Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation
1. Paper overview
In one sentence: DualVLN uses supervised pixel goals and learned latent queries to connect slow semantic planning with fast RGB-conditioned trajectories, improving navigation while leaving substantial dynamic-obstacle failures. e-identitye-rationalee-architecturee-goalse-ce-resultse-pe-resultse-social-results
| At a glance | What to know |
|---|---|
| Research problem | Source description Frequent VLM calls for short discrete actions can fragment motion and delay reactions. DualVLN asks how a semantic planner can guide continuous navigation while a faster policy responds to changing observations. e-rationale |
| Core mechanism | Source description The interface combines interpretable pixel-goal supervision with four learnable queries that extract task-relevant conditioning from a subsequently frozen VLM. e-goalse-training-local |
| A key reported result | R2R-CE unseen instruction-following: 64.3%; 58.5; 4.05m SR ↑; SPL ↑; NE ↓. Habitat VLN-CE, R2R validation-unseen; single-view RGB policy. StreamVLN: 56.9%; 51.9; 4.98m. SR improves by 7.4 percentage points. Success means stopping within 3m; SPL penalizes long paths. This is simulated navigation. e-ce-results |
| Reading caution | Reader analysis Local control tolerates small direction-preserving pixel errors, but large or semantically wrong goals remain failure cases, especially near obstacles. Scaling approaches saturation around 10% of System-2 trajectories; attributing this solely to a System-2 ceiling is the authors’ interpretation. e-diagnostics |
Core contributions
- Source description
The interface combines interpretable pixel-goal supervision with four learnable queries that extract task-relevant conditioning from a subsequently frozen VLM. e-goalse-training-local
- Source description
Social-VLN inserts Habitat 3.0 humanoids along R2R-CE routes, checks that paths remain passable, and adds Human Collision Rate (HCR). Its collection pipeline produces 763K social-navigation episodes across 60 MP3D scenes. e-social-design
Figure 2. An adaptive latent interface connects slow semantic grounding to fast visual trajectory updates. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the purple System-2 branch. Instruction, history and the current view enter QwenVL; its outputs include view adjustments and a pixel goal. The orange trajectory queries produce the latent goal carried along the conditioning arrow. In the blue branch, RGB at time t is the image associated with that goal, while RGB at t+k is the newer observation. Their ViT features are fused, compressed by the Q-Former, and concatenated with the latent goal. The resulting features supply keys and values to DiT cross-attention; the noisy trajectory supplies the trajectory stream. Section 3.2 and Appendix A.2 clarify that queries follow generated pixel text. e-architecturee-goalse-flowe-training-local
What it supports. The fast policy can refresh its trajectory while the semantic goal remains unchanged. The pixel coordinate is supervised explicitly, but its influence reaches DiT through contextual latent conditioning. This gives the local planner both task direction and newer visual information without waiting for another complete VLM decision.
Where the evidence stops. The figure is schematic: it does not show the full autoregressive timing or frozen-weight schedule. The plus-in-circle means concatenation under Eq. (2), not arithmetic addition. Appendix A.2 supplies the training sequence needed to interpret the interface.
2. Motivation
2.1 The problem and the proposed response
Frequent VLM calls for short discrete actions can fragment motion and delay reactions. DualVLN asks how a semantic planner can guide continuous navigation while a faster policy responds to changing observations. e-rationale
2.2 What this reading follows
A navigation instruction might require a robot to pass through a doorway, turn at a landmark and stop in another room. DualVLN gives the semantic decisions to a large vision-language model and lets a smaller diffusion policy refresh the route between those decisions. The connection carries both explicit pixel-goal information and learned contextual features. This edition follows that interface from supervision through asynchronous execution, then checks what the experiments actually establish. Static unseen-environment gains are substantial, but human-collision rates and physical-controller failures remain important limits. The reviewed artifact is the supplied December 2025 arXiv v1, not a separately verified ICLR 2026 edition. e-identitye-rationalee-architecturee-goalse-ce-resultse-pe-resultse-social-results
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 | VLA |
| Architecture | Dual-system |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The architecture supports hierarchical, dual-system VLA: a VLM produces grounded goals, and a distinct diffusion transformer produces trajectories. No learned future-world observation rollout or inverse-dynamics action extraction is specified. The recorded prediction-paradigm/quadrant Not applicable is appropriate; shared conditioning does not make this One Model. e-architecturee-flowe-training-local
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 First teach the planner where a useful goal can be seen
System 2 is trained to answer a spatial question before the trajectory policy is trained. The supervision projects an expert route into the egocentric image, discards occluded points using depth, and selects the farthest visible waypoint. When the route is not visible, the model instead learns view adjustment; it also learns when to output STOP. This makes looking around part of the navigation interface rather than assuming that every useful waypoint is already in view. Stage 1 fully finetunes QwenVL. Only afterward does the paper freeze that backbone for local-policy learning. Depth therefore has a clear training role even though the learned navigation inputs are RGB. Figure 2 should be read alongside Appendix A.1: its outputs summarize an iterative process, not one simultaneous prediction of every action type. e-goalse-training-globale-architecture
5.2 Then learn what the fast policy should extract from the goal
Once the pixel coordinate has been generated, four learnable query embeddings are appended to its context. The frozen VLM transforms those queries into latent goals, and the trajectory loss trains the queries to extract information useful to System 1. A separate visual branch combines the image associated with the goal and the latest RGB image, compressing their fused features into conditioning tokens. DiT uses both kinds of conditioning while learning the flow-matching velocity objective. This is the operational meaning of the explicit-plus-implicit interface: supervised coordinate text remains in the context, while learned queries select richer information for motion. Figure 7’s yellow variant retains frozen pixel-token features, so calling it a complete removal of latent information would misdescribe the experiment. e-training-locale-architecturee-flowe-goal-ablation
5.3 Finally separate a fresh trajectory from a successful execution
Reader interpretation: the two observation times are essential to understanding the asynchronous design. A latent goal can remain fixed while the robot moves or an obstacle changes position; current RGB gives System 1 a way to update its trajectory under that older instruction-grounded goal. Execution then depends on the deployment stack, which uses odometry and MPC after remote inference. Consequently, the architecture’s update-frequency labels are not a full measurement of sensor-to-motion latency. The evaluations probe different parts of this chain: VLN-CE tests instruction-following, VLN-PE adds physical-control failures, and Social-VLN exposes collisions with moving humanoids. The results justify a stronger navigation policy in these settings, but the 35.4% HCR prevents treating fast replanning as a demonstrated solution to safe human-aware navigation. e-architecturee-overviewe-deploymente-ce-resultse-pe-resultse-social-results
5.4 Training and inference
During training
Stage 1 fully unfreezes QwenVL-2.5-7B, including vision and language backbones, following StreamVLN’s referenced data recipe. The main text specifies one epoch; Appendix A.1 gives AdamW, learning rate 2×10⁻⁵, batch 128 conversations and 14,000 steps. e-training-global
Stage 2 freezes QwenVL and trains queries and the diffusion policy on pixel-goal samples. Discrete waypoints become 32 interpolated fixed-interval waypoints. AdamW uses 10⁻⁴, batch 128 trajectories and 15,000 steps. The RGB encoder uses DepthAnythingV2-Small; DiT has width 384, 12 layers and six heads, with goal features projected 3584→768. e-training-local
During inference
System 2 alternates view adjustment, goal prediction and stopping; System 1 updates trajectories asynchronously. Deployment streams RGB-D from D455 cameras on Turtlebot4, Go2 and G1 to an RTX 4090 server using 20GB. Odometry transforms outputs into world coordinates, and MPC tracks them. RGB-only describes policy inputs, not the complete sensing/control stack. e-goalse-architecturee-deployment
Figure 1 labels System 2 at 2Hz and System 1 at 30Hz. Deployment reports cached System-2 token inference of 0.7s, down from 1.1s, and System 1 generating '32 trajectories' in 0.03s with TensorRT. These are not a unified end-to-end latency measurement; 32 trajectories must not be silently equated with 32 waypoints. e-overviewe-deploymente-architecture
5.5 Implementation flow
- Ground a visible waypoint
Depth-filtered trajectory projections identify the farthest visible waypoint for supervision. System 2 predicts that coordinate; if no future waypoint projects into view, it predicts turns. The method describes 15° turn/look adjustments; Appendix A.1 caps turn chunks at four and supervises terminal STOP. e-goals
- Extract an adaptive latent goal
Four learnable queries Z follow the pixel text in context X. Passing [X; Z] through the frozen VLM yields Z′. The explicit goal reaches the local policy through contextual latent features, without requiring depth conversion into a point goal. e-architecturee-training-locale-local-planners
- Update motion under an older goal
System 1 holds Z′ from time t while refreshing observations at t+k. A ViT encodes both frames, self-attention fuses them, and a Q-Former compresses them into 32 visual tokens F. Concatenated Z′ and F condition DiT cross-attention. e-architecture
6. Experiments & results
DualVLN separates instruction-grounded pixel-goal planning from fast trajectory generation. A VLM supplies waypoint text and learned latent goals to an RGB-conditioned diffusion policy. Unseen navigation improves, but dynamic-human collisions and incomplete implementation details limit stronger conclusions.
6.1 Read the original evidence
Table 1. Unseen instruction-following improves across both benchmarks, with sensor regimes kept visible. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the observation columns before comparing scores: Pano., Odo., Depth and S.RGB separate panoramic images, odometry, depth and single-view RGB. The last block is the most direct sensor-level comparison for DualVLN. Within each dataset, lower navigation error (NE) is better; higher success rate (SR) and success weighted by path length (SPL) are better. On R2R, the OS column also reports oracle success, which considers the closest point reached rather than only the stopping point. Follow the final two rows across R2R first, then RxR. The retained asterisk note identifies methods using an external waypoint predictor; dashes preserve missing measurements. e-ce-results
What it supports. DualVLN reaches 64.3% SR and 58.5 SPL on R2R, versus StreamVLN’s 56.9% and 51.9. On RxR, it reaches 61.4% and 51.8, versus 52.9% and 46.0. These are 7.4- and 8.5-percentage-point SR improvements on validation-unseen splits, with lower final navigation error as well.
Where the evidence stops. These are Habitat navigation results, not physical-robot completion rates. SR accepts stopping within 3m. Matching RGB inputs does not equalize training data, model capacity or compute, and the table supplies no uncertainty estimates.
Table 2. Physical simulation reveals a completion advantage alongside residual control failures. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First distinguish the training regimes. The upper rows use VLN-PE training, whereas the lower block transfers from VLN-CE; the retained caption explains the plus and dagger markers. Next choose the right half for validation-unseen environments. SR and SPL measure navigation completion and path efficiency, while FR and StR measure falls and inability to move. The source evaluates a simulated Unitree H1 with a physical locomotion controller. Compare DualVLN with NaVid within the transfer block, reading the failure columns as carefully as the bold success numbers. This controller differs from the flash controller used in the local-planner substitution experiment on page 9. e-pe-resultse-local-planners
What it supports. DualVLN records 51.60% unseen SR and 42.49 SPL, versus NaVid’s 22.42% and 18.58, without VLN-PE finetuning. However, DualVLN’s fall rate is 12.32% versus 8.61%, and stuck rate is 2.23% versus 0.45%. Better route completion does not mean every physical-control metric improves.
Where the evidence stops. The result is physically simulated transfer, not an H1 hardware trial. Table 4’s 63.62% SR uses a flash controller and cannot replace this table’s physical-controller score or be pooled with it.
Table 3. Moving humanoids expose a much larger gap than the static benchmark suggests. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across a row to see how the same method changes from standard R2R to Social-VLN. Then compare rows within the dynamic block. The benchmark places humanoids along reference routes, increasing the chance of interaction while checking that the route remains physically passable. HCR is the extra human-collision metric at the far right; lower is better. SR and SPL still reward finishing the instruction and avoiding unnecessary travel. This layout separates task completion from human-contact failures. It also lets you calculate the static-to-dynamic success drop directly, without relying on the approximate percentages stated in the surrounding prose. e-social-designe-social-results
What it supports. DualVLN’s SR falls from 64.3% to 37.2%, a 27.1-percentage-point decline. It still exceeds StreamVLN’s 31.4% dynamic SR and 29.1 SPL, reaching 35.8 SPL. Yet its HCR is 35.4% versus 36.4%: the measured collision improvement is only one percentage point.
Where the evidence stops. These values do not establish dependable human-safe navigation. The paper provides no confidence intervals for the collision difference, and targeted simulated encounters do not characterize all pedestrian behaviors. The prose’s approximate 27% drop should be read as percentage points.
Figure 5. Hardware trials support improved room-to-room navigation within a small evaluation. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the shared legend to track the cyan DualVLN bars and yellow-green StreamVLN bars across the three scenarios. The left chart gives success rates; the right chart gives final navigation error in meters. Hallway, single-bedroom and room-to-room office are the paper’s easy, medium and hard scenario labels. The surrounding text specifies 20 trials per scenario per model, so each success changes a scenario’s percentage by five points. Read both charts together: a small navigation error alone does not establish that all requirements for success were met. These panels report scenario-level results rather than a separate quantitative table for each robot platform. e-real-resultse-deployment
What it supports. DualVLN has 100%, 95% and 70% success across hallway, bedroom and office, with errors of 0.2m, 0.3m and 0.4m. StreamVLN reaches 100%, 85% and 60%, with 0.2m, 0.3m and 0.5m. Thus the office gain corresponds to two additional successes out of 20 trials.
Where the evidence stops. No confidence intervals, per-robot breakdown or explicit real-world success-distance threshold are supplied here. The paper describes remote inference, RGB-D streaming, odometry and MPC; these bars do not establish an entirely onboard RGB-only deployment.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| R2R-CE unseen instruction-following Habitat VLN-CE, R2R validation-unseen; single-view RGB policy. | 64.3%; 58.5; 4.05m SR ↑; SPL ↑; NE ↓ | StreamVLN: 56.9%; 51.9; 4.98m. SR improves by 7.4 percentage points. Success means stopping within 3m; SPL penalizes long paths. This is simulated navigation. e-ce-results |
| RxR-CE unseen instruction-following Habitat VLN-CE, RxR validation-unseen. | 61.4%; 51.8; 4.58m SR ↑; SPL ↑; NE ↓ | StreamVLN: 52.9%; 46.0; 6.22m. SR improves by 8.5 percentage points. Sensor columns distinguish methods; training data and compute are not equalized. e-ce-results |
| Physical-simulation transfer VLN-PE, Unitree H1, R2R validation-unseen, physical locomotion controller; DualVLN transferred without VLN-PE finetuning. | 51.60%; 42.49; 12.32%; 2.23% SR ↑; SPL ↑; fall rate ↓; stuck rate ↓ | NaVid: 22.42%; 18.58; 8.61%; 0.45%. Higher completion coexists with higher fall and stuck rates than NaVid; this is not across-metric dominance. e-pe-results |
| Social-VLN dynamic-human navigation R2R validation-unseen with strategically placed moving humanoids. | 37.2%; 35.8; 35.4% SR ↑; SPL ↑; HCR ↓ | StreamVLN: 31.4%; 29.1; 36.4%. DualVLN gains 5.8 SR points but only 1.0 HCR point. Its SR falls 27.1 percentage points from static R2R, not 27.1% relatively. e-social-designe-social-results |
| Real-world hallway, bedroom and office navigation 20 trials per scenario per model; Figure 5. | Hallway: 100%, 0.2m; bedroom: 95%, 0.3m; office: 70%, 0.4m. SR ↑; NE ↓ | StreamVLN: 100%, 0.2m; 85%, 0.3m; 60%, 0.5m. Office improvement is 10 SR points in a small evaluation. Per-robot breakdowns and confidence intervals are not provided. e-real-results |
| Goal-interface ablation Section 5.3, Figure 7; metrics matching the main R2R-CE result, without a separately restated split in the caption. | DualVLN 64.3%; w/o Sys.2 Train 55.2%; w/o Pixel Goal 62.2%; w/o Latent Goal 60.9%. SR ↑ | Pixel-text removal costs 2.1 points; frozen pixel-token hidden states instead of learned queries cost 3.4 points. The one-stage variant also removes explicit intermediate goals, so its 9.1-point drop does not isolate training order. e-goal-ablation |
| Local-planner substitution VLN-PE R2R validation-unseen with the source-labeled flash controller; point-goal baselines receive depth-projected goals. | System 1: 63.62%; 56.49. SR ↑; SPL ↑ | NavDP: 58.72%; 50.98. iPlanner: 47.07%; 41.09. Planner changes also change conditioning and trajectory distributions. These values must not be pooled with Table 2 physical-controller results. e-local-planners |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 7. Both explicit goal text and learned latent queries matter, but the largest ablation changes several factors. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow one color through all four panels. Green is full DualVLN. Orange, labeled w/o Sys.2 Train, jointly trains both systems in one stage and also omits explicit intermediate pixel goals. Gray removes pixel-goal text from context before latent queries are appended, testing whether those queries benefit from the explicit prediction. Yellow replaces learned latent queries with frozen VLM hidden states from the generated pixel text; despite its name, it does not eliminate all latent conditioning. Higher SR, SPL and OS are better; lower NE is better. The red labels show changes from the full model, in percentage points for the rate metrics. e-goal-ablation
What it supports. Full DualVLN reaches 64.3% SR. Removing pixel text lowers it to 62.2%, while replacing learned queries lowers it to 60.9%; SPL also falls from 58.5 to 55.8 and 55.1. The one-stage, no-explicit-goal configuration performs worst at 55.2% SR and 51.5 SPL.
Where the evidence stops. The 9.1-point SR loss cannot isolate sequential training because explicit goal supervision also changes. Figure 7 does not separately restate the evaluation split or show uncertainty. The authors’ slower-convergence claim is not accompanied by a learning curve here.
7. Analysis & limitations
7.1 What the evidence leaves open
Local control tolerates small direction-preserving pixel errors, but large or semantically wrong goals remain failure cases, especially near obstacles. Scaling approaches saturation around 10% of System-2 trajectories; attributing this solely to a System-2 ceiling is the authors’ interpretation. e-diagnostics
Projected trajectory/goal agreement and appendix attention maps are descriptive diagnostics. They do not establish causal semantic reasoning or safe execution; collision results are the more direct dynamic-navigation test. e-diagnosticse-attentione-social-results
7.2 Questions for discussion
- Would fast RGB updates still help when goal age is controlled independently of VLM accuracy?
- Would sequential training retain its advantage if every compared model kept the same pixel-goal supervision?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the referenced training-data mixture, projection, interpolation, frozen-VLM query tuning, paired-frame encoder, DiT and evaluation controllers. Social data use a human-mask threshold and modified A* replanning, but the threshold value and detailed modifications are unspecified. e-goalse-training-globale-training-locale-social-design
Exact flow schedules, integration steps, waypoint interval and controller/network settings remain incomplete. Appendix pseudocode names noise prediction although Eqs. (2)–(3) train velocity; relative-pose encoding needs reconciliation with world-coordinate output wording. Separate ViT/Q-Former trainability is not explicit, and training hardware/software versions are unreported. e-flowe-training-locale-deployment
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: Separate sequential training from explicit-goal supervision
Reader-proposed check, not performed: run a matched factorial comparison of sequential versus joint backbone/policy training and pixel-goal supervision present versus absent. Keep initialization, data, model capacity, trajectory budget and evaluation episodes fixed; define the no-goal branch explicitly so token capacity is controlled. Within goal-supervised models, also mask the generated goal text only when extracting trajectory queries. Measure held-out pixel-goal accuracy, trajectory validation loss and navigation SR/SPL across multiple seeds. If sequential training helps only when goal supervision also changes, the original 9.1-point gap does not establish an independent benefit from training order. e-training-globale-training-locale-goal-ablation
Check 2: Test fresh RGB conditioning under controlled goal age
Reader-proposed check, not performed: hold the trained VLM and policy fixed, control how long each latent goal is reused, and compare normal paired-frame conditioning with a control that repeats the goal-time image in place of current RGB. Use identical simulated routes, humanoid schedules, execution controller and random seeds. Sweep imposed goal delays and record SR, HCR, goal-to-trajectory alignment and measured observation-to-command latency. A reliable advantage from current RGB at matched goal age would support the proposed adaptation mechanism; unchanged collision rates would weaken the claim that this branch explains dynamic avoidance. Report static routes separately to distinguish obstacle response from ordinary progress estimation. e-architecturee-social-designe-social-resultse-diagnosticse-deployment
8.3 Reading coverage
Visual audit: PDF page numbers are used throughout; supplementary printed pages 1–3 correspond to PDF pages 15–17. Page 1 supplied title, authors, affiliations, version and Figure 1; pages 2–5 supplied rationale, Figure 2, all equations, architecture and Social-VLN construction; pages 6–9 supplied every quantitative table, real-world protocol, Figures 4–7 and controller distinctions; page 10 supplied goal-error, scaling and consistency diagnostics; page 11 supplied InternVLA-N1 context; pages 15–17 supplied all appendix training settings and Figure 11. Figures 1–11 and Tables 1–4 were visually read. Six final original PNG crops were individually viewed; Table 1 and Table 2 retain short caption notes needed to interpret method markers. Figure 2’s conditioning arrows and concatenation marker were cross-checked against Section 3.2, Eqs. (1)–(3) and Appendix A.2. The source’s velocity/noise terminology, relative/world-coordinate wording, and overview/deployment timing differences remain disclosed in the base report. Reference-only pages 12–14 were read in the complete text chunks; separate video and linked resources were not viewed.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 15, 16, 17. 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
- Abstract (PDF p. 1)
- 1 Introduction (p. 2)
- 2 Related Work (pp. 2–3)
- 3 Method, Sections 3.1–3.3 and Eqs. (1)–(3) (pp. 3–5)
- 4 Social Vision-and-Language Navigation Benchmark (p. 5)
- 5.1 Simulation Experiments (pp. 6–7)
- 5.2 Real-World Cross-Embodiment Experiments (pp. 7–8)
- 5.3 Ablation Study and diagnostics (pp. 8–10)
- 6 Conclusion (p. 10)
- 7 Contributions and Acknowledgments (p. 11)
- References (pp. 11–14)
- Supplementary A.1 System 2 (PDF p. 15, supplement p. 1)
- Supplementary A.2 System 1 and B Attention Map Analysis (PDF pp. 16–17, supplement pp. 2–3)
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/version: the title and all 11 authors match the catalog author string. The artifact is arXiv:2512.08186v1, dated 9 December 2025 (e-identity). The catalog lists ICLR 2026; no separate venue edition was supplied, so equivalence to that edition is unverified. The catalog BibTeX says Peng, Peng; the title page verifies Jiaqi Peng.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by inspecting original PDF figures, tables, method and appendix pages.
- Separate supplemental material availability has not been fully verified. The three supplementary pages embedded in this PDF were read; the separately referenced video was not supplied or viewed.
- All five text chunks were read individually, without sampling. Figures 1–11 and Tables 1–4 were visually inspected. Reference-only PDF pages 12–14 were read as text.
- Code, linked models and datasets were not inspected; no experiments were reproduced. No external sources or later revisions were consulted.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author/affiliation block and arXiv margin
The observed title matches the supplied title; all 11 authors and four affiliations are printed. The margin identifies arXiv:2512.08186v1 [cs.RO], 9 Dec 2025.
Go to primary source ↓e-rationalePDF p. 2, Introduction, limitations of discrete actions and dual-system motivation
The authors motivate separate high-level planning and low-level execution by fragmented motion, VLM latency and dynamic obstacles.
Go to primary source ↓e-overviewPDF p. 1, Figure 1 and caption
The schematic labels a 7B VLM at 2Hz, diffusion policy at 30Hz and low-level controller at 200Hz; it does not provide a timing distribution.
Go to primary source ↓e-architecturePDF pp. 3–4, Figure 2 and Section 3.2
VLM context supplies learned latent goals; paired RGB frames at t and t+k pass through ViT, self-attention and a 32-token Q-Former. Concatenated conditions feed DiT cross-attention; output contains 32 dense waypoints.
Go to primary source ↓e-goalsPDF p. 4, Section 3.1; PDF p. 15 (supplement p. 1), A.1
Depth filters occluded projected waypoints; supervision chooses the farthest visible goal. View adjustment precedes grounding when needed. A.1 specifies turn chunks of at most four actions and STOP supervision.
Go to primary source ↓e-flowPDF pp. 4–5, Section 3.2, Eqs. (1)–(3)
Noisy trajectories combine ground-truth waypoints and Gaussian noise using unspecified decreasing/increasing schedules. DiT predicts velocity conditioned on latent goals and visual features under squared-error loss.
Go to primary source ↓e-training-globalPDF p. 5, Section 3.3; PDF p. 15 (supplement p. 1), A.1 Training
QwenVL-2.5-7B is fully finetuned following StreamVLN’s data recipe for one epoch. Appendix settings are AdamW, 2e-5 learning rate, batch 128 and 14,000 steps.
Go to primary source ↓e-training-localPDF p. 5, Section 3.3; PDF p. 16 (supplement p. 2), A.2
Specifications include four latent queries, DepthAnythingV2-Small ViT, DiT width 384/layers 12/heads 6 and 3584-to-768 projection. Stage 2 freezes QwenVL and trains queries/policy on interpolated 32-waypoint samples: AdamW 1e-4, batch 128, 15,000 steps. Pseudocode says noise_pred and relative poses; prose says world-coordinate trajectories.
Go to primary source ↓e-social-designPDF p. 5, Section 4 and Figure 3
Social-VLN places Habitat 3.0 humanoids along R2R-CE routes and checks passability. Human-mask-triggered modified A* collection yields 763K episodes across 60 MP3D scenes; the threshold value is unstated.
Go to primary source ↓e-ce-resultsPDF p. 6, Table 1, DualVLN and StreamVLN rows; Section 5.1 metrics
R2R SR/SPL/NE are 64.3/58.5/4.05 versus 56.9/51.9/4.98; RxR gives 61.4/51.8/4.58 versus 52.9/46.0/6.22. Both splits are validation-unseen; SR uses a 3m stopping threshold. Observation columns distinguish sensor regimes.
Go to primary source ↓e-pe-resultsPDF p. 6, VLN-PE Benchmark & Metrics; p. 7, Table 2 and preceding paragraph
Unitree H1 physical simulation evaluates zero-shot transfer. Unseen DualVLN SR/SPL/FR/StR are 51.60/42.49/12.32/2.23 versus NaVid 22.42/18.58/8.61/0.45. The table separately labels VLN-PE-trained baselines.
Go to primary source ↓e-social-resultsPDF p. 7, Table 3 and Social-VLN Experiment
Social-VLN DualVLN versus StreamVLN SR is 37.2 versus 31.4, SPL 35.8 versus 29.1, HCR 35.4 versus 36.4. Static SR is 64.3 versus 56.9. Prose describes approximately 27%/26% drops, but the table supports percentage-point differences.
Go to primary source ↓e-deploymentPDF p. 7, Section 5.2 Experimental Setup
Turtlebot4, Go2 and G1 use downward-angled D455 cameras, remote RTX 4090 inference occupying 20GB, RGB-D streaming, odometry and MPC. KV-cache changes 1.1s to 0.7s; TensorRT System 1 is described as generating 32 trajectories in 0.03s.
Go to primary source ↓e-real-resultsPDF p. 8, Figure 5 and Section 5.2 Quantitative Analysis; p. 9, Figure 6
Twenty trials per scenario per model are reported. DualVLN hallway/bedroom/office SR is 100/95/70 and NE 0.2/0.3/0.4m; StreamVLN has 100/85/60 and 0.2/0.3/0.5m. Figure 6 shows selected qualitative comparisons, without per-embodiment quantitative breakdowns or intervals.
Go to primary source ↓e-goal-ablationPDF pp. 8–9, Section 5.3; p. 9, Figure 7 and caption
Full/one-stage/no-pixel-text/frozen-pixel-hidden-state SR is 64.3/55.2/62.2/60.9; SPL is 58.5/51.5/55.8/55.1. The one-stage variant both jointly trains systems and removes explicit intermediate pixel goals.
Go to primary source ↓e-local-plannersPDF p. 9, Table 4 caption, unseen columns and following paragraph
Under the flash controller, System 1/NavDP/iPlanner have unseen SR 63.62/58.72/47.07 and SPL 56.49/50.98/41.09. Alternatives remove latent conditioning and project pixel goals through additional depth; the authors discuss distribution mismatch.
Go to primary source ↓e-diagnosticsPDF p. 10, Figures 8–10 and Section 5.3 continuation
Figure 8 shows tolerance to small direction-preserving errors but failures for large or semantically incorrect goals. The authors describe near-saturation around 10% training trajectories. Figure 10 uses 1000 random samples to compare pixel distances and angular deviations.
Go to primary source ↓e-attentionPDF p. 16, Appendix B; PDF p. 17, Figure 11
Language and visual attention examples at layers 6, 15 and 24 are interpreted as progressing toward the goal region and emphasizing STOP at deeper layers. These are visualizations rather than intervention experiments.
Go to primary source ↓e-resource-contextPDF p. 1, resource links; p. 11, Section 7
The title page names InternNav, InternVLA-N1 and InternData-N1. Section 7 describes an elaboration with improvements to the dual-system VLN component within InternVLA-N1; linked artifacts were not independently inspected.
Go to primary source ↓8.5 Primary sources
Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation ↗
PDF · 8,204 extracted words
Source fingerprint
c535e225421e186c8553ae3930cd4c4f5f787c0a5030da56c3bb74d200c50768