WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN

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

Authors: Yuehao Huang; Yunzi Wu; Xiaotao Zhang; Xinhai Li; Jiankun Dong; Jiajun Lv; Chi Zhang; Chenjia Bai; Yong Liu; Xuelong Li

Affiliations: Institute of Artificial Intelligence, China Telecom; Zhejiang University; Tongji University; Shanghai Jiao Tong University

Source: 2608.07267 ↗ · Project page ↗ · Catalog record

Reading: 64 / 558 · 5 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: A geometry-aware history prefix improves joint video–action navigation on GN-Bench, while the curriculum shows that expert correction must precede effective reward refinement in the tested configurations. e-problem e-architecture e-inference e-results e-ablation

At a glance	What to know
Research problem	Author claim Continuous vision-language navigation changes the agent's visual evidence after every action. The authors argue that action-only supervision leaves observation evolution unconstrained, while appearance-based world-action histories lack persistent geometric context. Their proposed solution makes history-derived scene features a shared condition for future-view and action generation under partial observability. e-problem e-architecture
Core mechanism	Source description A frozen VGGT-Ω encoder and trainable adapter inject RGB-derived scene context throughout joint video–action denoising, including at inference. e-architecture e-adapter
A key reported result	GN-Bench Unseen closed-loop navigation: WNM-3D: SR 53.5%, SPL 46.6%; NE 4.1, OS 62.2%, TL 10.2. SR / SPL; NE, OS and TL. Official Unseen split, 5,000 episodes; same full-curriculum WNM evaluation protocol and seed. WNM-2D: SR 45.9%, SPL 42.8%. FPV-only GN-BAE: 38.9% / 37.3%. WNM-3D gains 7.6 SR and 3.8 SPL points over WNM-2D. Improved Unseen scores do not establish a smaller Seen-to-Unseen gap. e-results e-protocol
Reading caution	Reader analysis Evaluation is in GN-Bench's 3DGS simulation. The paper supplies point estimates with no multi-seed variation or bootstrap intervals, and does not establish physical-robot transfer or deployment latency. It explicitly notes that geometry conditioning does not reduce the Seen-to-Unseen gap. e-problem e-results e-protocol

Core contributions

Source description
A frozen VGGT-Ω encoder and trainable adapter inject RGB-derived scene context throughout joint video–action denoising, including at inference. e-architecture e-adapter
Source description
A three-stage curriculum combines expert demonstrations, corrective supervision at policy-visited states, and reward refinement with localized counterfactual denoising credit. e-curriculum e-counterfactual
Reader analysis
The WNM-2D comparison and training-stage ablation test the complete conditioning design and curriculum, rather than isolating each adapter component. e-results e-ablation e-configuration

Three-panel architecture: alternative RGB-history encoders produce a prefix for a shared video–action DiT; modality decoders and a block-causal visibility matrix appear to the right. — **Figure 1.** History-derived scene tokens condition the same transformer that generates visual and action variables. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

2. Motivation

2.1 The problem and the proposed response

Author claim

Continuous vision-language navigation changes the agent's visual evidence after every action. The authors argue that action-only supervision leaves observation evolution unconstrained, while appearance-based world-action histories lack persistent geometric context. Their proposed solution makes history-derived scene features a shared condition for future-view and action generation under partial observability. e-problem e-architecture

2.2 What this reading follows

WNM-3D asks how a navigation policy can use the geometry implicit in its past RGB views while predicting both motion and what that motion will reveal. A frozen geometry encoder and a learned adapter turn the observed history into scene tokens that remain available throughout joint video–action denoising. The paper then separates two training problems: recovering from states reached by an imperfect policy, and making successful behavior more efficient. Its strongest evidence comes from the WNM-2D comparison and the training-stage table. Read these alongside the inference rule: only actions execute, and only newly observed images enter the next history. e-problem e-architecture e-inference e-results e-ablation

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 dimension	Recorded classification
Major category	WAMs
Architecture	One Model
Prediction paradigm	Joint prediction
Quadrant	Q1 · One Model × Joint prediction

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The recorded One Model × Joint prediction category is supported by a single shared DiT with interacting visual and action tokens and modality-specific heads. Actions are jointly generated rather than recovered by a separate inverse-dynamics controller. The frozen geometry encoder supplies inference-time context, so its presence does not turn the joint generator into a separate world-model/planner pipeline. e-architecture e-flow e-inference

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

Inputs	Outputs
Language instruction and current monocular egocentric RGB frame 33 RGB observations sampled uniformly from the observed trajectory; no policy input of depth, pose, BEV or metric map	Four aligned blocks totaling 32 future RGB-frame targets represented as visual latents and 32 continuous actions Physical action increments (Δx, Δy, Δψ); only the first eight actions are eligible for execution before replanning

4.2 Equations and their role

C_t^{\mathrm{3D}}=T_\phi(E_{\mathrm{geo}}(\mathcal{H}_t))\in\mathbb{R}^{N_c\times d}

Equation (3): history H_t passes through frozen geometry encoder E_geo and trainable adapter T_phi. C_t is the clean scene prefix, N_c its token count, and d the DiT width; the implementation uses 450 tokens of width 3072. e-architecture e-configuration

x_{\sigma_b}^{m,b}=(1-\sigma_b)x_{\mathrm{data}}^{m,b}+\sigma_b\epsilon^{m,b},\qquad u_{\sigma_b}^{m,b}=\epsilon^{m,b}-x_{\mathrm{data}}^{m,b}

Equation (5): modality m is visual v or action a, b indexes temporal blocks, sigma_b is the flow timestep, x_data the target, and epsilon independent Gaussian noise. The DiT learns velocity u along this path using timestep-weighted squared error; visual and action terms have equal coefficients. e-flow

5. Method in detail

5.1 Turn multiview history into shared conditioning

Reader analysis

At a replanning step, the agent has several past views but no depth or pose input to the policy. WNM-3D samples 33 RGB observations across that history and runs the frozen geometry encoder on resized versions. The adapter is responsible for translating its cross-view feature lattice into the DiT's conditioning format. Its coarse path preserves pooled scene content, while its detail path retrieves local features around structured anchors and refines them spatially and temporally. The resulting prefix contains 450 tokens, matching WNM-2D's interface. The reader interpretation is that this makes geometric evidence accessible to both modalities without forcing the downstream generator to consume the full upstream feature grid. The paper's comparison tests that complete conditioning design; it does not separate geometric information from added representation capacity or input-resolution effects. e-adapter e-configuration e-results

Three stacked adapter stages show gated geometry features, pooled and structured target queries with local deformable sampling, and spatial–temporal refinement with detail and coarse heads. — **Figure 2.** The adapter compresses a history feature lattice into a fixed scene-token interface. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

5.2 Keep prediction, execution and observation separate

Source description

The joint generator predicts four aligned blocks, each representing eight future frames and eight actions. During supervised learning, a visual block and its action block share a noise level, but draw independent noise. Their tokens interact inside one temporal block while remaining unable to attend to later blocks. At deployment, sampling produces both streams, but only the first action block is eligible for execution. Action decoding restores the three physical increments from checkpoint quantiles. XY increments share the local frame fixed at replanning; accumulated yaw does not rotate subsequent XY increments again. STOP checks may terminate before an action is applied. After execution, real simulator observations update history and the prefix is recomputed. Thus the policy uses a jointly sampled visual representation without treating its imagined frames as new sensory evidence. e-flow e-architecture e-configuration e-actions e-inference

Training schematic progressing from A* supervised joint prediction, through policy rollout and expert corrective labeling, to counterfactual branches with separate visual, navigation and stopping rewards. — **Figure 3.** Expert imitation, corrective supervision and reward refinement address different training conditions. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

5.3 Distinguish coverage correction from reward credit

Reader analysis

DAgger first collects states actually reached by the imperfect Stage-I policy and supplies expert-corrected futures there. Counterfactual DanceGRPO then asks a different question: which local denoising decisions produce better completed samples? For each context, four noise strata each supply a four-branch group. Branches share initial latents and all external noise except at one intervention. Separate reward streams compare completed videos and action chunks with fixed references; advantages are standardized within the corresponding group and gradients replay only the intervention. Visual rewards combine image and temporal quality with motion consistency, while navigation and stopping rewards shape progress, collision avoidance and settling. The reader interpretation of Table 2 is that correction establishes a useful starting distribution for refinement. That explanation remains provisional because branch diversity and reward spread were not directly measured. e-curriculum e-counterfactual e-stage3-data e-visual-reward e-nav-reward e-stop-reward e-ablation

5.4 Training and inference

During training

Source description

Stage I trains for 20 epochs on 16K A* demonstrations with trajectory-matched future images and actions. Stage II rolls out each Stage-I policy on the same tasks and queries A* corrections, producing approximately 633K WNM-3D versus 691K WNM-2D chunks for five supervised epochs. Both stages minimize weighted visual/action velocity error with padded action dimensions masked. e-curriculum e-supervised e-flow

Source description

Stage III uses about 96K fixed records per configuration, recollected from Stage II for the full curriculum. It does not step the interactive simulator or add a supervised loss. Four noise strata each generate four branches sharing all randomness except one intervened transition; terminal rewards are evaluated, but gradients replay only that transition. e-stage3-data e-counterfactual

Source description

Visual rewards combine reconstruction, temporal quality and optical-flow/action consistency. Navigation rewards score progress, route adherence, collision and orientation; stopping rewards favor settling near the goal. Advantages are standardized per context and stratum. Visual credit uses visual transition ratios, while navigation/stopping credit uses action ratios, all updating the shared DiT. e-visual-reward e-nav-reward e-stop-reward e-counterfactual

During inference

Source description

Recompute history tokens from actual observations, jointly sample visual/action variables from Gaussian noise, decode actions, and execute at most the first block. Predicted visual latents are neither decoded to RGB nor appended to history. STOP precedes execution if the block's XY path length is strictly below 0.15 m; otherwise an individual action with every component at most 10⁻³ in absolute value stops before that action. e-inference e-actions e-protocol

5.5 Implementation flow

Build the history prefix
VGGT-Ω features from four encoder levels are projected and fused with per-location gates. Pooled scene content initializes structured queries; anchored deformable resampling retrieves local details. Spatial then temporal attention refines them, and detail/coarse heads produce 450 tokens. The resampling lattice is history–height–width, not an explicit metric voxel map. e-adapter e-configuration
Couple the predicted modalities
The clean prefix enters DiT self-attention; current-frame and language encodings enter cross-attention. Each visual/action pair shares a flow timestep but independent noise. The mask allows bidirectional interaction within a temporal block and access to earlier blocks, while masking later blocks. Separate modality heads share the transformer. e-architecture e-flow
Represent executable motion
XY increments use a fixed local frame anchored at the current camera pose. Yaw accumulates separately and does not recursively rotate XY increments. Three physical dimensions are scaled by four, quantile-normalized and padded to width 32; decoding uses checkpoint-associated 1% and 99% quantiles. e-actions

6. Experiments & results

WNM-3D conditions a shared video–action diffusion transformer on geometry-aware tokens extracted from RGB history. It jointly predicts future visual latents and continuous navigation actions, then executes only the first action block before observing again. On GN-Bench, the full curriculum improves both Seen and Unseen success over WNM-2D. The evidence supports the combined geometry-conditioned design and staged training, while leaving component attribution, physical deployment and uncertainty unresolved.

6.1 Read the original evidence

Full GN-Bench comparison table with Depth, BEV and FPV observation flags, Seen and Unseen metrics, and highlighted WNM-2D and WNM-3D rows; the dagger definition is retained. — **Table 1.** Geometry-conditioned WNM-3D improves success and SPL on both benchmark splits. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

6.2 Results and evaluation conditions

Task & protocol	Reported result	Comparison & interpretation
GN-Bench Seen closed-loop navigation Official Seen split, 1,000 episodes; full three-stage WNM checkpoints, monocular RGB; one evaluation per episode, inference seed 1140.	WNM-3D: SR 88.7%, SPL 80.9%; NE 1.9, OS 94.2%, TL 10.6. SR / SPL; NE, OS and TL	WNM-2D: SR 75.6%, SPL 72.9%. GN-BAE with BEV+FPV: 58.6% / 58.6%. Gains over WNM-2D are 13.1 SR and 8.0 SPL percentage points. The prior-method comparison differs in observation space and training. e-results e-protocol
GN-Bench Unseen closed-loop navigation Official Unseen split, 5,000 episodes; same full-curriculum WNM evaluation protocol and seed.	WNM-3D: SR 53.5%, SPL 46.6%; NE 4.1, OS 62.2%, TL 10.2. SR / SPL; NE, OS and TL	WNM-2D: SR 45.9%, SPL 42.8%. FPV-only GN-BAE: 38.9% / 37.3%. WNM-3D gains 7.6 SR and 3.8 SPL points over WNM-2D. Improved Unseen scores do not establish a smaller Seen-to-Unseen gap. e-results e-protocol
WNM-3D training curriculum ablation Same GN-Bench episodes and receding-horizon evaluation across Stage I, Stage II, full Stage III and direct Stage-I-to-DanceGRPO configurations.	Unseen Stage I→II→III: SR 43.1→49.7→53.5%; SPL 41.6→38.2→46.6%; TL 5.9→13.2→10.2. Seen SR: 56.7→80.7→88.7%. SR / SPL / TL on Unseen; SR on Seen	Skipping DAgger and applying DanceGRPO directly gives Seen/Unseen SR 47.1%/39.2%, below Stage I. DAgger's success gains can accompany inefficient paths; subsequent reward refinement improves both. The direct-refinement failure motivates, but does not prove, the authors' hypothesis about weak within-group behavioral diversity. e-ablation e-stage3-data e-protocol

6.3 Ablations and diagnostic examples

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

Training-stage checkmarks and Seen/Unseen TL, NE, OS, SR and SPL for four WNM-3D configurations followed by four WNM-2D configurations. — **Table 2.** DAgger raises success, but reward refinement is needed to recover path efficiency in the Unseen comparison. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Evaluation is in GN-Bench's 3DGS simulation. The paper supplies point estimates with no multi-seed variation or bootstrap intervals, and does not establish physical-robot transfer or deployment latency. It explicitly notes that geometry conditioning does not reduce the Seen-to-Unseen gap. e-problem e-results e-protocol

Reader analysis

WNM-2D controls the backbone and prefix length, but WNM-3D adds a geometry encoder and 23.97M-parameter adapter, uses higher-resolution geometry inputs, and has different policy-specific training data and batch size. No component-level adapter ablation or quantitative video-quality evaluation isolates why navigation improves. e-configuration e-supervised e-stage3-data e-results e-ablation

Author claim

The proposed explanation for direct DanceGRPO failure remains an author hypothesis: action diversity and reward spread within counterfactual groups were not measured. e-ablation

7.2 Questions for discussion

Does geometry conditioning retain its gain when both variants train on identical corrective records?
Which reward stream changes stopping and path efficiency after DAgger?
Does the visual latent rollout improve executed navigation beyond what the history prefix alone provides?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Required resources include Wan2.2-TI2V-5B initialization, VGGT-Ω, GN-Bench environments and A* annotations. The backbone operates at 160×320 and geometry inputs at 512×512. Stages I/II use 16 H100 GPUs, BF16, AdamW at 10⁻⁵, and WNM-3D global batch 192. The geometry encoder, VAE and image/language encoders are frozen. e-architecture e-configuration e-supervised

Reader analysis

Stage III specifies 1,500 AdamW updates at 5×10⁻⁶, condition batch 64, 16 sampler transitions and CFG scale 5. Navigation chunk weights are (8,4,2,1)/15; stopping chunks are averaged. Visual flow calibration uses 480 ground-truth clips. Exact quantiles, trajectory-smoothing parameters, ridge regularization and software versions are not supplied as reproducible values. e-stage3-settings e-visual-reward e-actions e-supervised

Reader analysis

Reader-proposed checks: match training records and update budgets when replacing the scene prefix, then measure paired navigation outcomes; separately measure reward spread and branch diversity before and after DAgger under identical counterfactual noise. These would test conditioning attribution and the proposed explanation of refinement failure. e-configuration e-supervised e-counterfactual e-ablation

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 the scene prefix help under matched corrective data?

Reader-proposed experiment, not performed: train WNM-2D and WNM-3D from their common backbone initialization on the same Stage-I examples and a fixed pooled set of corrective records, with equal optimizer-update counts and effective batch size. Preserve the 450-token prefix and common action decoder, STOP rules and evaluation episodes. First compare after supervised adaptation, before policy-specific Stage-III recollection adds another data difference. Report paired episode SR/SPL differences, inference variation, memory and latency. Then test history-prefix shuffling at evaluation as a diagnostic for dependence on scene-specific information, recognizing that shuffling creates distribution shift. Persistence of the navigation gain under matched training would support conditioning attribution; disappearance would weaken the claim that prefix representation alone explains the original comparison. e-configuration e-supervised e-actions e-results e-protocol

Check 2: Does DAgger create more informative counterfactual groups?

Reader-proposed experiment, not performed: sample the Stage-I and Stage-II checkpoints on an identical held-out subset of annotated navigation contexts. Use the same four strata, four branches, shared initial latents and matched intervention noise. Before any update, measure within-group action-path diversity, each reward stream's standard deviation, tie frequency and valid-branch fraction. Then run a matched-budget refinement comparison while holding record selection fixed, evaluating with the same STOP rule. The authors' explanation predicts that correction should yield more feasible and informative groups and improved subsequent refinement. Similar reward spread and diversity despite opposite refinement outcomes would weaken that mechanism and motivate examination of reward calibration or optimization instead. Keep this diagnostic separate from the paper's full curriculum, which recollects Stage-III records using Stage II. e-ablation e-counterfactual e-stage3-data e-visual-reward e-nav-reward e-stop-reward e-stage3-settings e-protocol

8.3 Reading coverage

Visual audit: Visually inspected the title/author/version page, the motivation page, every method and results page, and all implementation/evaluation appendix pages. Figures 1–4 and Tables 1–8 were read in the original PDF render. All five final crops were separately viewed: Figures 1–3 and Tables 1–2. Table 1 retains its observation and dagger legend; the other crops exclude long captions and body prose. The narrow adapter crop was rendered at 350 DPI and the numerical tables at 250 DPI. Page 3 and reference pages 12–13 were read as text but were not part of the visual pass. No separately supplied supplement was available.

PDF pages inspected for this edition: 1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 14, 15, 16, 17, 18, 19, 20, 21, 22. Appendix coverage: reviewed.

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, authors, abstract and version stamp (p. 1)
1 Introduction (pp. 1–2)
2 Related Work (p. 3)
3 Method, including 3.1–3.5 (pp. 4–8)
4 Experiments, including 4.1–4.3 (pp. 8–11)
5 Conclusion (p. 11)
References (pp. 11–13)
A.1 Architecture and Input Configuration (pp. 14–15)
A.2 Action Encoding, Decoding, and Execution (p. 15)
A.3 Stage-I and Stage-II Training (pp. 15–16)
A.4 Counterfactual DanceGRPO, including A.4.1–A.4.6 (pp. 16–22)
B.1 GN-Bench Closed-Loop Evaluation (p. 22)

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 ten authors match the catalog. This report reads arXiv:2608.07267v2, stamped 20 August 2026; the catalog lists submission on 19 August 2026. No earlier revision was supplied, so changes from v1 cannot be established.
All seven supplied text chunks were read individually. Text extraction alone omits figure images; the retained PDF was additionally rendered and inspected, including all four figures and all eight tables.
Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
Code, project pages and model weights were not inspected; experiments were not reproduced. The source's links establish stated provenance only.
The reference list was read, but its cited works were not independently reviewed.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title, author/affiliation block and left-margin arXiv stampInspect

Exact catalog title and ten authors appear; the PDF identifies arXiv:2608.07267v2, 20 Aug 2026, with four affiliations.

Go to primary source ↓

e-problemPDF pp. 1–2, Abstract and Section 1Inspect

The authors motivate shared history-derived geometric context for continuous VLN and describe evaluation in GN-Bench's 3D Gaussian Splatting environments.

Go to primary source ↓

e-architecturePDF pp. 4–5, Figure 1, Sections 3.1–3.2, Eqs. (3)–(4)Inspect

Frozen geometry/history conditioning and current-frame/language context feed a shared world-action DiT. Visual/action tokens interact within temporal blocks under a block-causal mask; conditioning stays clean. Figure 1 marks frozen and trained modules.

Go to primary source ↓

e-adapterPDF p. 6, Section 3.4, Figure 2, Eqs. (7)–(9)Inspect

The adapter gates projected encoder levels, initializes queries from pooled content plus learned structure, resamples around history–height–width anchors, and applies spatial/temporal refinement with detail/coarse residual output.

Go to primary source ↓

e-flowPDF p. 5, Section 3.3, Eqs. (5)–(6); p. 16, A.3, Eqs. (16)–(17)Inspect

Joint flow matching uses linear data-to-noise paths and modality velocity targets; paired blocks share timesteps, noise is independent, weights have unit mean on a shifted 1,000-index schedule, padded action residuals are masked, and the action coefficient is one.

Go to primary source ↓

e-curriculumPDF p. 7, Figure 3 and Section 3.5Inspect

A* supervised initialization is followed by policy rollout and expert correction with trajectory-consistent views, then Counterfactual DanceGRPO without simulator stepping or an auxiliary supervised objective.

Go to primary source ↓

e-counterfactualPDF p. 8, Section 3.5, Eqs. (10)–(12); p. 17, A.4.1, Eqs. (18)–(20)Inspect

Four strata {0,1,2}, {6}, {10}, {13,14,15} each yield four branches with noise changed only at the intervention. Rewards are standardized within groups; ties yield zero advantage. Gradients replay the intervened transition and reward streams use modality-specific ratios.

Go to primary source ↓

e-inferencePDF p. 8, Section 3.5, Receding-Horizon InferenceInspect

History is uniformly resampled from actual observations and the prefix recomputed. Joint Gaussian-initialized sampling yields visual/action rollouts; only the first action block executes. Predicted visual latents are not decoded or added to history.

Go to primary source ↓

e-configurationPDF pp. 14–15, A.1 and Table 3Inspect

Wan2.2-TI2V-5B initialization; 33 history frames; 160×320 backbone and 512×512 geometry inputs; four blocks of eight frames/actions. VGGT-Ω blocks 5/12/18/24 feed width-512 memory; a 9×5×10 lattice yields 450 tokens of width 3072, matching WNM-2D. The adapter has 23.97M parameters. No depth, pose, BEV or metric map is a policy input.

Go to primary source ↓

e-actionsPDF p. 4, Section 3.1, Eqs. (1)–(2); p. 15, A.2, Eqs. (13)–(15) and ordered STOP ruleInspect

Smoothed trajectories define fixed-frame XY and tangent-heading increments. Actions scale by four, use checkpoint 1%/99% quantiles and pad to 32 dimensions. XY is rotated by the replanning-frame rotation only; yaw accumulates separately. STOP uses strict block path length <0.15 m, then componentwise ≤10⁻³. Occupancy handling distinguishes raw plans from execution.

Go to primary source ↓

e-supervisedPDF pp. 15–16, A.3 and Table 4Inspect

16K A* trajectories support 20 Stage-I epochs. Policy-specific DAgger supplies about 691K/633K chunks for WNM-2D/3D and five epochs. AdamW starts at 10⁻⁵ with 5% warmup; 16 H100 GPUs use BF16/TF32 and global batches 256/192.

Go to primary source ↓

e-stage3-dataPDF pp. 16–17, A.4 opening paragraphsInspect

About 96K fixed records per configuration include reference views, expert trajectories, occupancy and metadata. Direct refinement reuses the Stage-I-policy collection underlying DAgger; full refinement recollects with Stage II. No interactive simulator stepping or supervised world-action loss occurs during optimization.

Go to primary source ↓

e-visual-rewardPDF pp. 17–19, A.4.2, Eqs. (21)–(30), Table 5Inspect

Visual reward combines arithmetic pyramid SSIM, Charbonnier reconstruction/temporal error and degradation gating with a flow/action bonus. DIS optical-flow descriptors are mapped to motion by a fixed ridge regressor fit on 480 ground-truth clips; inferred yaw is diagnostic only. This is not standard multiplicative MS-SSIM.

Go to primary source ↓

e-nav-rewardPDF pp. 19–20, A.4.3, Eqs. (31)–(36), Table 6Inspect

Chunk rewards combine geodesic progress, path-length agreement, goal entry/potential, collision, route deviation, backward motion and yaw penalties. Navigation-active segments can end before the full raw predicted chunk used for path-length agreement.

Go to primary source ↓

e-stop-rewardPDF pp. 20–21, A.4.4, Eqs. (37)–(41), Table 7Inspect

Stopping credit combines goal-potential change, translation/rotation energy, STOP success/failure and an exit penalty. Table 7 gives a 1.5 m goal radius; navigation and stopping rewards are clipped.

Go to primary source ↓

e-stage3-settingsPDF pp. 21–22, A.4.5–A.4.6, Eqs. (42)–(44), Table 8Inspect

Navigation losses weight chunks (8,4,2,1)/15 and stopping losses average them. Stage III uses condition batch 64, 16 sampler transitions, CFG 5, AdamW at 5×10⁻⁶, 1,500 updates, action-side weight 0.25 and stopping weight 0.50.

Go to primary source ↓

e-protocolPDF p. 22, B.1 GN-Bench Closed-Loop EvaluationInspect

Seen/Unseen have 1,000/5,000 episodes. Each WNM checkpoint runs once per episode with inference seed 1140, executes at most eight actions before replanning, and shares STOP rules. Metrics follow GN-Bench; no multi-seed standard deviations or bootstrap intervals are computed.

Go to primary source ↓

e-resultsPDF pp. 9–10, Section 4.2 and Table 1, WNM-3D, WNM-2D and GN-BAE rows, Seen/Unseen columnsInspect

WNM-3D Seen TL/NE/OS/SR/SPL is 10.6/1.9/94.2/88.7/80.9; Unseen is 10.2/4.1/62.2/53.5/46.6. WNM-2D SR/SPL is 75.6/72.9 Seen and 45.9/42.8 Unseen. GN-BAE BEV+FPV gives 58.6/58.6 Seen; FPV-only gives 38.9/37.3 Unseen. The prose explicitly disclaims a reduced Seen-to-Unseen gap.

Go to primary source ↓

e-ablationPDF pp. 10–11, Section 4.3 and Table 2, WNM-3D rowsInspect

Seen SR across I/II/III is 56.7/80.7/88.7. Unseen SR is 43.1/49.7/53.5, SPL 41.6/38.2/46.6, and TL 5.9/13.2/10.2. Direct Stage-I-to-DanceGRPO yields Seen/Unseen SR 47.1/39.2. Weak branch diversity/reward spread is hypothesized, not measured.

Go to primary source ↓

8.5 Primary sources

WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN ↗

PDF · 10,637 extracted words

Source fingerprint

fb921446edb659ab1cb46214157e935ea1c17a3e2760e3958c1893ab535a0b30