WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation
1. Paper overview
In one sentence: WAM-Nav couples a long action trajectory to one-step latent visual foresight inside a shared generator, improving reported navigation performance while retaining task-balancing and embodiment limits. e01e02e03e04e09e10e12e18
| At a glance | What to know |
|---|---|
| Research problem | Reader analysis The authors target mapless Image-Goal, Point-Goal and No-Goal navigation. Their concern is that reactive policies lack explicit future perception, while sequential imagination–policy pipelines add latency and error. The proposed alternative couples action and visual prediction inside generation; learned geometric consistency is a motivation, not a formal collision-avoidance guarantee. e20e04e05 |
| Core mechanism | |
| A key reported result | Image-Goal navigation: 50.2 / 48.2 Success rate (SR) / success weighted by path length (SPL), percent; higher is better. Zero-shot IsaacSim Dingo evaluation on ClutterScenes and InternScenes; original pretrained weights, no finetuning. Section 4.1 specifies 100 episodes per scene, 6,000 total. NavDP: 43.4 / 41.4. The SR gain is 6.8 percentage points, approximately 15.7% relative (reader calculation). For the Ours/NavDP Image-Goal SR and SPL entries, equal means of the four displayed scene categories agree with the printed averages after rounding. This does not establish the aggregation rule for every Table 3 cell or justify relabeling these summaries as pooled episode rates. e07e09 |
| Reading caution | Source description Camera height and field of view can hide low nearby obstacles; camera-clear paths can still collide with the robot body because embodiment is not explicitly modeled. e18 |
Core contributions
Figure 2. Goal information steers both perception and motion context before joint generation. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read left to right. The blue observation branch enters the frozen vision encoder; the yellow motion branch enters the learned motion encoder. The central goal module routes visual query g_V upward and geometric query g_G downward. These colors distinguish the two functions, rather than two independent navigation policies. Condition fusion produces C, the context supplied to the shared Diffusion Transformer. On the right, action and latent tokens enter the same generator and leave through distinct output heads. Equations (2)–(5) explain the operations compressed by the arrows: residual visual gating, geometric-goal retrieval of motion history, then motion-biased visual retrieval. e02e03e04e08e17
What it supports. The architecture makes future perception part of action generation. Its two outputs correspond to a planned action sequence and latent visual foresight; the default horizons are 24 action steps and one visual step. The schematic supports a unified generator even though perception encoders, output heads and the physical controller are separate components.
Where the evidence stops. The caption calls the history RGB-D, but Section 3.2 explicitly describes RGB encoding and does not specify a depth-fusion path. The diagram’s t at the generator denotes the flow timestep represented by tau in Eqs. (6)–(8); it should not be read as a new observation.
2. Motivation
2.1 The problem and the proposed response
The authors target mapless Image-Goal, Point-Goal and No-Goal navigation. Their concern is that reactive policies lack explicit future perception, while sequential imagination–policy pipelines add latency and error. The proposed alternative couples action and visual prediction inside generation; learned geometric consistency is a motivation, not a formal collision-avoidance guarantee. e20e04e05
2.2 What this reading follows
A navigation policy must decide where to move while its viewpoint changes with every action. WAM-Nav brings a prediction of the next visual state into the same generative process that proposes the path. The important design choice is asymmetric: the action plan extends farther than the imagined visual future. This reading follows goal routing and motion-conditioned perception into the shared Transformer, then checks the claim against simulation, ablations and physical trials. The results favor coupling, but they do not establish guaranteed collision avoidance, optimal goal balancing, or a fully specified reproduction protocol. The inspected source is arXiv v2, dated 13 June 2026. e01e02e03e04e09e10e12e18
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
The recorded One Model × Joint prediction quadrant is supported by architecture: action and future-visual tokens share self-attention and generative parameters, rather than merely sharing a training dataset. Foresight participates in inference-time generation. Separate frozen encoders and downstream MPC do not imply separate world/policy generators. The latent-prediction subcategory fits VAE latents; no JEPA-specific objective is demonstrated. e04e06e17
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Use two goal queries to organize one navigation context
An image goal contains visual identity, whereas a point goal supplies relative geometry. WAM-Nav preserves this difference in its initial encoders, then creates two functional projections for either input: g_V asks which visual patches are relevant, and g_G asks how recent motion relates to the target direction. The first query reinforces visual memory with a sigmoid residual. The second retrieves a kinematic summary from relative pose history. That summary biases learned queries which attend to visual memory, so fusion is directional: motion guides visual retrieval. Reader interpretation: this is an inductive bias for continuity, not an explicit momentum constraint enforced by a dynamics solver. Table 6 exposes the cost of unification: single-query variants win their favored goal task, while the combined version offers more balanced performance and the strongest No-Goal coverage. e02e03e05e12
Figure 5. The place where action and visual tokens interact is inside each DiT block. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the bottom, where action noise and image noise enter separate adaLN branches. The central time embedding modulates both. Both QKV arrows then enter one self-attention block, making action–visual exchange explicit. Residual paths continue upward into normalization and cross-attention; C supplies keys and values to both streams. The Shared marker identifies tied cross-attention and feed-forward weights, while the two streams retain separate modulation. Appendix C.2 and Eq. (7) confirm that the stacked blocks predict joint flow velocities. The figure uses t for time embedding, while the equations use tau for flow time. e03e04e06e11
What it supports. This is architectural evidence for One Model × Joint prediction. Coupling occurs during token processing, beyond sharing a conditioning vector or adding a visual loss to an otherwise separate policy. Table 5 is directionally consistent: Image-Goal SR is 50.2% for full sharing, 47.5% for partial sharing and 45.9% for decoupled DiT.
Where the evidence stops. These comparisons do not report parameter- or compute-matched controls or variation across seeds. The figure does not specify an action/visual causal mask; the separately described causal encoder belongs to motion history.
5.2 Predict visual consequences inside action generation
Training begins with a paired action trajectory and future observation latents. Gaussian action and visual noise are linearly interpolated toward those targets at a sampled flow time. The shared DiT regresses both velocity fields, while a contrastive term aligns equivalent goal representations. Frozen DINOv2 and the VAE supply visual representations; the navigation-specific modules learn from the simulated trajectories. At inference, ten Euler integration steps jointly update both prediction streams. The visual branch is therefore present in generation, rather than serving only as a discarded auxiliary training head. Crucially, the method samples sixteen candidates but selects the first; it does not describe a visual critic ranking those candidates. Reader interpretation: the proposed benefit lies in internal coupling and training regularization, which must be tested separately from any hypothetical foresight-based candidate selection. e04e06e07e08
5.3 Separate anticipation, trajectory prediction and physical execution
The asymmetric horizon gives the policy a long action sequence and a short predicted visual future at each decision. New observations then rebuild the context for receding-horizon control. The action history is maintained from executed actions and converted into the current frame, so it is not a supplied global map. On the physical G1, an MPC tracker turns predicted trajectories into velocity commands; policy inference, camera acquisition and motor control operate on different schedules. The reported controller frequency and command duration still need clarification. Reader interpretation: navigation success measures the whole deployed loop, including sensing and tracking, rather than visual prediction alone. The failure cases make this boundary concrete: a path can look passable from the camera while the body strikes a side obstacle, and a low obstacle can disappear from view. e04e05e13e17e18
5.4 Training and inference
During training
Training uses VLN-N1: over 400 simulated hours and 200K trajectories from six scene-asset families with domain randomization. Reported cost is approximately 8 × 120 GPU-hours. DINOv2 ViT-S/14 and the Stable Diffusion VAE remain frozen; the goal image encoder, fusion decoder, motion encoder and DiT train from scratch. e07e08
AdamW uses batch 256, peak learning rate 1.5 × 10^-4, weight decay 10^-4, cosine decay and 5% warmup. Joint flow-velocity regression is weighted with a symmetric InfoNCE goal-alignment term; visual and alignment weights are 0.25 and 0.1. e04e08
During inference
At each receding-horizon decision, rebuild C and jointly integrate action and latent noise with ten Euler steps. The policy samples sixteen trajectories and executes the first sampled trajectory; the supplied method does not describe foresight-based ranking of candidates. e04e08
Physical deployment uses a Unitree G1, D455 RGB-D camera at 30 Hz and 20° downward pitch, and an onboard RTX 4060. Policy inference is configured at 1 Hz with approximately 1.3 GB GPU memory. MPC tracks predicted trajectories and generates motor velocity commands. e17
5.5 Implementation flow
- Preserve goal modality
A scratch-trained image ViT, sinusoidal coordinate encoding, or masked zero-state produces a base goal embedding. Separate learned projections produce visual-semantic query g_V and geometric query g_G. e02
- Retrieve visual and kinematic context
Frozen DINOv2 encodes visual patches. Sigmoid goal relevance adds a patchwise residual, reinforcing rather than deleting tokens. Executed actions update a pose buffer; positions are recentered and rotated into the current egocentric frame, with wrapped heading differences and front zero-padding. e02e05e08
- Let motion guide visual retrieval
A causal Transformer encodes relative motion. The geometric goal cross-attends to it to form o_kin; its projection biases learned queries that retrieve the goal-modulated visual memory through a Transformer decoder, producing context C. e03
- Denoise both prediction streams together
Action and VAE-latent tokens share self-attention at every DiT layer, with separate timestep modulation branches and shared cross-attention/FFN weights. Both attend to C. Defaults use eight history steps, 64 conditioning tokens, 16 DiT layers and horizons H_act=24, H_vis=1. e04e06e08
6. Experiments & results
WAM-Nav jointly generates navigation trajectories and short-horizon visual latents using one shared Diffusion Transformer. Goal-conditioned visual and motion histories guide a 24-step action horizon with one future visual state. Reported zero-shot navigation improves over NavDP, but task balancing sacrifices specialized peaks, and physical failures expose camera visibility and body-clearance limits. Method, results and boundaries are traced in e02–e18.
6.1 Read the original evidence
Table 3. Read each task with its own metric pair and preserve the four scene categories. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the task headings: M1/M2 mean success rate (SR)/success weighted by path length (SPL) for Image-Goal and Point-Goal, but area/time for No-Goal. All arrows point upward. Follow the Ours and NavDP rows across the two ClutterScenes columns and two InternScenes columns before reading Average. Section 4.1 specifies zero-shot evaluation in IsaacSim with a Dingo, original pretrained weights and no additional finetuning. There are 100 episodes per scene across 10 easy, 10 hard, 20 home and 20 commercial scenes. For Ours/NavDP Image-Goal SR/SPL, equal category means agree with the printed averages after rounding; this reader calculation does not establish the aggregation rule for every cell or pooled episode rates. e07e09e21
What it supports. Image-Goal SR rises from 43.4% to 50.2%: 6.8 percentage points, or approximately 15.7% relative. Point-Goal rises from 77.8% to 80.4%, a 2.6-point increase. Source-reported No-Goal explored area rises from 167.2 to 171.1 m², subject to the baseline discrepancy below. These are simulation outcomes, and the Home Image-Goal SR remains only 30.7%.
Where the evidence stops. NavDP’s No-Goal area entries (315.6, 274.1, 34.1, 43.9) average to 166.925 m², about 166.9, whereas the table prints 167.2; the discrepancy is unresolved. Baseline training data and compute are not controlled, and uncertainty is absent. No-Goal time has a 120-second cap and early stuck termination; it is not a speed measure.
Figure 4. Physical trials demonstrate transfer while leaving substantial visibility and clearance limits. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as one environment, with example robot frames and inset first-person trajectory overlays on the left. At the right, blue Ours and gray NavDP bars show successful trials out of ten; use the printed counts rather than estimating bar heights. The rows are meeting room, warehouse, lobby and parking lot. Section 4.2 identifies a Unitree G1 with a D455 camera. Appendix G explains the execution chain: the onboard model predicts trajectories at 1 Hz, while MPC converts them into velocity commands. The images therefore document a policy-plus-controller deployment. e15e16e17e18
What it supports. The visible counts are 8, 9, 10 and 7 successes for WAM-Nav, versus 4, 6, 7 and 2 for NavDP. Summing equal-sized scene groups gives 34/40, matching the reported 85%; NavDP is 19/40 or 47.5% by reader calculation. The parking lot remains the weakest WAM-Nav scene here.
Where the evidence stops. Only forty trials per method are shown, without uncertainty or a stated goal-modality protocol. Appendix H reports low-obstacle and body-clearance collisions. These physical results must remain separate from the simulated G1/H2 comparisons in Table 8.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Image-Goal navigation Zero-shot IsaacSim Dingo evaluation on ClutterScenes and InternScenes; original pretrained weights, no finetuning. Section 4.1 specifies 100 episodes per scene, 6,000 total. | 50.2 / 48.2 Success rate (SR) / success weighted by path length (SPL), percent; higher is better | NavDP: 43.4 / 41.4. The SR gain is 6.8 percentage points, approximately 15.7% relative (reader calculation). For the Ours/NavDP Image-Goal SR and SPL entries, equal means of the four displayed scene categories agree with the printed averages after rounding. This does not establish the aggregation rule for every Table 3 cell or justify relabeling these summaries as pooled episode rates. e07e09 |
| Point-Goal navigation Same zero-shot Dingo benchmark and four-category Table 3 summary. | 80.4 / 78.0 SR / SPL, percent | NavDP: 77.8 / 74.8. SR improves 2.6 percentage points, approximately 3.3% relative (reader calculation). e07e09 |
| No-Goal exploration Same benchmark; time capped at 120 seconds with early termination when locally stuck. | 171.1 / 75.6 Explored area (m²) / episode time (s), higher is better | NavDP, source-reported Average cells: 167.2 / 72.5. These measure coverage and episode persistence, not goal-reaching success. Reader calculation: NavDP’s four displayed area values (315.6, 274.1, 34.1, 43.9) average to 166.925 m², approximately 166.9, whereas the table prints 167.2. The reported value is retained; the source does not explain this discrepancy or provide an aggregation rule that resolves it. e07e09e21 |
| Image-Goal component and horizon ablations Appendix F.2–F.3 Image-Goal evaluation. | Neither component: 42.1; latent only: 45.7; motion only: 43.9; both: 50.2. With H_act=24, H_vis=1/4/8/24 gives 50.2/46.2/39.6/30.4. SR, percent | Shared DiT: 50.2; partially shared: 47.5; decoupled: 45.9. Consistent with useful coupling and short foresight; missing repeat variability and compute-matched controls limit causal attribution. e10e11e13 |
| Inference efficiency Table 1 per-decision comparison; timing hardware/profiling protocol unspecified. | 0.26 / 0.7; 234.9M total and 129.2M trainable parameters. Latency (s) / TFLOPs | NavDP: 0.16 / 1.3; NWM: 1.43 / 8.3. Lower arithmetic cost does not make WAM-Nav faster than NavDP; reported policy latency differs from configured deployment frequency. e14e17 |
| Physical zero-shot navigation G1 in meeting room, warehouse, lobby and parking lot; ten trials per scene; goal modality not specified. | 8/10, 9/10, 10/10, 7/10; reported average 85%. Successful trials / SR | NavDP: 4/10, 6/10, 7/10, 2/10; pooled 47.5% by reader calculation. A small physical execution study, distinct from simulated G1/H2 transfer in Table 8. No uncertainty interval is supplied. e15e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. Motion history and latent foresight contribute differently across environments. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the first two columns as a two-factor comparison: crosses remove motion history or latent prediction, and checkmarks retain them. First compare latent-only with neither component to see the visual contribution. Then compare motion-only with neither, looking separately at ClutterScenes and InternScenes. Finally read the bottom row to assess the combination. The overall columns show whether a local improvement survives averaging. This layout matters because a single overall number hides the motion-only variant’s loss on InternScenes. The table is an Image-Goal ablation; it does not report physical trials or the other goal settings. e10
What it supports. Latent prediction alone increases overall SR from 42.1% to 45.7%; motion alone reaches 43.9%; both reach 50.2%. Motion alone improves ClutterScenes SR from 51.0% to 57.9% but reduces InternScenes SR from 33.2% to 30.0%. The combined result supports complementarity more strongly than a claim that either component helps everywhere.
Where the evidence stops. Component removal changes available inputs or supervision as well as the resulting computation. The table supplies no uncertainty or direct smoothness metric, so a precise causal interaction strength and guaranteed kinematic improvement remain unestablished.
Table 7. Long action planning performs best with short visual foresight in this sweep. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the Setting column carefully: H_act stays at 24 in every row, while H_vis changes from 1 to 4, 8 and 24. This isolates the nominal prediction horizon being varied; it does not mean that the amount of computation or the optimization problem stays constant. Move horizontally to SR and SPL, both higher-is-better navigation metrics. The first row is the default configuration in Table 2. The final row makes the two horizons equal. Thus the table tests the paper’s asymmetric design through executed navigation performance, rather than reporting an image reconstruction score. e08e13
What it supports. SR declines from 50.2% with one visual step to 46.2%, 39.6% and 30.4% as the visual horizon increases; SPL declines in parallel. This supports the choice of near-future foresight for the tested policy. It does not show that all long-horizon world models would perform worse.
Where the evidence stops. The authors attribute degradation to visual error accumulation under viewpoint changes, but this table has no horizon-wise latent-error or latency measurements. Loss normalization and matched training budgets need checking before assigning the entire decline to that mechanism.
7. Analysis & limitations
7.1 What the evidence leaves open
Camera height and field of view can hide low nearby obstacles; camera-clear paths can still collide with the robot body because embodiment is not explicitly modeled. e18
Goal balancing is a tradeoff: visual-only queries reach 55.6% Image-Goal SR versus 50.2 jointly; geometric-only reaches 83.5% Point-Goal SR versus 80.4 jointly. The joint variant leads No-Goal coverage, not every task-specific maximum. e12
Table 9 lists NavDP SPL above SR in its Medium and Hard subsets; its Average columns are not explained by the displayed subset means. Absolute SR gaps also do not grow monotonically with difficulty. These entries need clarification before adopting the authors’ difficulty trend. e16
Pretrained baselines share an evaluation environment, not a controlled training budget. Qualitative decoded foresight examples do not quantify dynamics accuracy or establish physical safety guarantees. e07e19
7.2 Questions for discussion
- Does inference-time cross-modal coupling improve navigation beyond the regularization supplied by the visual training loss (e06, e11)?
- Would explicit body-clearance conditioning preserve transfer while reducing the failures in Appendix H (e18)?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the specified VLN-N1 data, frozen visual checkpoints, Table 2 architecture, and original baseline weights. The source omits training GPU model, total update count, exact success thresholds, simulator version and detailed InfoNCE pairing/temperature. RGB-D is named, but a depth-fusion pathway is not specified. e02e04e07e08
Deployment reports a 50 Hz control loop and 0.1-second command application without explaining their scheduling relationship. MPC parameters, action scaling and physical goal protocol also need clarification; Table 1 timing hardware cannot be inferred from the RTX 4060 deployment description. e14e15e17
Proposed checks: compare shared and attention-decoupled predictors under matched training budgets; repeat the visual-horizon sweep with per-horizon loss normalization and latency measurements. Both should use fixed evaluation episodes and repeated seeds. e06e11e13
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 cross-modal attention help beyond visual supervision?
Reader-proposed, not run: use the reported default history and horizons to train a fully shared DiT and a control with a block-diagonal action/visual self-attention mask. Keep data, parameter count, context, both losses, training updates, Euler steps and first-candidate selection fixed. Evaluate identical Image-Goal episodes over at least three seeds, logging SR/SPL, collisions and visual velocity error. Also disable cross-stream attention only at inference in the shared model, labeling that intervention as distribution shift. A repeatable full-model advantage over the trained masked control would support inference-relevant coupling; parity would weaken that claim even if visual supervision remains useful. e04e06e07e08e11
Check 2: Is the visual-horizon penalty prediction error or training allocation?
Reader-proposed, not run: repeat H_vis=1,4,8,24 with H_act=24, the same data, seeds, evaluation episodes and ten Euler steps. Explicitly compare mean-per-latent-element loss against summed visual loss, keeping the stated visual weight fixed, because the paper does not give the reduction convention across horizons. Run one equal-update and one equal-training-compute comparison. Record SR/SPL, collision rate, per-offset latent prediction error, token count and end-to-end latency. Persistent degradation alongside increasing future error would strengthen the authors’ account; recovery after normalization or budget adjustment would implicate optimization or compute allocation. e04e07e08e13
8.3 Reading coverage
Visual audit: Pages 1–8 and 12–19 were rendered and actually viewed, covering identity, every method/equation, training and evaluation details, all figures and Tables 1–9. Figure 2 arrows and frozen-encoder marker were checked against Eqs. (1)–(5); Figure 5 self-attention, condition KV and shared-weight markers were checked against Eq. (7) and Appendix C.2. The figures label diffusion time t while the equations use tau; captions describe RGB-D while explicit encoding prose specifies RGB. All six final original crops were viewed and retain complete visual content, legends and table headers. Uncropped pages include Algorithm 1, all implementation settings, the goal-query ablation, timing table, Table 9 anomalies, hardware/control details and the failure explanation continuing on p. 19. Reference pages 9–11 were fully read as text but were not visually inspected. Separate supplements, code and external links were not inspected. Table 3’s NavDP No-Goal area entries and printed Average cell were rechecked on p. 15 and the final crop; their arithmetic discrepancy is disclosed without replacing the source-reported value.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18, 19. 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 and abstract (p. 1)
- 1 Introduction (pp. 1–3)
- 2 Related Work, 2.1–2.2 (p. 3)
- 3 Methodology, 3.1–3.4 (pp. 3–6)
- 4 Experimental Results, 4.1–4.2 (pp. 6–8)
- 5 Conclusion and Limitations (p. 8)
- References (pp. 9–11)
- Appendices A–B and Algorithm 1 (p. 12)
- Appendix C.1–C.2 (p. 13)
- Appendices D–E (p. 14)
- Appendix F.1–F.4 (pp. 15–17)
- Appendix G.1–G.2 (pp. 17–18)
- Appendix H (pp. 18–19)
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.
- The extraction-only visual limitation was addressed by inspecting the retained PDF: all figures and tables were visually read, and six final crops were inspected.
- Identity: the inspected title page identifies arXiv:2606.04907v2 [cs.RO], 13 June 2026. Title and all twelve authors match the catalog after name-order normalization. The catalog does not specify a revision; v1 and revision differences were not supplied or compared (e01).
- All five supplied text chunks were read individually. Reference pages 9–11 were read as text; visual inspection covered pages 1–8 and 12–19.
- No code, external linked resources, or separate supplements were inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author/affiliation block and arXiv margin
WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation; twelve named authors and seven affiliations; arXiv:2606.04907v2, 13 Jun 2026.
Go to primary source ↓e02PDF pp. 3–4, Sections 3.1–3.2, Figure 2 and Eqs. (1)–(3)
Modality-specific goal encoders form two queries. RGB patch memory receives sigmoid-weighted residual reinforcement. Figure 2 depicts a frozen vision encoder and shared generator; its caption describes RGB-D.
Go to primary source ↓e03PDF p. 5, Section 3.2, Eqs. (4)–(5)
The geometric goal queries causal motion features; the resulting kinematic token biases learned decoder queries over visual memory to form C.
Go to primary source ↓e04PDF pp. 5–6, Sections 3.3–3.4, Eqs. (6)–(8)
VAE future latents and actions share flow interpolation and DiT attention. The loss combines two velocity regressions and symmetric InfoNCE. Online control samples sixteen trajectories and selects the first.
Go to primary source ↓e05PDF pp. 12–13, Appendix B, Algorithm 1 and Appendix C.1
Action-accumulated poses use a sliding, front-zero-padded buffer, current-frame recentering/rotation and wrapped relative headings.
Go to primary source ↓e06PDF p. 13, Appendix C.2 and Figure 5
Separate adaLN-Zero branches feed shared action/image self-attention; both branches use context C with shared cross-attention and FFN weights.
Go to primary source ↓e07PDF p. 6, Section 4.1, training/evaluation/metrics paragraphs
VLN-N1 supplies over 400 hours and 200K trajectories. Cost is 8 × 120 GPU-hours. Evaluation uses 10 easy, 10 hard, 20 home and 20 commercial scenes, 100 episodes each; No-Goal time is capped at 120 seconds.
Go to primary source ↓e08PDF p. 14, Table 2 and Appendix D
Default history 8, action horizon 24, visual horizon 1, condition length 64, hidden dimension 384, sixteen DiT layers. Frozen DINOv2/VAE, scratch-trained remaining named modules, AdamW and ten Euler steps are specified.
Go to primary source ↓e09PDF p. 15, Table 3, all three task blocks, Ours/NavDP rows and Average columns
Image-Goal: 50.2/48.2 versus 43.4/41.4; Point-Goal: 80.4/78.0 versus 77.8/74.8; No-Goal area/time: 171.1/75.6 versus 167.2/72.5.
Go to primary source ↓e10PDF p. 15, Table 4, all component rows and scene-group columns
Overall SR/SPL: neither 42.1/40.8, latent only 45.7/43.8, motion only 43.9/42.9, both 50.2/48.2. Motion-only InternScenes SR is 30.0 versus 33.2 without either.
Go to primary source ↓e11PDF p. 16, Table 5 and Appendix F.3 discussion
Memory 4/8/16 gives SR 46.6/50.2/44.9; removing goal injection gives 44.5; decoupled/partially shared/full DiT gives 45.9/47.5/50.2.
Go to primary source ↓e12PDF p. 16, Table 6, all goal-query variants
Visual-only Image-Goal SR is 55.6; geometric-only Point-Goal SR is 83.5. Joint queries yield 50.2/80.4 respectively, and the largest No-Goal area/time, 171.1/75.6.
Go to primary source ↓e13PDF p. 16, Table 7 and final Appendix F.3 paragraph
At action horizon 24, visual horizons 1/4/8/24 yield SR 50.2/46.2/39.6/30.4 and SPL 48.2/43.8/37.3/28.1.
Go to primary source ↓e14PDF p. 8, Table 1; p. 7, Q3 discussion
Ours/NavDP/NWM latency is 0.26/0.16/1.43 seconds and TFLOPs 0.7/1.3/8.3. WAM-Nav has 234.9M total and 129.2M trainable parameters; table timing hardware is not identified.
Go to primary source ↓e15PDF p. 8, Figure 4(a–b) and Section 4.2
Physical G1 evaluation has ten trials in each of four scenes. Figure bars show Ours 8/9/10/7 successes and NavDP 4/6/7/2; text reports 85% average success.
Go to primary source ↓e16PDF p. 17, Tables 8–9 and Appendix F.4
Table 8 is IsaacSim Dingo/G1/H2 transfer. Table 9 gives NavDP Medium 33.8 SR/35.6 SPL and Hard 17.4/21.3; Ours Easy/Medium/Hard SR is 52.5/42.4/25.0, versus 46.2/33.8/17.4.
Go to primary source ↓e17PDF pp. 17–18, Appendix G.1–G.2 and Figure 7
G1 carries D455 and RTX 4060. Deployment uses 20° camera pitch, 30 Hz observations, 1 Hz inference, about 1.3 GB memory and MPC; it reports both 50 Hz control and 0.1-second command application.
Go to primary source ↓e18PDF p. 8, Limitations; pp. 18–19, Appendix H and Figure 8
Failures involve low obstacles missed near the camera and side-body collisions despite a passable camera trajectory.
Go to primary source ↓e19PDF p. 7, Figure 3 and Q2 discussion
Two qualitative scenes compare NavDP/WAM-Nav trajectories and GT/decoded latent foresight; visual fidelity and smoother avoidance are author interpretations of these examples.
Go to primary source ↓e20PDF pp. 2–3, Introduction and Sections 2.1–2.2
The authors motivate joint modeling through the latency/error costs of decoupled imagination and action pipelines and the limited anticipation of reactive navigation.
Go to primary source ↓e21PDF p. 15, Table 3, Task 3: No-Goal Exploration, NavDP row, Easy/Hard/Home/Commercial M1 and Average M1 cells
The four displayed NavDP explored-area entries are 315.6, 274.1, 34.1 and 43.9; the source prints 167.2 in the Average area cell. No aggregation formula or explanation of these entries is supplied.
Go to primary source ↓8.5 Primary sources
WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation ↗
PDF · 7,878 extracted words
Source fingerprint
fa2bafda5333466692978418ebfb9a420da91bf52acf84671a127bc5f54824a5