STRIDER: Navigation via Instruction-Aligned Structural Decision Space Optimization

Weakness 1:

Thank you for the thoughtful comment. To disentangle the contribution of STRIDER’s framework from the underlying pretrained models, we conduct controlled comparisons using the same LLM and VLMs across STRIDER and baseline methods. As shown in Table 1, STRIDER outperforms the other two baselines under identical LLM and VLM settings, confirming that the performance gain stems from our framework design.

Table 1: Comparison Under Identical VLM and LLM

Weakness 2:

Thank you for raising this important point. We agree that the robustness of STRIDER depends in part on the reliability of its underlying VLM and LLM components. To evaluate the impact of incorrect feedback, we conduct a controlled experiment where we randomly corrupt the progress feedback provided to the LLM during reasoning. As shown in Table 2, this experiment offers a concrete quantification of STRIDER’s tolerance to feedback noise and supports its practical usage even in imperfect settings.

Table 2: Robustness to incorrect feedback

Question 1:

STRIDER introduces a per-step latency of approximately 34.2754 seconds, only slightly higher than Open-Nav’s 32.1105 seconds. In terms of overall runtime, STRIDER takes 3.9074 minutes per episode, while Open-Nav takes 3.6606 minutes. Notably, STRIDER’s waypoint generation is faster, requiring only 0.4001 seconds compared to 0.8892 seconds for Open-Nav. These results demonstrate that STRIDER remains efficient and practical for real-world use.

Question 2:

We clarify that the capabilities of the underlying VLM/LLM are fundamental to the success of STRIDER.

Our method is designed to leverage and amplify the strengths of powerful models like GPT-4o, which help improve decision robustness. This makes the method less sensitive among models of similar capability, e.g., between GPT-4 and GPT-4o.

We present ablation results in Table 3 using different VLMs. The results demonstrate that while stronger models do yield better performance, STRIDER maintains competitive results even with smaller or different models, confirming that the design generalizes well across models of similar capacities and does not rely on any single pretrained model.

Table 3: Ablation on different VLMs

Question 3:

We have indeed observed occasional cases where the skeleton graph misses viable paths or where feedback from the VLM introduces minor errors.

To mitigate these issues, we enforce a maximum step-size limit of 1.5 meters during navigation, which ensures fine-grained control and prevents large deviations. As a result, even if an error occurs at one step, the agent does not stray far from the intended path and can often correct itself in subsequent steps after gaining a broader view.

Our robustness evaluation, shown in Table 2 in the supplemental material, further supports this claim: when subjected to mid-trajectory perturbations that displace the agent from its planned waypoint, STRIDER demonstrates better recovery.

Thank you for the comment. We would like to clarify that sensitivity to model capacity is expected and inherent in any zero-shot method based on frozen VLMs and LLMs. This is not unique to STRIDER; it is a general limitation of the paradigm. Our goal is not to eliminate this dependency, but to bring out the strengths of powerful models while ensuring that weaker models still yield usable results.

As demonstrated in Table 1 from Open-Nav[1] and Table 2 from CA-Nav[2], even with an identical architecture, different LLMs can lead to large differences. Our results in Table 3 of the rebuttal further support that changing only the VLM (perception) leads to performance differences. This reflects a well-understood fact in the field: zero-shot methods rely heavily on the quality of the pretrained models.

Table 1: Comparison in Open-Nav[1]

Table 2: Comparison in CA-Nav[2]

What we emphasize is that STRIDER remains robust and competitive across model scales. The system remains functional even with smaller models. Under fair comparison across models, the consistent performance gains of STRIDER sufficiently demonstrate the strength of our contribution. We hope this provides further clarity and addresses the reviewer’s concerns.

Reference:

[1] Yanyuan Qiao, Wenqi Lyu, Hui Wang, Zixu Wang, Zerui Li, Yuan Zhang, Mingkui Tan, and Qi Wu. Open-nav: Exploring zero-shot vision-and-language navigation in continuous environment with open-source llms. arXiv preprint arXiv:2409.18794, 2024.

[2] Kehan Chen, Dong An, Yan Huang, Rongtao Xu, Yifei Su, Yonggen Ling, Ian Reid, and Liang Wang. Constraint-aware zero-shot vision-language navigation in continuous environments. arXiv preprint arXiv:2412.10137, 2024.

Weakness 1.1:

We select to use Qwen-VL-Max (latest) because it is the most recent (updated on 25-04) and stable release. In our experiments, it produces clear, concise descriptions under our prompt design, showing comparable performance with other VLMs.

For the LLM, although most baselines adopt GPT-4, we choose GPT-4o due to its more consistent formatting, reduced verbosity, faster and more stable response under API usage.

Weakness 1.2:

We would like to clarify that STRIDER is fundamentally architecture-agnostic. Its design centers around a structured planning-regulation pipeline that can operate with a variety of models.

We present ablation results in Table 1 using different VLMs. The results demonstrate that while stronger models do yield better performance, STRIDER maintains competitive results even with smaller or different models, confirming that the design generalizes well across models of similar capacities and does not rely on any single pretrained model.

Table 1: Ablation on different VLMs

Weakness 2.1:

Thank you for pointing this out. In the paper, our reference to the “original waypoint generator” corresponds to the learning-based waypoint predictor used in prior VLN works. This predictor is typically trained on the R2R dataset using RGB or depth inputs to regress a set of candidate waypoints (angles and distances). It operates as a perception network with no explicit structural grounding, relying entirely on visual cues and learned priors. We use the version from Open-Nav in all comparison experiments in the full paper.

In contrast, STRIDER replaces this component with a geometry-based, layout-consistent skeleton extraction process that requires no training and offers interpretable and spatially constrained waypoint candidates.

Weakness 2.2:

We appreciate the reviewer’s suggestion, and we would like to clarify that Table 3 in the full paper already includes the requested controlled comparison.

Specifically, the first row (marked with "×" under SWG) uses the original learning-based waypoint predictor from Open-Nav, while keeping other components unchanged. In contrast, the remaining rows apply our structured waypoint generator (SWG) under different configurations. As shown, enabling SWG instead of the original waypoint predictor yields consistent improvements across all metrics, confirming the effect of structured waypoint design.

Weakness 3.1:

Thank you for pointing this out. Our prompting setup for GPT-4o follows the prompt templates from Open-Nav. We intentionally adopt these prompts without major modifications to ensure a fair comparison with prior methods and to isolate the contribution of our structural planning architecture. Since our focus is not on prompt engineering for LLMs, we kept this component consistent across experiments.

Weakness 3.2:

STRIDER adopts the same prompt as Open-Nav, which includes historical observations, previous decisions, and prior reasoning steps, along with the current view. While SmartWay supports explicit backtracking when the agent detects failures, STRIDER follows a forward-only design.

Weakness 4.1:

We acknowledge the performance gap between STRIDER and fully supervised models. However, it is important to highlight that STRIDER operates entirely in a zero-shot manner without any task-specific training. This design is particularly valuable for practical deployments, where collecting annotated navigation data is often infeasible.

Despite this constraint, STRIDER achieves strong results across two challenging datasets (R2R-CE and RxR-CE). As shown in Table 1 and Table 2 in the full paper, STRIDER improves SR and SPL significantly (e.g., +20.6% SR and +34.9% SPL over baseline on R2R-CE; +11.5% SR and +52.3% SPL on RxR-CE), demonstrating its robust generalization to diverse environments and languages.

We believe this performance offers a complementary perspective to zero-shot VLN.

Weakness 4.2:

We appreciate the reviewer’s broader reflection on the maturity of zero-shot VLN-CE. While we agree that this research area is still developing, we believe this is what makes STRIDER a timely and meaningful contribution.

STRIDER offers a structural shift in zero-shot VLN by combining skeletonized spatial abstraction and action regulation. It complements existing approaches by offering an interpretable alternative that can generalize across environments without fine-tuning. While not aiming to replace fine-tuning or hybrid systems, STRIDER opens a new direction for zero-shot, interpretable navigation that we believe will grow increasingly relevant as foundation models continue to evolve.

Question 1:

Thank you for pointing out the typos. We appreciate your attention to detail and have carefully proofread the manuscript to fix such issues.

Question 2:

In STRIDER, the LLM’s role extends beyond ranking visual cues. GPT-4o is responsible for interpreting and decomposing the full instruction, reviewing the agent’s historical observations, past decisions, and prior reasoning. In this sense, the LLM performs structured, high-level reasoning grounded in both spatial layout and linguistic intent.

Weakness 1:

We understand the reviewer’s concern. While STRIDER’s pipeline involves multiple calls to a VLM and an LLM, our measurements show that the overall inference time per episode remains comparable to the Open-Nav baseline. Specifically, STRIDER requires 3.9074 minutes per episode, compared to 3.6606 minutes for Open-Nav. We believe this minor overhead is acceptable given the substantial gains in navigation performance, and it demonstrates that STRIDER is practical for offline applications.

Weakness 2:

Thank you for the helpful suggestion. To assess how much of STRIDER’s performance gain comes from model design rather than model choice, we conduct controlled experiments where STRIDER and prior methods use the same LLM and VLM. As shown in Table 1, STRIDER outperforms the other two baselines under identical LLM and VLM settings, confirming that the performance gain stems from our framework design.

Table 1: Comparison Under Identical VLM and LLM

Question 1:

Our method is designed with a focus on indoor navigation scenarios, where environments typically exhibit well-defined spatial constraints and structured topologies (e.g., hallways, rooms, doorways). In these settings, skeleton extraction yields meaningful connectivity graphs that align with human spatial reasoning and support robust, layout-consistent planning.

We agree that in open or non-standard layouts, such as large open spaces without clear boundaries, the extracted skeleton may become overly simplified or lack informative branching. However, for the indoor environments targeted in VLN-CE benchmarks like R2R-CE and RxR-CE, which are exactly our tasks, our abstraction is well-suited and highly effective.

Question 2:

Yes, the Task-Alignment Regulator is explicitly designed to handle fine-grained instruction alignment by generating feedback based on scene changes. At each step, the regulator compares pre- and post-movement observations, and reasons over whether the intended semantic progression (e.g., fully entering a room vs. partially entering) has been achieved.

This allows the agent to detect subtle differences and adjust accordingly in the next decision step. As shown in Fig. 5 in the full paper, the regulator distinguishes between visually similar but semantically distinct states.

Question 3:

We appreciate the reviewer’s attention to detail. The candidate waypoints in Figure 2 are intended for illustrative purposes only, to convey the general idea of skeleton-based waypoint sampling. As shown, the figure includes various keypoints along the skeleton, including some with degree > 1, to better visualize the underlying graph structure. However, as described in Section 3.2 (and shown in Figure 3), our actual method only retains candidate waypoints with degree = 1, i.e., skeleton endpoints.

Question 4:

The performance without using the Task-Alignment Regulator (TAR) is reported in Table 4 in the full paper, specifically in the first row labeled “×”. We intentionally separate this analysis from the Structured Waypoint Generator (SWG) ablation in Table 3 in the full paper to isolate the individual contribution of each component.

Weakness 1:

While NE (Navigation Error) is a meaningful metric, it is not the most decisive indicator of navigation quality, as shown in recent works such as Smartway, which accepts slightly higher NE (7.01), in exchange for improved SR, SPL, and NDTW. This is because NE measures the Euclidean distance from the agent's final position to the goal, but it does not account for whether the agent successfully followed instructions or generated a coherent, goal-directed path. In contrast, SR (Success Rate) reflects whether the agent stops next to the goal, SPL (Success weighted by Path Length) evaluates path efficiency while accounting for success, and NDTW (Normalized Dynamic Time Warping) captures trajectory fidelity concerning the reference path. These metrics are more indicative of instruction-following accuracy, semantic alignment, and task-level success, especially in instruction-following navigation. To this end, recent works mainly use SR, SPL, and NDTW to evaluate the overall performance of the methods.

Weakness 2:

While STRIDER’s OSR (39) on R2R-CE is lower than SmartWay's (51), we believe this difference is largely attributable to SmartWay’s use of a backtracking mechanism and significantly longer trajectories (TL = 13.09) as shown in Table 1. OSR considers a path successful if any visited point comes within a threshold distance of the goal. Therefore, methods with longer and more exploratory trajectories inherently have higher OSR, even if they do not stop correctly or follow instructions precisely. In contrast, STRIDER produces concise, forward-directed trajectories (TL = 8.47) without backtracking, yet still achieves a strong OSR. This suggests that our agent’s paths are well-aligned with the goal, not by chance, but through accurate perception and structured reasoning. STRIDER also ranks highest in SR and SPL, further indicating its strength in task completion and trajectory quality, rather than relying on incidental proximity.

We see this as a design trade-off: STRIDER avoids unnecessarily long or exploratory paths, and instead focuses on semantically faithful, efficient navigation.

Table 1: Comparison of TL and OSR

Question 1:

We would like to emphasize that our architecture adopts a modular design, where VLMs are used for perception and LLMs for reasoning. This design is not tied to any specific models and is model-agnostic across components of similar capabilities.

To verify this, we conduct an ablation study across multiple VLMs of varying sizes and providers (Table 2). The results demonstrate that while stronger models do yield better performance, STRIDER maintains competitive results even with smaller or different models, confirming that the design generalizes well across models of similar capacities and does not rely on any single pretrained model.

Table 2: Ablation on different VLMs

Question 2:

We measure the average inference time per episode for both STRIDER and the Open-Nav baseline. STRIDER takes 3.9074 minutes per episode compared to 3.6606 minutes for Open-Nav. Despite incorporating additional reasoning steps via the Task-Alignment Regulator, STRIDER introduces only a minor runtime overhead. We believe this slight increase is acceptable given the clear performance improvements, and it confirms that STRIDER remains computationally practical for real-world applications.

Question 3:

Thank you for the insightful suggestion. We evaluate the effect of feedback quality by conducting an ablation study using open-source VLMs of varying sizes and capabilities, including Qwen2.5-VL-72B, 32B, and 7B (Table 2). The results confirm that higher-quality feedback (e.g., from larger or more capable models) improves performance. However, STRIDER continues to offer improvements over baselines even when paired with smaller models, suggesting that our structural guidance provides added value regardless of model scale.

Question 4:

STRIDER does not rely on the model generating an explicit “stop” signal.

Specifically, the instruction is first broken down into a series of subtasks by the LLM, and the maximum number of execution steps is set to be equal to the number of subtasks, with a hard minimum of 6 steps to ensure completion. Once the max step is reached, the system stops.

This approach avoids reliance on potentially unreliable stop predictions from the model and ensures stable behavior across tasks.