI's waiting-time and accuracy tradeoff

Reader-proposed check: start from one fixed CLWM checkpoint and compare sequential inference with several SAI split points using the same initial noises, total solver evaluations, task episodes, and hardware. Introduce a controlled object displacement during action execution, with an unperturbed control. Record success, blocking latency, total computation, and speculative-versus-observed feature error. Over a 2,000-step run, also log memory footprint and failures. The useful-speedup claim would be weakened if waiting falls only by accepting lower success, or if calibration cannot recover after displacement. This check was not run. e07e09e14

Check 2: Separate streaming delivery from access to more unique data

Reader-proposed check: reproduce the three-task, 5,000-iteration, batch-64 comparison while logging unique trajectory IDs and generation cost. Compare ODS10, the 1,500-demo static baseline, and an offline pool containing the same trajectories eventually consumed by ODS10. Match task mixture, augmentations, learner updates, and replay counts where possible; repeat with independent seeds and held-out evaluation episodes. If the matched offline pool reaches ODS10 performance, the gain is explained by fresh-data access without requiring online delivery. Report success uncertainty and total wall time rather than inferring a universal scaling law. This check was not run. e10e11e15e16

8.3 Reading coverage

Visual audit: The title/author/version page and every method, training, evaluation, and compute page supporting retained claims were rendered and visually inspected. This includes Figures 1–7, Tables 1–4, the SAI diagram on page 8, compute details on page 12, and the protocol paragraphs around the cropped tables. All six final original crops were individually viewed. Figure 1–2 flow-time labels were checked against Equations (4)–(10), the memory-update text, and Section 3.3; the inconsistency is disclosed. Table 4's missing standard errors and mixed training protocols are preserved as limitations. References on pages 16–19 were read as text; reference-only pages 17–19 were not visually inspected. No separate supplement, execution video, code, or reproduced experiment was inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16. 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 and abstract (p. 1)
  • 1 Introduction (pp. 1–2)
  • 2.1 Vision-Language-Action Models; 2.2 World Action Models; 2.3 Conditional Flow Matching (pp. 3–4)
  • 3.1 Autoregressive Generation of Latent Video Features and Actions (pp. 5–6)
  • 3.2 Constant-Memory Autoregression via Test-Time Training (pp. 6–7)
  • 3.3 Speculative Asynchronous Inference (pp. 7–8)
  • 4.1 Generative Simulation for Robot Learning Environment (pp. 9–10)
  • 4.2 Data Scaling via Domain Expansion; 4.3 Efficient Scaling via Online Data Streaming (pp. 10–11)
  • 5.1 Dataset Curation; 5.2 Implementation & Training Details (p. 12)
  • 5.3.1 Simulation Results on RoboTwin; 5.3.2 Efficiency Analysis (pp. 12–14)
  • 5.3.3 EmbodiChain Experiments; 5.3.4 Real-world Deployment (pp. 14–15)
  • 6 Conclusion (p. 16)
  • References (pp. 16–19)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • Identity: the inspected title and all three authors match the catalog. The supplied PDF is explicitly arXiv:2604.16484v1 [cs.CV], 13 April 2026; no revision or edition difference is observed in the supplied material, and no other version was inspected.
  • The complete supplied text was read. Text extraction does not reconstruct figure images; this limitation was addressed by inspecting original PDF pages 1–16 and all six final crops.
  • Separate supplemental material availability has not been fully verified; no separate supplement was supplied. No appendix is present in the supplied 19-page PDF.
  • Code, external links, and linked references were not inspected; no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title block and arXiv marginInspect

Exact title: DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks. Authors: Yueci Deng, Guiliang Liu, Kui Jia. Affiliation: DexForce AI. Version: arXiv:2604.16484v1 [cs.CV], 13 Apr 2026.

Go to primary source ↓
e02PDF pp. 1–2, Abstract and Section 1Inspect

The authors motivate semantic rather than pixel prediction, constant-size historical memory, and speculative inference to address representation, cache, and latency costs.

Go to primary source ↓
e03PDF p. 3, Section 2.2, Eqs. (2)–(3); pp. 4–5, Figure 1 caption and Section 3.1 stagesInspect

The source describes predicting a future visual state followed by inverse dynamics; CLWM instantiates this sequence with semantic features and an action flow model.

Go to primary source ↓
e04PDF p. 4, Figure 1; p. 5, Section 3.1, Eqs. (6)–(7)Inspect

Figure 1 marks DINOv3 and the language model with freezing symbols. DINOv3 base uses patch size 16. Video/action branches share Wan2.2-5B-initialized transformer blocks, with separate timestep embeddings and linear projections.

Go to primary source ↓
e05PDF pp. 3–4, Section 2.3, Eqs. (4)–(5); pp. 5–6, Section 3.1, Eqs. (8)–(10) and action interpolationInspect

Flow interpolation maps Gaussian noise at s=0 to data at s=1, with velocity regression losses for features and actions. The action model conditions on future semantics and generates chunks of size 16.

Go to primary source ↓
e06PDF p. 5, Eq. (9) and adjoining history-augmentation text; p. 12, Section 5.2Inspect

Historical features are interpolated with Gaussian noise with probability 0.5, using s_aug in [0.5,1]. The action loss additionally writes an augmented future-feature condition without an equally explicit separate future-feature sampling rule.

Go to primary source ↓
e07PDF pp. 6–7, Figure 2 and Section 3.2, Eqs. (11)–(16)Inspect

TTT reconstructs projected tokens using residual two-layer MLP weights. Projections train in the outer loop; weights update online. Long-term memory uses observed features and executed actions; its clone receives predicted features. The MLP uses GELU and 4× expansion; the gate starts at 0.1.

Go to primary source ↓
e08PDF p. 4, Figure 1 update label and Eq. (5) interpolation; p. 5, Eq. (8); pp. 6–7, Figure 2 and Section 3.2 intermediate update; p. 8, Section 3.3 Phases 1–2Inspect

Figures 1–2 and the intermediate-update text identify a completed feature with s=0, while the stated interpolation begins at noise at s=0 and the SAI text completes at s=1. The source does not reconcile these conventions.

Go to primary source ↓
e09PDF p. 8, Figure 3 and Section 3.3, Phases 1–2 and Theoretical Synergy with History AugmentationInspect

SAI pre-denoises using predicted context while actions execute, swaps to observed context, and finishes from s_mid to 1. The split is only constrained between 0 and 1. The source asserts stability from augmentation without supplying a proof on this page.

Go to primary source ↓
e10PDF pp. 9–11, Sections 4.1–4.2, Figures 5–7Inspect

EmbodiChain refines physical assets and layouts, samples reachable task-space diversity, reintegrates corrective motions, and applies temporally consistent visual variation while preserving physical parameter consistency.

Go to primary source ↓
e11PDF p. 9, Figure 4; p. 11, Section 4.3Inspect

The Efficiency Law figure is schematic. ODS is described as asynchronous generation into a shared CPU/GPU circular buffer, zero-copy consumption, bounded reuse, and supervised optimization. No empirical threshold or throughput curve is supplied.

Go to primary source ↓
e12PDF p. 12, Sections 5.1.1–5.2Inspect

Pretraining datasets are RoboMind, Agibot World Beta, and InternData-A1; action vectors combine 7 end-effector pose, 7 joint, and 1 gripper entries per arm. Post-training is synthetic. AdamW uses 1e-4, batch 128, about 20 epochs; training uses 64 H100 GPUs for about 20 days.

Go to primary source ↓
e13PDF p. 13, Table 1, Average, Hanging Mug, and Turn Switch rows; p. 14, Training SpecificationsInspect

Average success: π0.5 76.76%, X-VLA 72.84%, Motus 87.02%, LingBot-VA 91.55%, CLWM 94.00%. CLWM Hanging Mug is 40%; Turn Switch is 65% versus Motus 78%. Fine-tuning uses 25,000 synthetic trajectories, 40k iterations, and 1e-5 learning rate.

Go to primary source ↓
e14PDF p. 14, Section 5.3.2; p. 2, Section 1 SAI paragraphInspect

The efficiency section reports flat memory over 2,000 steps and about 50% lower blocking latency, without numeric memory/time tables or curves. The introduction names RoboTwin and LingBot-VA; the efficiency section itself describes a sequential baseline more generally.

Go to primary source ↓
e15PDF p. 14, Section 5.3.3 Domain Expansion and Table 2, all rowsInspect

Three tasks use 2,000 trajectories each. ID/OOD success increases cumulatively from 64/25 to 75/42, 81/56, and 95/82 percent. OOD includes unseen objects, textures, lighting, and layouts.

Go to primary source ↓
e16PDF pp. 14–15, Section 5.3.3 Online Data Streaming; p. 15, Table 3, all rowsInspect

The static 1,500-demo baseline and ODS variants train for 5,000 iterations at batch 64. Replay caps are 213, 50, 10. In Hanging Mug/Turn Switch/Stack Bowls order, success is 62/85/88, 60/84/85, 92/92/96, and 96/98/98 percent.

Go to primary source ↓
e17PDF p. 15, Section 5.3.4, Training Specifications and Table 4, all rows and caption; p. 12, Section 5.1Inspect

CobotMagic results in task-column order are CLWM 95/90/80/65, Sim2Real-VLA 80/80/40/35, π0 25/20/20/5, GR00T N1.5 35/20/15/5 percent. Synthetic budgets match between CLWM and Sim2Real-VLA; other baselines receive 50 real demonstrations per task. The caption promises standard errors but cells contain only percentages; Section 5.1 documents public robot pretraining.

Go to primary source ↓

8.5 Primary sources

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