acceptodds
Under review as a conference paper at ICLR 2027

GroundingVLN: Reasoning and Acting with Grounding for Vision-Language Navigation

Abstract

Although vision-language models (VLMs) possess strong visual understanding and reasoning capabilities, vision-and-language navigation (VLN) agents built upon them still struggle to bridge semantic reasoning and spatial execution. This disconnection stems from two coupled limitations: intermediate reasoning is not explicitly grounded in visual evidence, and high-level decisions lack precise spatial goals to guide low-level motion. Inspired by human navigation, which naturally links cognition and locomotion through landmarks and spatial goals, we propose **GroundingVLN**, a framework that uses visual grounding as a shared interface between reasoning and action. GroundingVLN first reasons with grounding by anchoring task-relevant visual evidence to precise image locations throughout structured reasoning. It then acts through grounding by predicting a progress-aligned pixel goal that a geometric planner translates into primitive actions. To learn these capabilities, we construct **GroundingCOTVLN-188K**, a dataset of temporally aligned grounded reasoning traces, and introduce **Grounded and Execution-Aware Reinforcement Learning (GEAR)**, which aligns grounded reasoning and spatial decisions with downstream execution. Experiments demonstrate that GroundingVLN achieves state-of-the-art performance (**69.9%** SR on R2R-CE and **75.1%** SR on RxR-CE) with superior data efficiency, requiring only **0.9%** of the training data used by the strongest baseline. It also generalizes strongly across datasets, attaining **59.9%** SR on RxR-CE when trained solely on R2R, a gain of **20.1%** over the strongest baseline. The full code, dataset, and model will be released after review.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language navigation, grounding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eiGROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

GroundingVLN addresses the challenge of Vision-and-Language Navigation (VLN) in continuous environments by employing a shared "grounding" interface that bridges structured reasoning with spatial execution. The framework integrates coordinate-based visual evidence with pixel-level goals, which are dynamically adjusted as navigation progresses and executes these goals via an RGB-D geometric planning module. The authors constructed a dataset, GroundingCOTVLN-188K, which features temporally aligned grounding annotations, and introduced GEAR: a reinforcement learning phase based on a "group-relative" mechanism that combines "execution-aware" goal rewards with supervisory signals for actions, subtasks, and grounding. Empirical validation includes evaluations on the R2R-CE and RxR-CE benchmarks, cross-dataset transfer, component ablation studies, real-world deployment, and supplementary analyses regarding reasoning efficiency and the impact of depth noise.

Critical strengths
  1. The design specifically tailored for "grounding" has been validated through comprehensive component ablation studies (Table 3). These comparative experiments provide compelling evidence that both the grounding-based supervision mechanism and the "execution-aware" optimization strategy contribute to improved navigation performance, thereby confirming the technical value of the training design proposed in this paper.
  2. This method demonstrates robust navigation performance and strong cross-dataset transferability. On the "unseen" validation splits (val_unseen) of R2R-CE and RxR-CE, the method outperform ABot-N0 by 3.5 and 5.8 percentage points. When trained exclusively on R2R data, the method attains a 59.9% success rate on RxR-CE, compared to only 39.8% for AwareVLN. These results confirm the effectiveness of the complete system and its capacity for generalization.
Critical weaknesses

The "progress-aligned pixel-goal selection" method lacks an ablation study that isolates this specific component. Section 3.3 notes that restricting goal candidates to the current subtask segment is the key distinction between this method and "farthest-visible-point-based supervision." However, Table 3 does not include a variant in which only this "semantic boundary constraint" is removed. In the "w/o Subtask Decomposition" ablation, the paper does not explicitly clarify how goal segmentation and subtask rewards are adjusted, making it difficult to disentangle the contribution of "goal selection" from other effects of the decomposition. This, to some extent, weakens the empirical support for the study's core contribution.

Questions for authors
  1. In the "w/o Subtask Decomposition" variant, how are the target-segment constraints and subtask rewards handled? Could the authors isolate the effect of "progress alignment" by removing only the boundary constraints for the current subtask—while keeping the remaining target-construction constraints, training procedure, and execution pipeline intact?
  2. Can the training-data comparison be expressed using a common counting unit and comparable task coverage? Otherwise, please explicitly qualify the 0.9% figure as a ratio of heterogeneous reported training counts.
3 comments
AuthorsOpenReview ↗

W1&Q1: Independent progress-alignment ablation

We thank the reviewer for the feedback. Our responses are provided below.

W1&Q1: Independent progress-alignment ablation

We have added an independent w/o current-subtask boundary ablation to Table 3 in the main paper and expanded the analysis in Section 4.3.

Original decomposition ablation. w/o Subtask Decomposition removes only the initial decomposition step at inference, with unchanged model weights and no retraining. The checkpoint was trained with the original progress-aligned targets and subtask rewards; the remaining reasoning format is unchanged.

Independent boundary ablation. To isolate progress alignment, we retain the original reasoning annotations, subtask labels and reward, all other target-construction constraints, the SFT + GEAR training procedure, and the executor. We replace subtask-bounded pixel goals with the farthest visible path points, update target-related annotations and goal-reward maps consistently, and retrain the model.

Table R2. Progress-alignment ablation on R2R-CE val_unseen.

VariantNE ↓ (m)SR ↑ (%)OSR ↑ (%)SPL ↑ (%)
Full model3.6669.974.864.1
w/o Subtask Decomposition4.0568.174.962.2
w/o current-subtask boundary4.6458.274.547.3

As shown in Table R2, removing the target boundary leaves OSR nearly unchanged (74.8% vs. 74.5%) but lowers SR from 69.9% to 58.2%, a drop of 11.7 percentage points. Case analysis shows that the additional failures mainly involve farther goals beyond the destination or continued motion after arrival. The agent can still enter the goal vicinity but fails to stop reliably, supporting the role of progress alignment in judging task boundaries.

AuthorsOpenReview ↗

Q2: Comparable training-data counts

Q2: Comparable training-data counts

We agree that the original 0.9% figure combines heterogeneous counting units. We report the trajectory-count comparison alongside comparisons of instruction-following training samples and reasoning examples:

Comparison basisGroundingCOTVLN-188KABot-N0 (Chu et al., 2026)Our count / ABot-N0
Training trajectories≈21K R2R/RxR trajectories16.9M reported trajectories≈0.12%
Instruction-following training samples188K R2R/RxR decision-level examples2.8M instruction-following samples≈6.71%
Reasoning examples188K grounded image–CoT examples5.0M additional reasoning examples3.76%

Our 188K decision-level examples come from approximately 21K R2R/RxR trajectories, include the 16K GEAR subset. ABot-N0's 2.8M instruction-following samples cover R2R/RxR, door traversal, person search, and short-horizon navigation (Section 3.2.2); these are clip-derived training instances, while its 5.0M reasoning corpus has broader coverage. Using the reported trajectory counts gives 21K / 16.9M ≈ 0.12%. These ratios therefore describe reported counts with different task coverage and supervision.

Dear Reviewer,

I hope this message finds you well. With less than a day left in the discussion period, I wanted to kindly check whether our responses have addressed your concerns, or whether there are any remaining points you would like us to clarify. We would be happy to respond promptly. If our response has been satisfactory, a brief confirmation would be very helpful.

Thank you again for your time and for reviewing our paper.

Reviewer_HeH2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GroundingVLN, a framework for vision-and-language navigation in continuous environments. A high-level VLM performs structured reasoning, associates instruction-relevant visual evidence with image coordinates, and predicts a high-level action and a pixel goal aligned with the current instruction subtask. A low-level planning module converts goals into executable actions.

The paper introduces GroundingCOTVLN-188K, containing grounded reasoning and decision supervision derived from R2R/RxR trajectories. Its annotation pipeline combines subtask–trajectory alignment, geometric target construction, and temporal grounding alignment. Following supervised fine-tuning, GEAR applies GRPO with action, subtask, grounding, and geometry-based target rewards.

GroundingVLN achieves SOTA on the two main benchmarks.

Critical strengths
  1. The experiments are comprehensive. The paper evaluates on both R2R-CE and RxR-CE, and includes cross-dataset transfer, component ablations, and real-world robot experiments.

  2. The method is well motivated and achieves SOTA. Using grounded visual evidence to support reasoning provides a clear connection between understanding the instruction and deciding where to move. The resulting framework achieves the best performance among the strong baselines on R2R-CE and RxR-CE.

  3. GroundingCOTVLN-188K is a useful contribution. The dataset connects instruction progress, grounded visual evidence, and executable pixel goals. The temporal-alignment ablation supports the value of this annotation design.

Critical weaknesses
  1. The main comparison does not adequately separate the contribution of the proposed high-level perception from low level executor. Although DualVLN (S2)+SPF provides a useful partial control, the absence of a comparison using the same executor limits attribution of the performance gains to the proposed perception methods. Lack an ablation experiment about the executor. This makes it unclear whether the performance of the method mainly depends on the executor or the grounding method.

  2. The training-data efficiency claim is stronger than the evidence supports. Appendix A.1 explicitly compares sample counts in their original units: the proposed method uses decision-level examples, while ABot-N0 combines trajectories and reasoning samples across several navigation tasks. The 0.9% ratio therefore does not provide a directly comparable measure of data efficiency. Appendix A.2 usefully compares methods trained without additional navigation corpora, but differences in backbone, supervision, and execution remain. The results show excellent performance with a relatively small data, but this small dataset requires a lot of additional information annotation.

Questions for authors
  1. Ablation experiment about executor. Could you provide the ablation experiments for the executor to study its impact on performance? The executor can be the same as the baseline that uses pixel-based planning methods.

  2. Clarification of the ablation: What exactly changes in “w/o Subtask Decomposition”? Does this variant replace progress-aligned targets with farthest-visible targets, or does it also change the reasoning format and reward computation?

  3. Are all the real world results obtained using the checkpoints from simulation training, without using any extra real world demonstrations, real world finetuning, or adaptations for the testing environment? If so, please specify the data size, collection method, and training process.

  4. Clarification for the data-efficiency claim: Why is your method's "TRAINING-DATA EFFICIENCY" so high? An analysis of the reasons for thi is needed. For instance, when compared with Abot-N0, your method differs in many aspects: the VLM backbone, the executor, the training stages, the data corpus, etc. Which of these factors have a greater impact on data efficiency?

8 comments
AuthorsOpenReview ↗

W1&Q1: Executor ablation

We thank the reviewer for the feedback. Our responses are provided below.

W1&Q1: Executor ablation

Executor ablation

We have added this executor ablation to Appendix A.3 of the revised paper.

Table R1(b). GroundingVLN executor ablation on R2R-CE val_unseen.

ExecutorS.RGBPano.DepthOdo.NE ↓ (m)OSR ↑ (%)SR ↑ (%)SPL ↑ (%)
RTAB-Map + A*✓✓3.6674.869.964.1
RGB-only VGGT + PnP✓3.8972.868.162.7
Habitat SPF✓✓✓3.5475.270.364.4

The core of GroundingVLN is a VLM for high-level decision-making that requires only RGB observations as visual input. Depth and pose requirements depend on the action executor. To compare navigation performance under the same RGB-only input setting, we implemented a VGGT + PnP executor. It predicts depth with VGGT (Wang et al., 2025), recovers metric scale using camera calibration and visible ground, and estimates pose via ORB matching and PnP with keyframe relocalization. These estimates enable mapping, pixel-goal back-projection, planning, and path following without measured depth or ground-truth pose.

We compare RGB-only, RGB-D, and idealized Habitat SPF executors on R2R-CE val_unseen. As shown in Table R1(b), replacing the default executor with VGGT + PnP reduces SR/SPL by 1.8/1.4 percentage points, while SPF improves them by 0.4/0.3 points. With VGGT + PnP, approximately 1% of all evaluated episodes fail because depth-estimation errors disrupt path planning. SPF further removes minor errors introduced by RTAB-Map pose estimation, yielding a slight increase in SR. The policy therefore remains effective across all three executors.

Cross-method comparisons

We further compare GroundingVLN with baselines sharing the same executor, three-view RGB setting, or pixel-goal interface.

Table R1(c). Cross-method comparisons by shared setting. SR / SPL (%) are reported on val_unseen. Blank performance cells indicate unevaluated variants.

Shared settingMethod / executorS.RGBPano.DepthOdo.R2R-CE SR / SPL ↑RxR-CE SR / SPL ↑
Same SPF executorDualVLN (S2) + SPF (Wei et al., 2026)✓✓✓60.9 / 55.263.5 / 55.0
Same SPF executorGroundingVLN with SPF✓✓✓70.3 / 64.4
Three-view RGBABot-N0 (Chu et al., 2026)✓66.4 / 63.969.3 / 60.0
Three-view RGBGroundingVLN with RTAB-Map✓✓69.9 / 64.175.1 / 62.0
Three-view RGBGroundingVLN with VGGT + PnP✓68.1 / 62.7
Pixel-goal interfaceGoal2Pixel-2B (Bao et al., 2026)✓✓✓54.1 / 52.542.7 / 39.5
Pixel-goal interfaceGoal2Pixel-8B (Bao et al., 2026)✓✓✓53.9 / 52.748.1 / 44.7
Pixel-goal interfaceGroundingVLN with RTAB-Map✓✓69.9 / 64.175.1 / 62.0

Same executor. As shown in Table R1(c), GroundingVLN with SPF achieves 70.3% SR / 64.4% SPL, versus 60.9% / 55.2% for DualVLN System 2 + SPF (Wei et al., 2026): gains of 9.4/9.2 points.

Same three-view input. Against ABot-N0 (Chu et al., 2026), GroundingVLN with RTAB-Map improves R2R-CE SR/SPL by 3.5/0.2 points and RxR-CE by 5.8/2.0 points. With the RGB-only executor, GroundingVLN reaches 68.1% SR on R2R-CE, compared with 66.4% for ABot-N0 (Chu et al., 2026). Thus, the advantage persists without measured depth.

Same pixel-goal interface. Goal2Pixel (Bao et al., 2026) also predicts image-space goals for geometric execution. Compared with Goal2Pixel-8B (Bao et al., 2026), GroundingVLN with RTAB-Map improves SR and SPL by 16.0 and 11.4 percentage points on R2R-CE, and by 27.0 and 17.3 points on RxR-CE, respectively.

Attribution to visually grounded reasoning. The advantage persists with the same SPF executor and with three-view RGB-only inputs, reducing the likelihood that it comes solely from execution or extra sensing. Within GroundingVLN, removing grounding tags lowers SR from 69.9% to 66.1% (Table 3 in the main paper), while the no-CoT SFT + GEAR variant reaches 61.7%. Together, these results support a contribution from CoT with Grounding, which connects instruction-relevant visual evidence to landmark localization and point-goal prediction.

AuthorsOpenReview ↗

Q2: Subtask-decomposition ablation

Q2: Subtask-decomposition ablation

We have added the w/o current-subtask boundary ablation to Table 3 in the main paper and expanded the analysis of both ablations in Section 4.3.

We clarify that w/o Subtask Decomposition removes only the initial subtask-decomposition step at inference. The model weights remain unchanged, and no retraining is performed.

For w/o current-subtask boundary, we retain the original reasoning annotations, subtask labels and reward, and other geometric constraints, but replace subtask-bounded pixel goals with the farthest visible path points. We update target-related annotations and reward maps consistently, and retrain using the same SFT + GEAR procedure and executor.

Table R2. Progress-alignment ablation on R2R-CE val_unseen.

VariantNE ↓ (m)SR ↑ (%)OSR ↑ (%)SPL ↑ (%)
Full model3.6669.974.864.1
w/o Subtask Decomposition4.0568.174.962.2
w/o current-subtask boundary4.6458.274.547.3

As shown in Table R2, removing the pixel-goal boundary leaves OSR nearly unchanged (74.8% vs. 74.5%) but lowers SR from 69.9% to 58.2%, a drop of 11.7 percentage points. Our case analysis shows that the additional failures mainly involve much farther predicted goals: the model often selects points beyond the destination or continues moving after arrival. The agent can therefore still enter the goal vicinity but fails to stop there reliably, explaining the similar OSR despite lower SR. Aligning pixel goals with task progress thus improves the model's judgment of task boundaries.

AuthorsOpenReview ↗

Q3: Real-world training and adaptation

Q3: Real-world training and adaptation

All real-world results use the simulation-trained checkpoints, with SFT on R2R/RxR and GEAR on the 16K-example subset. We use no real-world demonstrations, additional real-world training examples, or environment-specific fine-tuning. These results support GroundingVLN's ability to generalize from simulation to real-world navigation without additional training.

Reviewer_u8yEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GroundingVLN generates structured, visually grounded reasoning and pixel-goal actions with a Qwen3.5-4B model. Goals are back-projected using depth and executed through RTAB-Map and A*. Training combines teacher-annotated grounded reasoning data with GEAR, a GRPO procedure using execution-aware goal rewards. The paper reports strong R2R-CE/RxR-CE results, cross-dataset transfer and robot evaluation.

Critical strengths
  1. Grounded reasoning and pixel goals offer an interpretable interface between instruction understanding and execution. The data construction aligns landmarks over time, while the execution-aware reward addresses the limitations of measuring only 2D goal distance. The ablation evidence makes this a promising design.

  2. The reported benchmark and transfer results are strong. The grounding/action analysis and robot study provide useful additional evidence beyond aggregate navigation success.

Critical weaknesses
  1. Sensor and execution differences make the main comparison difficult to interpret. The proposed agent uses three-view RGB-D observations, pose estimation, a point-cloud map and an A* executor, whereas several listed baselines use monocular RGB and low-level actions. Without explicit input/executor annotations and matched comparisons, the gain cannot be attributed cleanly to grounded reasoning. This is the main reason I currently place the paper on the weak-rejection side of borderline.

  2. Supervision accounting and component attribution need clarification. The 0.9% data comparison mixes teacher-distilled samples with baseline corpora of different types and omits the cost of the 397B teacher and tracking pipeline. GEAR contributes a large portion of the reported gain, so a non-grounded pixel-goal policy trained with the same reward would better isolate grounded CoT. Missing seed variation, runtime measurements and executor failure statistics also limit confidence in the practical conclusions.

Questions for authors
  1. Please annotate each baseline's views, depth, pose/odometry and executor. Which comparisons match the proposed input and control setup? Does the turn primitive follow the official action protocol, and what evidence separates gains from extra sensing/mapping from gains due to grounded reasoning?

  2. How are the 16K GEAR anchors chosen, and which rewards use reference paths or simulator geometry? Can the same reward be applied to a pixel-goal-only policy without grounded CoT under matched supervision and compute?

  3. Please compare supervision using trajectories/episodes as well as generated samples, and account for teacher annotation. What are the inference latency, episode wall-clock cost and mapping/executor failure rates? Were robot baselines deployed with the same sensors and executor?

12 comments
AuthorsOpenReview ↗

W1&Q1: Inputs, executor, and attribution

We thank the reviewer for the feedback. Our responses are provided below.

W1&Q1: Inputs, executor, and attribution

Baseline inputs and dependencies

We distinguish high-level policy inputs from geometry used for waypoint generation, mapping, and control. Our VLM receives three RGB views, the instruction, and history. Depth and online pose estimates are used by the default executor.

Table R1(a). Baseline observations and performance. Results follow Table 1 in the main paper, with Goal2Pixel (Bao et al., 2026) added for comparison. Scores are SR / SPL (%) on the reported val_unseen settings. S.RGB denotes single-view RGB; Pano. denotes panoramic or multi-view RGB. Depth and Odo. indicate externally supplied depth and pose/odometry used in the navigation pipeline; internally estimated geometry is excluded.

MethodS.RGBPano.DepthOdo.R2R-CE SR / SPL ↑RxR-CE SR / SPL ↑
HPN+DN (Krantz et al., 2021)✓✓✓36.0 / 34.0—
CMA (waypoint-based) (Hong et al., 2022)✓✓✓41.0 / 36.026.5 / 22.1
GridMM (Wang et al., 2023)✓✓✓49.0 / 41.0—
ETPNav (An et al., 2025)✓✓✓57.0 / 49.054.7 / 44.8
ScaleVLN (HAMT) (Wang et al., 2023)✓✓✓55.0 / 51.0—
InstructNav (Long et al., 2025)✓✓✓31.0 / 24.0—
SmartWay (Shi et al., 2025)†✓✓29.0 / 22.5—
Open-Nav (Qiao et al., 2025)†✓✓19.0 / 16.1—
HSGM (Li et al., 2026)‡✓✓✓47.9 / 32.841.8 / 25.1
VLN-R1 (Qi et al., 2025)✓30.2 / 21.822.7 / 17.6
NaVid (Zhang et al., 2024)✓37.4 / 35.9—
NaVILA (Cheng et al., 2025)✓54.0 / 49.049.3 / 44.0
Uni-NaVid (Zhang et al., 2025)✓47.0 / 42.748.7 / 40.9
StreamVLN (Wei et al., 2025)✓56.9 / 51.952.9 / 46.0
NavFoM (Zhang et al., 2026)✓61.7 / 55.364.4 / 56.2
MapDream (Lian et al., 2026)✓59.8 / 54.459.4 / 49.2
DualVLN (S2) + SPF (Wei et al., 2026)✓✓✓60.9 / 55.263.5 / 55.0
DualVLN (Wei et al., 2026)✓64.3 / 58.561.4 / 51.8
JanusVLN (Zeng et al., 2026)✓60.5 / 56.856.2 / 47.5
AwareVLN (Guo et al., 2026)✓65.4 / 55.167.6 / 56.1
ABot-N0 (Chu et al., 2026)✓66.4 / 63.969.3 / 60.0
Goal2Pixel-2B (Bao et al., 2026)✓✓✓54.1 / 52.542.7 / 39.5
Goal2Pixel-8B (Bao et al., 2026)✓✓✓53.9 / 52.748.1 / 44.7
GroundingVLN✓✓69.9 / 64.175.1 / 62.0

Executor ablation

We have added this executor ablation to Appendix A.3 of the revised paper.

Table R1(b). GroundingVLN executor ablation on R2R-CE val_unseen.

ExecutorS.RGBPano.DepthOdo.NE ↓ (m)OSR ↑ (%)SR ↑ (%)SPL ↑ (%)
RTAB-Map + A*✓✓3.6674.869.964.1
RGB-only VGGT + PnP✓3.8972.868.162.7
Habitat SPF✓✓✓3.5475.270.364.4

The core of GroundingVLN is a VLM for high-level decision-making that requires only RGB observations as visual input. Depth and pose requirements depend on the action executor. To compare navigation performance under the same RGB-only input setting, we implemented a VGGT + PnP executor. It predicts depth with VGGT (Wang et al., 2025), recovers metric scale using camera calibration and visible ground, and estimates pose via ORB matching and PnP with keyframe relocalization. These estimates enable mapping, pixel-goal back-projection, planning, and path following without measured depth or ground-truth pose.

We compare RGB-only, RGB-D, and idealized Habitat SPF executors on R2R-CE val_unseen. As shown in Table R1(b), replacing the default executor with VGGT + PnP reduces SR/SPL by 1.8/1.4 percentage points, while SPF improves them by 0.4/0.3 points. With VGGT + PnP, approximately 1% of all evaluated episodes fail because depth-estimation errors disrupt path planning. SPF further removes minor errors introduced by RTAB-Map pose estimation, yielding a slight increase in SR. The policy therefore remains effective across all three executors.

AuthorsOpenReview ↗

W1&Q1: Inputs, executor, and attribution

Cross-method comparisons

We further compare GroundingVLN with baselines sharing the same executor, three-view RGB setting, or pixel-goal interface.

Table R1(c). Cross-method comparisons by shared setting. SR / SPL (%) are reported on val_unseen; observation columns follow Table R1(a). Blank performance cells indicate unevaluated variants.

Shared settingMethod / executorS.RGBPano.DepthOdo.R2R-CE SR / SPL ↑RxR-CE SR / SPL ↑
Same SPF executorDualVLN (S2) + SPF (Wei et al., 2026)✓✓✓60.9 / 55.263.5 / 55.0
Same SPF executorGroundingVLN with SPF✓✓✓70.3 / 64.4
Three-view RGBABot-N0 (Chu et al., 2026)✓66.4 / 63.969.3 / 60.0
Three-view RGBGroundingVLN with RTAB-Map✓✓69.9 / 64.175.1 / 62.0
Three-view RGBGroundingVLN with VGGT + PnP✓68.1 / 62.7
Pixel-goal interfaceGoal2Pixel-2B (Bao et al., 2026)✓✓✓54.1 / 52.542.7 / 39.5
Pixel-goal interfaceGoal2Pixel-8B (Bao et al., 2026)✓✓✓53.9 / 52.748.1 / 44.7
Pixel-goal interfaceGroundingVLN with RTAB-Map✓✓69.9 / 64.175.1 / 62.0

Same executor. As shown in Table R1(c), GroundingVLN with SPF achieves 70.3% SR / 64.4% SPL, versus 60.9% / 55.2% for DualVLN System 2 + SPF (Wei et al., 2026): gains of 9.4/9.2 points.

Same three-view input. Against ABot-N0 (Chu et al., 2026), GroundingVLN with RTAB-Map improves R2R-CE SR/SPL by 3.5/0.2 points and RxR-CE by 5.8/2.0 points. With the RGB-only executor, GroundingVLN reaches 68.1% SR on R2R-CE, compared with 66.4% for ABot-N0 (Chu et al., 2026). Thus, the advantage persists without measured depth.

Same pixel-goal interface. Goal2Pixel (Bao et al., 2026) also predicts image-space goals for geometric execution. Compared with Goal2Pixel-8B (Bao et al., 2026), GroundingVLN with RTAB-Map improves SR and SPL by 16.0 and 11.4 percentage points on R2R-CE, and by 27.0 and 17.3 points on RxR-CE, respectively.

Attribution to visually grounded reasoning. The advantage persists with the same SPF executor and with three-view RGB-only inputs, reducing the likelihood that it comes solely from execution or extra sensing. Within GroundingVLN, removing grounding tags lowers SR from 69.9% to 66.1% (Table 3 in the main paper), while the no-CoT SFT + GEAR variant reaches 61.7% (Table R3). Together, these results support a contribution from CoT with Grounding, which connects instruction-relevant visual evidence to landmark localization and point-goal prediction.

AuthorsOpenReview ↗

Q1: TURN implementation

Q1: TURN implementation

We define high-level TURN actions as left/right rotations of 90° or 180°. Similar coarse turning actions are used in HSGM (Li et al., 2026) (90°/180°), SmartWay (Shi et al., 2025) (e.g., 90°), and VLMnav (Goetting et al., 2025) (180°). The official VLN-CE configuration uses 15° turn primitives; we execute each 90° or 180° turn as six or twelve consecutive 15° turns, respectively.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To all reviewers

We thank all reviewers for their careful reading and constructive feedback. We appreciate the recognition that grounded reasoning and pixel goals provide an interpretable link between instruction understanding and navigation decisions (Reviewers u8yE and HeH2). All three reviewers noted the strong benchmark performance, and Reviewers u8yE and eiGR also highlighted the transfer results. We further appreciate the positive assessments of GroundingCOTVLN-188K and its temporal grounding annotations (Reviewer HeH2), the ablation evidence for grounded supervision and execution-aware optimization (Reviewer eiGR), and the experimental validation, including real-world studies (Reviewers u8yE and HeH2).

In the following responses, we address concerns about sensor and executor differences, component attribution, training-data accounting, and practical evaluation. We present additional executor and progress-alignment ablations, no-CoT comparisons, and results across evaluation seeds, and clarify teacher supervision, inference runtime, and real-world deployment. The revised manuscript includes executor ablations in Appendix A.3, the independent progress-alignment ablation and analysis in Table 3 and Section 4.3, and clarified data-counting units in Appendix A.1. We appreciate the reviewers' time and suggestions, which have helped us improve the evidence and presentation of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.