PAPER REPORTENAll readings ↗

Valerant: An Automatic Navigable Game Map Generator via Action-Conditioned World Model Exploration

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

Authors: Yiran Qiao; Feng Wang; Jing Ma

Affiliations: Case Western Reserve University; Johns Hopkins University

Source: 2609.09418 ↗ · Catalog record

Reading: 18 / 558 · 4 original figures & tables · ~20 min ·

1. Paper overview

In one sentence: Valerant turns a frozen action-conditioned video model into a 3D map-building search procedure, with geometric checks compensating for generated views that can drift or pass through walls. e02e07e08e09e10e11e12e13e14e15

At a glanceWhat to know
Research problem
Source description

A video generator can show plausible movement without producing reusable 3D geometry. For game-map authoring, the environment itself must be constructed: unlike physical robot exploration, no independent world supplies corrective observations. Valerant therefore couples generated visual futures to explicit reconstruction and action selection, aiming to create persistent space from a single visual concept. e02e04

Core mechanism
Source description

The framework connects a frozen action-conditioned world model, visual SLAM and an exploration policy entirely at inference time, converting selected predicted observations into a persistent point cloud. e03e07

A key reported resultSingle-image game-map generation: human preference: Valerant: 83.33%

User Preference (UP), percentage of pairwise comparisons preferring each method; higher is better. VALORANT screenshots; the same initial frames for Valerant and HY-World 2.0. Pairwise human comparisons; image count, evaluator count and split are not specified.

HY-World 2.0: 16.67% Reported perceptual preference favors Valerant. Missing comparison counts and uncertainty prevent a significance assessment; UP is not a collision-free navigation rate. e12e13

Reading caution
Reader analysis

Figure 2 shows four generated trajectories and reconstructed maps, not ground-truth reconstruction error or independent navigation trials. The source provides no quantitative coverage, collision-rate or downstream game-engine test. Generalization beyond the evaluated game's screenshots remains untested here. e11e12e13e15

Core contributions

  • Source description

    The framework connects a frozen action-conditioned world model, visual SLAM and an exploration policy entirely at inference time, converting selected predicted observations into a persistent point cloud. e03e07

  • Source description

    Robust floor estimation, complementary geometric/depth collision tests and checkpoint rewind address drift, wall penetration and local exploration traps. The paper supports their roles with qualitative examples rather than aggregate ablation statistics. e08e09e10e15

Figure 1. Reconstruct candidate futures before deciding which observations enter the persistent map. Original paper, p. 3 ↗

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

How to read it. Follow the upper row from the initial image through candidate videos, spatial reconstruction and evaluation. Although the first block is labeled WAM, Sections 3.1–3.2 identify its video generator as a frozen WM and the overall prediction–selection coupling as the WAM. The dashed return leads to another exploration step; the text defines termination as exhaustion of the budget. In the lower panel, inspect the separate left, straight and right reconstructions and their red dashed paths. The green check highlights the straight branch. The formal method restores the common checkpoint before replaying that action, so unchosen geometry does not enter the persistent map. The upper output icon is schematic; the lower output depicts reconstructed geometry. e03e04e05e06e07

What it supports. The central mechanism is geometric evaluation before commitment. The policy sees reconstructed motion and scene structure for each generated alternative, then retains one branch. This links visual prediction to action selection and incremental mapping without training a new policy or an additional learned utility predictor. The loop is a modular inference procedure.

Where the evidence stops. The picture shows three branches and CF values of −100/0, while Equations (5)–(6) specify five actions and a hard binary collision constraint. The score-to-constraint correspondence is unexplained. Follow the equations for the verified selection rule; the displayed scores are schematic, not measured results.

2. Motivation

2.1 The problem and the proposed response

Source description

A video generator can show plausible movement without producing reusable 3D geometry. For game-map authoring, the environment itself must be constructed: unlike physical robot exploration, no independent world supplies corrective observations. Valerant therefore couples generated visual futures to explicit reconstruction and action selection, aiming to create persistent space from a single visual concept. e02e04

2.2 What this reading follows

Imagine starting with a screenshot and asking which movement would reveal useful new space. Valerant generates several possible video continuations, reconstructs each in 3D, and selects a branch before adding it to a persistent map. Its contribution lies in this inference loop: a pretrained video world model, SLAM, geometric rejection and backtracking work together without additional training. The visual evidence shows reconstructed routes and examples of three exploration failures. A perceptual comparison favors Valerant over HY-World 2.0, but the source leaves evaluation counts and major implementation settings unspecified. Read the figures as evidence for a map-authoring prototype whose navigation reliability still needs direct measurement. e02e07e08e09e10e11e12e13e14e15

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 Related resources. View the current classification.

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

The catalog is entirely unassigned, so there is no substantive classification to confirm or contradict. Architecture evidence supports a cascaded, modular system: an action-conditioned WM predicts video, SLAM reconstructs it, and an external policy selects candidates. It is neither one learned model jointly predicting futures/actions nor inverse-dynamics action extraction. This is a reader assessment, leaving the recorded taxonomy unchanged. e04e05e07

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
  • Single initial image; experiments use VALORANT screenshots
  • Visual history and five candidate action sequences at each exploration step
  • Committed generated observations and estimated camera trajectory
  • Persistent 3D point-cloud game map; optional global point-cloud refinement

4.2 Equations and their role

Otfpθ(Otc,At)\mathbf O_t^f \sim p_\theta(\cdot \mid \mathbf O_t^c,\mathbf A_t)
Equation (1): the WM with parameters θ samples future visual observations Oᶠ from context Oᶜ and supplied action sequence A at step t. The paper defines context length K and prediction horizon H. Conditioning on an action predicts its consequences; it does not select that action. e05
m=arg maxm{1,,5}:Bt,m=0(λcovCt,m+λprogPt,m)m^\star=\operatorname*{arg\,max}_{m\in\{1,\ldots,5\}:\mathcal B_{t,m}=0}\left(\lambda_{\mathrm{cov}}C_{t,m}+\lambda_{\mathrm{prog}}P_{t,m}\right)
Equation (6): m indexes candidate branches, C is exploration coverage, P is forward progress, and the two λ weights balance them. The binary collision indicator B must be zero. This is constrained selection; numerical weights and operational coverage/progress definitions are not supplied. e07
hˉi(p)=ni(pqi)ni(ciqi)\bar h_i(\mathbf p)=\frac{\mathbf n_i^\top(\mathbf p-\mathbf q_i)}{\mathbf n_i^\top(\mathbf c_i-\mathbf q_i)}
Equation (8): for keyframe i, q lies on the floor, n is its unit normal toward camera center c, and p is a reconstructed point. Floor height becomes zero and camera height one. Common translation and scale cancel; incorrect floor estimation itself is not corrected by this normalization. e08

5. Method in detail

5.1 Turn one visual history into several geometric choices

Source description

At a decision step, Valerant has a visual history, a camera pose and a map. It checkpoints the video model and SLAM state before evaluating five prescribed controls. For each control, the frozen model generates a video continuation from the same history. SLAM reconstructs that continuation, producing geometry and motion that the exploration policy can inspect. Collision screening removes inadmissible candidates; weighted coverage and forward progress rank the rest. The selected action is then replayed from the restored checkpoint, and only its observations and geometry update the persistent state. This explains why the model's ability to predict an action's visual consequences is only one part of the system. The selection rule supplies the decision, while reconstruction turns committed predictions into the map being authored. e04e05e06e07e09

5.2 Measure obstacles relative to the floor the camera currently sees

Reader analysis

A fixed global floor can become misleading when monocular reconstruction drifts: ground points may migrate into the height band used to identify an obstacle. Valerant instead estimates a supporting plane locally and expresses point height as a fraction of camera-to-floor distance. Equation (8) makes the floor zero and the camera one. Under a common translation or scale change, numerator and denominator change consistently, so this relative height remains stable. The collision test then restricts attention to a body-height band and normalized horizontal radius. A second test checks raw central depth together with continued forward motion, compensating for surfaces filtered out of the point cloud. As reader analysis, this normalization protects against a coordinate inconsistency; it cannot by itself repair an incorrectly fitted floor, missing wall geometry or contradictory generated views. e08e09

5.3 Backtrack the generated world while remembering the failed route

Source description

Even collision-aware local selection can enter a pocket where all translational choices are blocked or make negligible progress. The two in-place rotations may then keep producing new orientations without useful expansion. Valerant stores synchronized checkpoints along the committed route so that it can return to an earlier visual context, camera pose and map. It also retains the cells traversed during the discarded segment as failed-region memory. Equation (11) restores the earlier joint state while enlarging that memory, and subsequent candidates entering those cells become inadmissible. An earlier checkpoint can be tried again if the restored state remains trapped. Figure 3's Split example illustrates the intended outcome: escape from repeated turning toward another route. The paper gives this recovery logic, but leaves the numerical progress trigger, cell resolution and checkpoint-selection settings unspecified. e07e10e14e15e17

5.4 Training and inference

During training

Source description

Valerant introduces no additional training. Section 3.1 describes the underlying video generator using flow matching: interpolate a clean video latent with Gaussian noise and learn a conditional velocity toward the noise-minus-clean target. At inference the flow runs from noise toward a video latent, followed by decoding. These are pretrained-model preliminaries, not a new Valerant optimization stage; the video WM stays frozen. e05e07

During inference

Reader analysis

Planning, geometric rejection and map fusion are inference-time operations. 'Execution' means replaying the chosen action within the generated environment and committing its observations, not controlling a physical robot or validating the map in an independent game engine. Optional video2world post-processing globally refines the point cloud. e04e07e14

5.5 Implementation flow

  1. Checkpoint and branch

    Maintain visual context, camera pose and global map. Checkpoint the WM and SLAM states, then test uw (forward), jw/lw (left/right turn followed by forward movement), and jq/lq (in-place left/right rotation) from the same pre-action context. e07

  2. Generate and reconstruct

    Matrix-Game 3.0 generates each candidate video. MASt3R-SLAM estimates camera motion and branch geometry in a common coordinate system. This explicitly reconstructs every prospective future; no additional learned utility predictor is specified. e05e06e07e14

  3. Establish a local floor

    Fit a plane to lower-envelope support points for each keyframe, accepting sufficient support and a normal consistent with the vertical direction. If the floor is unobservable, propagate the latest reliable camera-to-floor relation. Normalize heights relative to that local floor and camera distance. e08

  4. Reject implausible translations

    The point-cloud test counts body-height points inside a normalized cylinder along sampled poses; its occupancy threshold suppresses isolated noise. Rotation-only candidates are exempt from this test. A complementary detector uses unfiltered depth from all rollout frames: a nearby central surface plus continued forward displacement indicates penetration. Either detector activates the collision flag. e09

  5. Select and commit

    Among collision-free candidates, maximize weighted exploration coverage and forward progress. Restore the pre-action state, replay the selected action, update visual context and fuse only its geometry. Discard other branches. Repeat until the exploration budget is exhausted. e07

  6. Recover from a dead pocket

    Persistent insufficient progress triggers restoration of an earlier joint WM–SLAM checkpoint. Record cells along the discarded trajectory in failed-region memory and reject future candidates entering them. Rewind farther if necessary; restoring context and map together keeps their histories synchronized. e10

6. Experiments & results

Valerant expands one game screenshot into a persistent 3D point-cloud map by generating alternative action-conditioned videos, reconstructing each with SLAM, and committing the best admissible branch. A frozen video model supplies futures; an external exploration policy selects actions. Perceptual comparison and qualitative ablations support this prototype, while geometric reliability and reproducibility remain incompletely measured.

6.1 Read the original evidence

Figure 2. Generated observations become explicit point-cloud routes in four illustrated scenes. Original paper, p. 8 ↗

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

How to read it. Read each scene group horizontally. The first frame is the conditioning screenshot, the middle block contains six sampled generated frames, and the right column shows the reconstructed map from above. The red dashed overlay denotes the agent's trajectory, according to the caption. Lotus and Pearl illustrate routes through outdoor architectural spaces; Fracture shows an indoor passage, while Abyss shows a broader open area. Compare the changing views with the shape of the corresponding point cloud to understand what is being accumulated. These are four examples of the same procedure, not stages of one shared map. Section 4.2 describes the trajectories as generated automatically without manual path specification or intermediate intervention. e07e11e12e16

What it supports. The examples support the claim that the system produces an explicit spatial artifact alongside its video observations. The top-down maps and trajectory overlays make the explored routes visible. They also help distinguish the paper's output from a stream of frames: selected observations have been consolidated into persistent reconstructed geometry.

Where the evidence stops. These static examples do not establish continuous collision-free motion, metric reconstruction accuracy or complete level coverage. No reference geometry or scale bars are supplied here. The conclusion leaves richer editable assets and improved geometric consistency to future work; the images show point clouds.

Table 1. Reported perceptual scores favor Valerant on the shared screenshot inputs. Original paper, p. 9 ↗

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

How to read it. Read the two methods by row and keep the two percentage columns separate. UP means User Preference: the proportion of pairwise comparisons in which human evaluators prefer a method's output. AS has a source naming inconsistency: the table caption calls it aesthetic score, while the nearby metric paragraph calls it Authenticity Score and describes a VLM evaluator. Preserve AS as the reported label rather than assuming a standardized metric. Section 4.1 states that both methods receive the same initial VALORANT screenshots. Its open-ended generation setting has no unique target map, which motivates perceptual evaluation but does not make these columns tests of geometric validity. e12e13e14

What it supports. Valerant reports 83.33% UP against 16.67% for HY-World 2.0, and 80.00% AS against 20.00%. Both reported comparisons favor Valerant. The strongest numerical statement supported here is preference or perceptual scoring under the paper's evaluation, not a navigation-success percentage, reconstruction-accuracy gain or statistically established population effect.

Where the evidence stops. The source omits screenshot and evaluator counts, uncertainty, the VLM identity and prompt, and AS aggregation details. It also does not specify matched compute or refinement settings for the comparison. These omissions limit reproducibility and prevent a significance assessment.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Single-image game-map generation: human preference

VALORANT screenshots; the same initial frames for Valerant and HY-World 2.0. Pairwise human comparisons; image count, evaluator count and split are not specified.

Valerant: 83.33%

User Preference (UP), percentage of pairwise comparisons preferring each method; higher is better

HY-World 2.0: 16.67%

Reported perceptual preference favors Valerant. Missing comparison counts and uncertainty prevent a significance assessment; UP is not a collision-free navigation rate. e12e13

Single-image game-map generation: AS evaluation

Same screenshot comparison; Section 4.1 describes a VLM evaluator but does not identify its model, prompt or scoring aggregation.

Valerant: 80.00%

AS (%), higher is better; Table 1 calls it aesthetic score, while the metric paragraph calls it Authenticity Score

HY-World 2.0: 20.00%

Preserve the AS label and naming discrepancy. The reported percentage is not a geometric accuracy measure or a demonstrated VLM win rate under a fully specified protocol. e12e13

Robust exploration component ablation

Figure 3 paired generated-frame examples: Heaven with/without RF, Icebox with/without COL, and Split with/without DP.

Authors describe improved forward traversal with RF, rejection of wall penetration with COL, and escape from repeated turning with DP.

Qualitative trajectory behavior; no numerical ablation metric

Corresponding module removed in the upper row of each scene pair; enabled in the lower row.

Examples illustrate complementary failure modes. They do not quantify average gains, failure rates, module interactions or sensitivity to sampling seeds. e15

6.3 Ablations and diagnostic examples

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

Figure 3. Three qualitative comparisons connect robustness modules to distinct exploration failures. Original paper, p. 9 ↗

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

How to read it. Start with the row labels on the right: w/o omits the named module, while w enables it. Compare vertically within each scene before following frames horizontally. In Heaven, as spelled in the source, RF means robust floor estimation; the lower sequence advances into the open area. In Icebox, COL means complementary collision detection; the upper sequence approaches and passes the orange wall, whereas the lower sequence follows free space. In Split, DP means dead-pocket rewind; repeated similar wall views above contrast with a different route below. The caption and Section 4.2 attribute these changes to floor-relative reasoning, rejection of infeasible candidates and restoration of an earlier exploration state, respectively. e08e09e10e15

What it supports. The ablation examples motivate three different interventions. RF addresses false obstacles associated with floor drift, COL screens apparent wall penetration, and DP provides recovery when local choices stop making progress. Their roles are complementary in the proposed loop, and the qualitative examples are consistent with the paper's stated mechanisms.

Where the evidence stops. Each module is illustrated in a different scene, with no aggregate failure-rate table, uncertainty or documented matching of random seeds and budgets. The examples therefore do not estimate additive gains or establish which component contributes most. Selected frames also do not supply full trajectory-level collision verification.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Figure 2 shows four generated trajectories and reconstructed maps, not ground-truth reconstruction error or independent navigation trials. The source provides no quantitative coverage, collision-rate or downstream game-engine test. Generalization beyond the evaluated game's screenshots remains untested here. e11e12e13e15

Reader analysis

The authors leave improved geometric consistency and conversion of point clouds into richer, editable game assets to future work. A point-cloud output should therefore not be equated with a finished, production-ready game level. e16

Reader analysis

Figure 1 depicts three moving branches and a CF score of −100 or 0, whereas Equations (5)–(6) specify five candidates and a hard binary collision constraint. The schematic does not establish a soft-penalty implementation or measured utility values; their exact correspondence is unexplained. e03e07

7.2 Questions for discussion

  1. How should geometric consistency and traversability be evaluated when the target map is intentionally open-ended?
  2. Can a selected stochastic rollout be replayed faithfully enough that its geometric collision verdict remains valid?
  3. Does failed-region memory suppress useful revisits when monocular reconstruction changes spatial scale or alignment?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

The reported stack is Matrix-Game 3.0 plus MASt3R-SLAM, optionally followed by video2world, on one NVIDIA H100 GPU. Exact software/checkpoint revisions, per-map runtime, peak memory and which reported outputs received refinement are unspecified. e14

Reader analysis

Reimplementation needs action durations, context/horizon lengths, exploration budget, utility definitions/weights, floor acceptance settings, collision thresholds, trajectory sampling and rewind/cell parameters. The paper gives symbolic tests but no numerical configuration; its hyperparameter appendix reference is unresolved and no appendix occurs in the supplied PDF. e05e07e08e09e10e14e17

Reader analysis

Proposed checks: compare fixed versus adaptive floor reasoning on identical reconstructed rollouts under controlled scale/translation drift; separately test whether checkpoint replay preserves the selected branch's geometry and collision decision when random state is restored. These would test normalization and commit reliability without claiming to reproduce the perceptual table. e05e07e08e09

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: Does floor normalization preserve collision decisions under drift?

Reader-proposed check, not performed: hold stored branch point clouds and camera poses fixed, then apply known positive uniform scales and vertical translations. First transform the accepted floor plane with the camera and points to test Equation (8) directly; normalized heights and point-cloud collision flags should remain unchanged. Then compare adaptive floor fitting with a floor calibrated only once, using the same transformed inputs and manually identified ground/obstacle regions. Log false obstacle flags and rejected forward candidates. A failure of the direct invariance test would expose an implementation inconsistency; a failure after refitting would locate the problem in estimation. Declare chosen thresholds because the source omits their values. e08e09e14e17

Check 2: Does checkpoint replay preserve the candidate that was evaluated?

Reader-proposed check, not performed: from one initial screenshot and checkpoint, generate the five candidate rollouts and record the selected branch's frames, reconstructed trajectory, utility and collision flags. Restore the joint WM–SLAM checkpoint, including random state where available, then replay the selected action with identical settings and optional refinement disabled. Repeat with deliberately fresh generator noise as a control. Compare the replayed observations, geometric verdict and utility using preregistered tolerances across several checkpoints. Changed admissibility after restoration would show that the screened candidate is not reliably the committed outcome. Report random-state handling explicitly: the source defines a stochastic WM and replay procedure but does not specify their determinism contract. e05e07e09e14

8.3 Reading coverage

Visual audit: All 14 supplied PDF pages were rendered and actually inspected, including the title/author/version block, complete method and Equations (1)–(11), experiment settings, all three figures, Table 1, conclusion and final references. All five text chunks were read individually in full. Figure 1 was checked against the five-action set and hard collision constraint; its schematic CF values are disclosed rather than converted into an implementation rule. Table 1 values and arrows were checked against the adjacent metric paragraph, preserving the AS naming inconsistency. The unresolved appendix reference and absence of an appendix were checked through the final page. All four final original crops were viewed after extraction; the narrow table was rendered at 450 DPI and recropped to exclude a caption fragment. No external works, code, checkpoint files or continuous supplementary videos were inspected, and no experiments were run.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14. 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, author block, arXiv version stamp and abstract (p. 1)
  • 1 Introduction (pp. 1–3)
  • 2 Related Work, including 2.1–2.3 (pp. 3–4)
  • 3 Method: 3.1 Preliminaries, 3.2 Overview, 3.3.1 Robust Floor Estimation, 3.3.2 Complementary Collision Detection and 3.3.3 Dead-Pocket Rewind (pp. 5–8)
  • 4 Experiments: settings, main results and ablation study (pp. 8–10)
  • 5 Conclusion and References (pp. 10–14)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Version scope: the supplied title page identifies arXiv:2609.09418v1 [cs.AI], 8 September 2026. Title and all three authors match the catalog; no revision or author discrepancy was observed. No other edition was supplied or compared.
  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This reading additionally inspected all 14 PDF pages and all final figure/table crops.
  • Separate supplemental material availability has not been fully verified.
  • The supplied PDF contains no appendix. Its implementation paragraph refers to the unresolved placeholder 'Appendix ??' for detailed hyperparameters.
  • Code, model checkpoints, external cited works and supplementary videos were not inspected. No experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title/byline, affiliations and arXiv margin stampInspect

Title matches the supplied observed title. Authors are Yiran Qiao, Feng Wang and Jing Ma; Qiao and Ma list Case Western Reserve University, Wang lists Johns Hopkins University. The stamp reads arXiv:2609.09418v1 [cs.AI], 8 Sep 2026.

Go to primary source ↓
e02PDF pp. 1–2, Abstract and Section 1, game-creation motivationInspect

The proposed task constructs persistent 3D game geometry from a single image by combining action-conditioned visual prediction, exploration and SLAM; the generated world lacks an independently existing physical substrate.

Go to primary source ↓
e03PDF p. 3, Figure 1, upper loop and lower branch-score panel/captionInspect

Candidate videos feed spatial reconstruction and evaluation/commit, with a budget-controlled loop and point-cloud output. The lower illustration shows left/straight/right branches and FP/EC/CF columns, including CF −100 for the left branch and 0 for the others; straight is selected. These are schematic values.

Go to primary source ↓
e04PDF p. 4, Sections 2.1–2.3, Valerant-specific paragraphsInspect

The paper explicitly calls Valerant a cascaded WAM: a frozen action-conditioned video model supplies observations, an exploration policy selects actions, and SLAM constructs explicit geometry. Observations arise from the model rather than sensors in an existing environment.

Go to primary source ↓
e05PDF p. 5, Section 3.1, Equations (1)–(3) and adjacent training/inference descriptionInspect

The conditional WM distribution uses visual history and a supplied action sequence, with context K and horizon H. Flow-matching preliminaries interpolate video latents and Gaussian noise, target noise minus clean latent, and integrate toward the latent at inference. Valerant keeps the video WM frozen.

Go to primary source ↓
e06PDF p. 5, Section 3.1, Visual SLAM, Equation (4)Inspect

SLAM estimates camera poses and a spatial map by visual/geometric consistency; each generated rollout is reconstructed and registered in a common 3D coordinate system.

Go to primary source ↓
e07PDF pp. 5–6, Section 3.2, Equations (5)–(7), candidate definitions and commit paragraphInspect

Five candidate sequences branch from checkpointed WM/SLAM states: forward, left/right then forward, and left/right in-place rotations. Equation (6) maximizes coverage/progress only where collision flag B is zero. Restore, replay and commit one branch; update context and fuse its map until the budget ends. Utility weights and numerical settings are not given.

Go to primary source ↓
e08PDF pp. 6–7, Section 3.3.1, floor fitting/propagation and Equation (8)Inspect

A locally supported, vertically consistent floor plane replaces fixed calibration; unreliable observations propagate the last camera-floor relation. Point height is divided by camera-floor distance so the floor is zero and camera one, canceling common translation/scale changes.

Go to primary source ↓
e09PDF p. 7, Section 3.3.2, Equations (9)–(10) and detector-combination paragraphInspect

Point-cloud collision requires enough points in a normalized cylindrical body band along sampled poses; rotation-only candidates are exempt from that test. Raw central depth across all frames complements filtered geometry by detecting a near surface followed by sufficient forward movement. Either positive detector sets the collision flag.

Go to primary source ↓
e10PDF pp. 7–8, Section 3.3.3, Equation (11) and rewind descriptionInspect

Persistent low progress triggers synchronized restoration of earlier WM/SLAM state. Cells along the discarded segment accumulate in failed-region memory and become inadmissible to subsequent candidates; earlier checkpoints can be tried repeatedly.

Go to primary source ↓
e11PDF p. 8, Figure 2/caption; pp. 9–10, Section 4.2, Main ResultsInspect

Four examples labeled Lotus, Pearl, Fracture and Abyss pair initial images, six sampled generated frames each, and final top-down point clouds with overlaid trajectories. Authors describe automatic exploration without manual trajectories/intervention. No geometric accuracy or coverage statistic accompanies these images.

Go to primary source ↓
e12PDF p. 9, Section 4.1, Dataset & Baseline and Metric paragraphsInspect

Inputs are VALORANT screenshots shared across both methods. HY-World 2.0 is the baseline. No unique reference map exists. AS is described here as VLM-assessed Authenticity Score; UP is the fraction of human pairwise preferences. Dataset/evaluator counts, split, VLM identity/prompt and aggregation are not specified.

Go to primary source ↓
e13PDF p. 9, Table 1, both rows and AS/UP columns, table caption versus Metric paragraphInspect

HY-World 2.0 reports AS 20.00% and UP 16.67%; Valerant reports AS 80.00% and UP 83.33%. Both arrows point upward. The caption expands AS as aesthetic score, conflicting with Authenticity Score in prose. No uncertainty or sample counts appear in the table.

Go to primary source ↓
e14PDF p. 9, Section 4.1, Implementation DetailsInspect

The stack is Matrix-Game 3.0, MASt3R-SLAM and optional video2world point-cloud refinement; all experiments use one NVIDIA H100. Detailed hyperparameters are deferred to the unresolved 'Appendix ??'. Runtime, peak memory, exact releases and refinement use per reported result are not supplied.

Go to primary source ↓
e15PDF p. 9, Figure 3, scene labels, paired row labels and caption; p. 10, Section 4.2, Ablation StudyInspect

Heaven contrasts without/with robust floor estimation (RF), Icebox without/with complementary collision detection (COL), and Split without/with dead-pocket rewind (DP). Upper rows omit the module; lower rows enable it. Prose attributes the changes to better progress, prevented wall penetration and escape from repeated turning, respectively. No numerical ablation table is supplied.

Go to primary source ↓
e16PDF p. 10, Section 5, final sentenceInspect

Future work targets improved geometric consistency and conversion of reconstructed point clouds into richer, editable game assets.

Go to primary source ↓
e17PDF p. 9, Implementation Details; pp. 10–14, Conclusion followed by References through the final pageInspect

The implementation text points to 'Appendix ??', but the supplied PDF proceeds from conclusion to references and ends on p. 14 without an appendix or hyperparameter table.

Go to primary source ↓

8.5 Primary sources

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