acceptodds
Under review as a conference paper at ICLR 2027

SuperNav: An Agentic Navigation System for Any Task in Any Scene

20 views

Abstract

General-purpose service robots need navigation systems that can handle diverse human requests in unfamiliar environments, combining task generality with scene generality. Some existing methods fine-tune multimodal large language models (MLLMs) to predict navigation actions, making their behavior dependent on the coverage of navigation training data and potentially limiting generalization to new requests and environments. Our key insight is to let the MLLM focus on what it does well—interpreting requests, understanding scenes, and making decisions—while preserving its general-purpose capabilities and delegating motion execution to navigation tools. To realize this idea, we introduce SuperNav, which equips a pretrained MLLM with a specialized agent harness without navigation-specific fine-tuning of the MLLM. Our harness supports these decisions with Navigation Skills, agent-oriented Tools for physical interaction, and task-progress and context management. A unified visual-point interface connects decision-making to motion by allowing the model to specify destinations directly in images and revise its decisions from execution feedback. Together, these components support sustained navigation across different task requirements and environments. SuperNav outperforms four evaluated baselines on instance-level, multi-object, and demand-driven tasks. Category-level evaluation on HM3D and deployment on a real quadruped robot further demonstrate its applicability across environments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied navigation, multimodal large language models, agentic systems, tool use, navigation skills

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_z4CdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents SuperNav, a system that lets a general AI model (GPT) control a navigating robot without training the model for navigation. The model makes the high-level decisions which is, it reads the request, looks at camera images, decides where to go, tracks progress and decides when the task is done. Movement is handled by separate tools. The model picks a point in one of four camera images, and a motion system drives the robot there. System has 4 main parts, Agent Loop, Navigation Skills, Tools and Two motion systems.

Critical strengths

Author has done a good job in separating “deciding” from “moving” through image points, it's simple and practical. The same decision system works with different motion systems. Improvement over baselines are very significant.

Critical weaknesses

The harness’s own contribution is not proven. SuperNav differs from the baselines in many ways at once: a stronger AI model, four cameras instead of one, a perfect map, and a longer time budget. It is unclear how much of the gain comes from the proposed harness itself. There are some evaluation concern as well. The agent had full file access and was only instructed not to read the answer files. SuperNav gets 1 hour per episode, while other methods get 500 steps, and SuperNav’s step count is not reported. The HM3Dv2 setup for SuperNav (number of episodes, stopping rule) is not described, yet its result is marked as the best. The ablation tests use 120 episodes, so most of the differences are within noise (about ±8 points). The title says “any task in any scene,” but the tests are indoor navigation only. The real-robot results are demos with no success rates.

Questions for authors
  1. Can you compare SuperNav with a simple fixed-rule system that uses the same GPT model, the same four cameras, the same motion planner and the same budget? This would show whether the harness itself causes the gains.

  2. How many steps does SuperNav use per episode, and what was the HM3Dv2 setup?

3 comments
AuthorsOpenReview ↗

Response: Controlled Baseline Comparison and HM3Dv2 Evaluation Details

Thank you for these suggestions!

For question1: We added an InstructNav-adapted fixed-sequence baseline using the same GPT-5.6 Terra, four cameras, same motion planner, and time budget on the same 120 OVON episodes. At 0.25 m/1 m, SuperNav achieves 68.33%/73.33% SR and 0.3837/0.4105 SPL, versus 41.67%/54.17% SR and 0.1865/0.2306 SPL for the InstructNav-adapted method, supporting the harness’s contribution beyond these shared components. Details are provided in Table 2 (Section 4.3) and Appendix B.7 of the revised paper.

For question2: SuperNav averages 239.62 steps per HM3D v0.2 episode. We evaluate ObjectNav on the HM3D v0.2 validation split (36 scenes, 1,000 tasks). Success requires the final position’s navmesh geodesic distance to the nearest goal viewpoint to be strictly <0.2 m; we additionally report results at 1 m. The evaluation protocol is detailed in Section 4.3 and Appendix B.5 of the revised paper. We would be happy to provide a more detailed statistics table if needed.

Thanks to the authors for addressing the concerns and new experiments.

The new comparison with the InstructNav baseline is helpful and addresses my main concern.

The additional details on HM3D evaluation and step counts also clear up my second question.

I still have some concerns about the small evaluation set, lack of quantitative results on real robots and the broader claims made in the paper. But overall, the response addresses my main concerns and I will keep my original rating of 3 (Weak Acceptance) with same confidence.

Thank you for taking the time to read our response and the new results. We’re glad the InstructNav comparison addressed your main concern and that the HM3D evaluation details clarified your second question.

Reviewer_4QPdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose SuperNav, a system that uses a pretrained MLLM to interpret requests and select navigation goals without navigation-specific fine-tuning. In simulation, SuperNav achieves higher success rates (SR) than the evaluated baselines on single-object, multi-object, and demand-driven navigation tasks. The authors also demonstrate deployment on a real quadruped robot. However, comparisons with closely related methods and analyses of path efficiency and execution time remain insufficient.

Critical strengths
  1. The proposed method supports multiple tasks and environments without additional training of the high-level MLLM and demonstrates high navigation success rates in simulation.
  2. In addition to ablations of Navigation Skills and tool design, the authors evaluate different MLLM configurations and motion execution backends, examining how these configurations affect performance.
Critical weaknesses
  1. Given the agentic nature of the proposed method, comparisons with existing agentic approaches, such as AgenticNav and HarnessVLN discussed in Related Work, are important. The authors should clarify whether SuperNav retains its advantage when these methods use the same backbone model, GPT-5.6 Terra, with matched reasoning settings and execution conditions.
  2. Since GT Planner uses privileged information, such as the ground-truth navigation mesh, the following discussion focuses on the Learned Executor results. In single-object navigation, SR improves substantially, whereas SPL is slightly lower than that of the best baseline. Path efficiency also remains a concern in multi-object navigation, and SPL on HM3D falls below the reported results of many existing methods. These results suggest inefficient navigation paths. The authors should analyze revisits, motion failures, and recovery behaviors to explain the factors.
  3. The concerns about path inefficiency noted above, together with repeated MLLM calls, raise concerns about execution time. The authors should report average execution time relative to the baselines, along with a breakdown of the time spent on MLLM inference and movement.
  4. The real-world evaluation is limited to qualitative demonstrations. The authors should report success rates with clearly specified trial counts and task conditions, and compare against existing methods under the same conditions.
Questions for authors
  1. How effective is the proposed method compared with agentic approaches discussed in Related Work, such as AgenticNav and HarnessVLN, when the backbone model and other conditions are matched?
  2. Why are improvements in SPL absent or limited despite the gains in SR?
  3. How does execution time compare with existing methods, and what is its detailed breakdown?
  4. What is the success rate in real-world robot experiments, and to what extent does the proposed method outperform existing approaches?
5 comments
AuthorsOpenReview ↗

Response to Question 1: Comparison with Agentic Navigation Systems

Thanks for the review and the question.

We added a HarnessVLN comparison using the same GPT-5.6 Terra backbone on the same 120 OVON val-unseen episodes under the same evaluation settings, including the same budget. At 1 m, SuperNav + GT Planner achieves 73.33% SR and 0.4105 SPL, compared with 53.33% and 0.2941 for HarnessVLN (Table 2, Section 4.3; Appendix B.6).

Conceptually, HarnessVLN is harness-centric—the MLLM proposes actions and stopping, while the harness programmatically validates them. SuperNav is MLLM-centric—the MLLM selects task-level actions and judges semantic completion, while Skills provide soft procedural guidance and motion tools perform geometric execution checks. AgenticNav’s implementation is not publicly available, preventing a faithful controlled reproduction. Show-Harness targets robotic manipulation rather than navigation, so we discuss it conceptually in Related Work instead of adding a navigation comparison.

AuthorsOpenReview ↗

Response to Question 2: SR and SPL

SPL jointly measures success and path efficiency, SPLi=SiLi/max⁡(Pi,Li)\mathrm{SPL}_i=S_iL_i/\max(P_i,L_i), whereas SR measures success alone. One explanation is that exploration, candidate rechecking, and recovery allow SuperNav to succeed on otherwise failed, harder episodes, but these successes can require longer executed paths and thus can lower SPL.

AuthorsOpenReview ↗

Response to Question 3: Runtime Comparison and Breakdown

We have added execution-time statistics in Appendix B.8 (Tables 9–10). SuperNav is generally slower than the four learned-policy baselines because of repeated MLLM reasoning and tool interaction. For example, SuperNav + GT Planner averages 384.00 s on single-object navigation, compared with 30.00–266.40 s for the baselines. And on HM3Dv2, its mean episode times are 445.25s; cumulative model time averages 383.05 s (approximately 86% of the mean). More details are provided in the tables. This indicates that faster upstream responses would substantially reduce SuperNav’s execution time.

Reviewer_6pRQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an agentic approach to embodied navigation tasks which can address diverse tasks. Navigation Skills, agent-oriented Tools, context/progress management, and a unified visual-point interface to connect high-level reasoning with geometric or learned motion execution. The method is evaluated across instance-level, multi-object, demand-driven, and category-level navigation tasks, as well as real-world robot demonstrations.

Critical strengths
  1. The results are good. The compared performance is strong on various benchmarks. Its exhibition is nice from simulations to real robots.

  2. The paper posepose a unified agentic framework that handles diverse navigation tasks without navigation-specific MLLM fine-tuning and achieves strong performance on the evaluated tasks. The visual-point tool interface provides a practical separation between high-level MLLM reasoning and motion execution,

Critical weaknesses
  1. Comparison fairness The main comparisons are not fully controlled, as the evaluated systems differ substantially in observation interfaces, model capabilities, action spaces, history mechanisms, and motion backends. This makes it difficult to isolate the contribution of the proposed agent harness.

  2. Generality evidence The empirical evidence for the claimed “any task in any scene” generality remains limited. While the paper covers several navigation settings, the instance-level tasks are evaluated on a newly constructed benchmark, demand-driven navigation uses an adapted protocol, and the real-world evaluation is limited to qualitative demonstrations on a single robot platform.

Questions for authors
  1. comparisons to agentic baseline approaches are missing. would it be possible to add such approaches based on reasoning?
  2. would it be possible to compare against baseline models using previously proposed benchmarks on the same protocols for at least 1-2 cases?
2 comments
AuthorsOpenReview ↗

Response to Questions 1–2: Agentic Baselines and Benchmark Comparisons

Thanks for your review and the questions!

1. Yes. We added two reasoning-based comparisons on the shared 120-episode OVON subset. HarnessVLN uses the same GPT-5.6 Terra backbone and evaluation settings, including the same budget; SuperNav + GT Planner achieves 73.33% SR and 0.4105 SPL, compared with 53.33% and 0.2941 for HarnessVLN. We also added InstructNav-adapted with the same GPT model, four cameras, GT Planner, and budget; SuperNav improves SR from 41.67% to 68.33% at 0.25 m and from 54.17% to 73.33% at 1 m.

AgenticNav’s implementation is not publicly available, preventing a faithful controlled reproduction. Show-Harness targets robotic manipulation rather than navigation, so we discuss it conceptually in Related Work instead of adding a navigation comparison. Details are provided in Table 2 (Section 4.3) and Appendices B.6–B.7.

2. We now evaluate on two established ObjectNav benchmarks: HM3D-OVON val-unseen and HM3Dv2 validation. The OVON comparisons above are local reruns on the same 120 episodes under matched evaluation settings. We additionally evaluate 1,000 HM3Dv2 episodes across 36 scenes, reporting SR/SPL at 0.2 m and 1 m; SuperNav + GT Planner achieves 80.30%/0.3338 and 86.50%/0.3638, respectively.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

Thank you for suggesting comparisons with agentic baselines on established benchmarks. We added two baseline comparisons on the same OVON episodes. Could you let us know whether these additions address your two questions, or whether anything needs clarification?

Best regards, The Authors

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

This paper presents SuperNav, an agentic navigation framework designed to support diverse navigation tasks across unfamiliar environments without navigation-specific fine-tuning of the underlying multimodal large language model (MLLM). The proposed system integrates a pretrained MLLM with a specialized agent harness consisting of Navigation Skills, agent-oriented Tools, task-progress tracking, and context management. A unified visual-point interface enables the MLLM to select navigation destinations directly from visual observations while delegating motion execution to either a geometric planner or a learned navigation policy. The framework is evaluated across instance-level, ordered multi-object, demand-driven, and category-level navigation tasks, together with real-world deployment on a Unitree Go2 robot. Experimental results demonstrate substantial improvements over several navigation baselines, suggesting that general-purpose MLLMs combined with structured agent execution mechanisms can provide a promising alternative to task-specific navigation models.

Critical strengths
  1. Relevant and timely research direction. The paper explores an important shift from navigation-specific model fine-tuning toward general-purpose MLLM-based agents equipped with reusable tools and execution mechanisms. This direction is particularly relevant to generalist embodied intelligence.
  2. Well-structured system architecture. The proposed separation between high-level reasoning and low-level motion execution is sensible. The integration of Navigation Skills, agent-oriented Tools, and context management provides a coherent framework for long-horizon navigation.
  3. Unified visual-point interface. The use of image-space destination selection with interchangeable geometric and learned motion backends is practically appealing and facilitates integration with different robotic platforms.
  4. Strong empirical performance. Table 1 reports substantial improvements over the evaluated baselines. SuperNav with GT Planner achieves 78.00% success on single-object navigation and 59.50% on demand-driven navigation, compared with 34.00% and 37.50% for the strongest respective baselines.
Critical weaknesses
  1. Although the proposed agent harness is well organized, its main components—including tool calling, procedural Skills, context management, visual grounding, and hierarchical navigation—are individually established techniques. More importantly, closely related approaches such as AgenticNav, HarnessVLN, Show-Harness, and DemandAgent are explicitly discussed in the paper. The authors should more precisely distinguish SuperNav from these methods and explain which components represent substantive methodological innovations rather than system-level integration.
  2. The title, An Agentic Navigation System for Any Task in Any Scene, suggests extremely broad task and environment generalization. However, the evaluation covers a limited number of navigation task families and primarily indoor environments. The authors should either moderate the title and corresponding claims or provide substantially broader evaluations involving different environmental conditions, task compositions, and navigation requirements.
  3. The proposed framework relies on a powerful pretrained MLLM and, in its strongest configuration, a geometric planner with access to depth, pose, and navigation geometry. In contrast, several evaluated baselines use different observation modalities and action interfaces. Although the authors preserve the baselines' native configurations, these differences complicate attributing performance improvements specifically to the proposed harness. More controlled comparisons using comparable perception inputs, motion execution capabilities, and computational budgets would strengthen the conclusions.
Questions for authors
  1. What are the fundamental differences between SuperNav and closely related agentic navigation systems such as AgenticNav, HarnessVLN, and DemandAgent? Can the authors provide direct experimental comparisons under identical evaluation conditions?
  2. How much of the performance improvement comes from the agent harness itself, rather than the stronger pretrained MLLM, multi-view observations, or privileged geometric navigation backend?
  3. Can the authors provide detailed statistics on inference latency, token consumption, tool-call frequency, total task execution time, and computational cost for different navigation tasks?
  4. How robust is the proposed framework to localization drift, inaccurate depth measurements, dynamic obstacles, and perception failures? Can the authors quantify these effects in real-world experiments?
  5. How sensitive is the system to the design and wording of Navigation Skills? Would simpler navigation prompts or automatically generated Skills achieve similar performance?
  6. Why does the Learned Executor exhibit substantially lower SPL than GT Planner despite achieving relatively competitive success rates? Could the authors provide a more detailed analysis of its trajectory quality and stopping behavior?
7 comments
AuthorsOpenReview ↗

Response to Question 1: Differences and Comparisons with Agentic Systems

We thank the reviewer for highlighting the importance of both conceptual differentiation and controlled empirical comparison with closely related agentic navigation systems.

The main differences lie in task scope and control allocation.

  • AgenticNav focuses on VLN-CE tool calling—pixel-goal motion, depth queries, and historical-image retrieval—whereas SuperNav addresses navigation task and scene generality through Navigation Skills, progress/context management, and interchangeable visual-point execution backends. AgenticNav’s implementation is not publicly available, preventing a faithful controlled reproduction.

  • HarnessVLN programmatically validates MLLM action and termination proposals, whereas SuperNav leaves task-level action selection and semantic completion to the MLLM, with Skills serving as soft procedural guidance. We added a matched comparison using the same GPT-5.6 Terra backbone, 120 OVON episodes, and evaluation settings: at 1 m, SuperNav + GT Planner achieves 73.33% SR and 0.4105 SPL, versus 53.33% and 0.2941 for HarnessVLN (Table 2; Appendix B.6).

  • DemandAgent translates high-level human demands into dependency-aware executable behaviors through its Adaptive Demand Ledger and persistent spatial memory. In contrast, SuperNav covers instance-level, category-level, multi-object, and demand-driven navigation, together with quadruped deployment. DemandAgent’s original protocol supports alternative object assignments and completion orders, whereas our adapted Demand-Bench protocol scores ordered trajectory arrivals. Our use of Demand-Bench is intended to evaluate navigation capability rather than dependency-aware completion of the underlying activities. Therefore, their reported scores are not directly comparable (Appendix B.2).

AuthorsOpenReview ↗

Response to Question 2: Harness, Model, and Backend Contributions

The revised Section 4 provides controlled evidence for the harness’s contribution.

With GPT-5.6 Terra and GT Planner fixed:

  • removing Navigation Skills reduces SR from 68.33% to 35.00% at 0.25 m and from 73.33% to 54.17% at 1 m.
  • Restricting observations and pointing to the front-only interface reduces SR to 52.50%/65.83% and increases mean explicit turns from 0.41 to 17.56, demonstrating the contribution of the combined multi-view Tool design.
  • Removing both Skills and agent-oriented Tool design yields 31.67%/54.17% SR.

Meanwhile, Learned Executor retains 65.00%/70.83% SR without depth-based grounding or navmesh planning, although its SPL decreases, indicating that privileged geometry improves path efficiency but is not the sole source of success.

Model configuration also matters: GPT-6 Astra achieves 71.67%/79.17% SR with the full harness. These results support complementary contributions from Skills, Tools, the model, and the execution backend rather than attributing the gains to any single factor (Sections 4.3–4.4, Tables 2–3; Appendix B.6).

The controlled InstructNav-adapted comparison provides additional evidence: with the same GPT-5.6 Terra, four cameras, GT Planner, and 3,600-s budget, SuperNav achieves 68.33%/73.33% SR at 0.25 m/1 m, compared with 41.67%/54.17% for InstructNav-adapted. Together, these results support complementary contributions from Skills, Tools, the model, and the execution backend rather than attributing the gains to any single factor (Sections 4.3–4.4, Tables 2–3; Appendices B.6–B.7).

AuthorsOpenReview ↗

Response to Question 3: Runtime and Resource-Usage Statistics

Sure,we added runtime and resource statistics in Appendix B.8 (Tables 9–11), covering total-time distributions, cumulative model time, tool-call counts/rates, and input/cached-input/output/reasoning tokens. For GT Planner on OVON, mean episode time is 174.78 s, including 103.95 s model time and 11.17 tool calls; input/output usage is 588.99k/4.99k tokens. Results for HM3Dv2, Learned Executor, and single-object, multi-object, and demand-driven tasks are also reported. These measurements quantify resource cost; monetary cost depends on the model provider and caching conditions.

Reviewer_hsnbOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. This paper presents SuperNav, which pairs a frozen pretrained MLLM with a navigation-specific harness and uses an agent loop, Tools, and Navigation Skills to handle diverse navigation tasks, without navigation-specific fine-tuning of the MLLM.
  2. The core interaction is a unified visual-point interface executed by either a GT Planner or a Learned Executor, with results reported on instance, multi-object, demand-driven, and HM3D settings, as well as real-robot deployment.
Critical strengths
  1. The problem setting covers both task generality and scene generality. Evaluation goes beyond a single ObjectNav benchmark to include ordered multi-object and demand-driven navigation, which is more complete than reporting only one task.
  2. Decision-making is separated from motion execution, and the visual-point interface can plug into different motion backends. Qualitative validation on a Unitree Go2 further supports the completeness.
  3. Ablations on Skills, Tool design, and decision-model substitution show that both harness components and MLLM choice affect performance.
Critical weaknesses
  1. The method is essentially a frozen MLLM plus an agent loop, i.e., harness engineering, without a new navigation model or learnable mechanism. Skills are on-demand textual guidance, while the only trained component, the Learned Executor, is deferred to the appendix and closely follows NoMaD. Given concurrent agentic navigation harnesses, the incremental contribution relative to this line of work needs to be articulated more clearly.
  2. The large gains in the main tables conflate a frontier MLLM, four-view interaction, and privileged geometry in the GT Planner, while baselines are smaller RGB action models. Switching the motion backend sharply reduces SPL, and changing the decision model also moves the numbers substantially, so the gains are hard to attribute to the harness design itself.
  3. Several evaluation protocols are misaligned with comparable work. For example, Demand-Bench is altered to ordered 2 m arrival with an explicit note of incomparability, and HM3D largely cites published numbers under a different evaluation subset and success criterion. Combined with heavy reliance on a closed-source API, both verifiability and robustness of the results are weakened.
Questions for authors
  1. It would help to report the gains from toggling only Skills, progress management, and four-view Tool design under a fixed motion backend and observation protocol, relative to the same frontier MLLM. This is important for separating how much of the main-table advantage comes from the harness versus model capability.
  2. The GT Planner uses privileged geometry, whereas baselines such as NaVid are primarily RGB action models. Clarifying whether the two sides are compared under similar geometric information would better explain the current large SR gap.
  3. The concrete content of the Navigation Skills, and whether they were iterated on for the evaluation tasks, is still underspecified. Clarifying whether the Skills used for evaluation were frozen, and how they relate to failure-case analysis, would make it easier to assess their generality.
  4. close(achieved) is the model’s own completion judgment, while evaluation SR is scored separately by distance-based rules. Reporting the rate of disagreement between the two, especially “declared complete but unsuccessful” versus “successful but not properly terminated,” would more directly reflect the reliability of semantic completion judgments.
  5. The demand-driven experiments use an adapted ordered 2 m arrival protocol and explicitly note incomparability with DemandAgent. Under this setting, clarifying which claim these experiments are primarily intended to support would help interpret the role of these results.
6 comments
AuthorsOpenReview ↗

Response to Question 1: Skills, Tools, and Progress-Management Ablations

The revised Section 4.4 isolates these contributions using the same GPT-5.6 Terra (high) and GT Planner.

  • Removing Navigation Skills while retaining the Tool interfaces reduces SR from 68.33% to 35.00% at 0.25 m and from 73.33% to 54.17% at 1 m.
  • Replacing four-view, cross-direction Tool interaction with a front-only interface reduces SR to 52.50%/65.83% and increases mean explicit turns from 0.41 to 17.56.
  • Removing only the progress management Tool on the same 150 multi-object episodes reduces SR from 34.00% to 31.33% and SPL from 0.1388 to 0.1198.

As further evidence, InstructNav-adapted uses the same GPT model, four cameras, GT Planner, and 3,600-s budget, but a fixed decision sequence. SuperNav achieves 68.33%/73.33% SR, compared with 41.67%/54.17% for InstructNav-adapted at 0.25 m/1 m. Together, these controlled comparisons support a substantial harness contribution beyond the shared MLLM and motion backend (Sections 4.3–4.4, Tables 2, 3, and 5; Appendices B.6–B.7).

AuthorsOpenReview ↗

Response to Question 2: Geometric Information and Backend Comparisons

The comparisons are not geometry-matched. GT Planner uses depth, camera poses, and navmesh geometry internally for motion execution, whereas NaVid and the other learned baselines retain their native observation and action interfaces. The upper-level MLLM does not receive evaluator targets or direct navmesh access. We therefore present the GT Planner results as system-level comparisons rather than geometry-controlled ablations.

And to examine this factor, Table 1 also evaluates Learned Executor, which performs image-goal navigation without depth-based grounding or navmesh planning. It achieves 68.00%/34.00%/45.00% SR on single-object, multi-object, and demand-driven navigation, respectively, exceeding the strongest baseline SR of 34.00%/4.00%/37.50%.These results indicate that the SR gains are not solely attributable to privileged geometry.

AuthorsOpenReview ↗

Response to Question 3: Skill Content, Freezing, and Generality

Section 3.3 summarizes the Skills’ search, recovery, and completion guidance. To illustrate the concrete Skill content, we have updated the Supplementary Material to include, for example, skills.zip, which contains the Learned Executor entry SKILL.md files and linked references.

The original Skills were frozen before evaluation and not tuned on evaluation episodes; failure feedback informs decisions without rewriting the Skills.

In a separate check to examine generality beyond manually refined guidance, we used GPT-6 Astra to generate Skills and a prompt in a fresh context without evaluation trajectories. Under identical settings, the generated guidance achieves 71.33% SR and 0.4542 SPL, versus 78.00% and 0.4127 for the original guidance (Section 4.4, Table 4).

Comments on the paper

3 comments
AuthorsOpenReview ↗

Title Revision

We thank the reviewers for noting that “any task in any scene” overstates the scope supported by our evaluation. Our experiments span diverse navigation tasks and environments, but they demonstrate broad applicability within the evaluated settings rather than universal generality.

We have revised the title to “SuperNav: An Agentic System for General Embodied Navigation.”

AuthorsOpenReview ↗

Response to the comparison result with HarnessVLN

Thanks for the response,SuperNav retains higher success with the Learned Executor. On the same OVON subset, SuperNav with the Learned Executor achieves 70.83% SR, compared with 53.33% for HarnessVLN. Though its SPL is lower (0.1443 vs. 0.2941),HarnessVLN also uses simulator-provided information for navigation. Its execution pipeline consumes simulator-provided depth and pose to construct maps, localize targets, and execute motion. These inputs provide privileged sensing in simulation.

AuthorsOpenReview ↗

Map Sources, Motion Executors, and Their Impact on SPL

Thank you for this helpful follow-up. We agree that the SPL gap at comparable SR requires direct analysis of path efficiency, and have added comparisons on jointly successful episodes.

1. Map sources and path planning.

Appendix B.9, Table 12 now details each method's map sources and execution mechanisms; Table 2 also identifies its motion executor. The table distinguishes complete GT NavMesh planning, online mapping with FMM/A*, learned local control, and waypoint teleportation. Several online mapping implementations use simulator-provided depth and pose, while their map coverage grows with observations. For example, MSGNav selects destinations using an online TSDF/scene graph, queries the GT NavMesh for shortest paths, and executes waypoint teleportation through direct updates to the agent's pose.

These geometric resources and execution mechanisms affect path efficiency across the compared methods.

2. How mapping and execution affect SPL.

Complete GT NavMesh access supports short routes between selected destinations and reduces local route-search detours. Online FMM/A* can also provide efficient routes, with performance depending on observed map coverage, obstacle representation, and replanning. Learned controllers depend on their observations and waypoint-following/stopping accuracy. Direct waypoint teleportation changes how planned routes are executed and recorded. Across these mechanisms, local stopping and the agent's sequence of destinations jointly determine total travel.

Following your suggestion, we analyzed 693 HM3Dv2 episodes successful under both GT Planner and Learned Executor. Learned Executor has longer mean trajectories (50.67 vs. 22.25 m), greater segment detouring (1.44 vs. 1.01), and more episodes with success-region re-entry (39.7% vs. 18.8%). These measurements identify additional local detours and repeated approaches alongside exploration and verification. Appendix B.9, Table 15 provides the full statistics and measurement definitions.

3. Backend effects and SuperNav's agentic contribution.

GT Planner benefits from GT NavMesh geometry and achieves higher SPL than our Learned Executor. We also replaced it with ApexNav's online mapping and planning, keeping the high-level agent and evaluation settings fixed on the same 120 OVON episodes (Appendix B.9, Table 13):

SuperNav motion backendSR at 0.25 m (%)SPL at 0.25 mSR at 1 m (%)SPL at 1 m
GT Planner68.330.383773.330.4105
ApexNav Executor65.000.392770.830.4281
Learned Executor65.000.135570.830.1443

ApexNav increases SPL, with SR lower by 3.33/2.50 percentage points. Its efficiency advantage persists on episodes successful under all three executors at each threshold (Table 14):

Success thresholdApexNav SPLGT Planner SPLLearned Executor SPL
0.25 m0.65540.60410.2317
1 m0.65940.58840.2208

The controlled comparison shows that online mapping and planning can yield higher SPL than GT NavMesh execution.We notice that in some tasks, with the ApexNav Executor, agent selects more direct waypoints and makes fewer repeated approaches to the target, resulting in shorter overall trajectories and a higher average SPL.

In an additional episode-level check, we fixed Agent's original decision points and replaced only its executor with GT execution. Path lengths were close (6.015 m for ApexNav Executor vs. 5.758 m for GT planner). Both succeeded, stopping 0.137 vs. 0.054 m from the goal. It shows similar path efficiency with fixed decisions between the two executors;

SuperNav contributes an agentic navigation framework for diverse tasks and scenes.Its unified visual-point interface supports motion backends suited to available maps and sensors, including geometric planning, online depth/pose mapping, and RGB-based learned control. With the same GT Planner, GPT model, four cameras, and time budget, SuperNav also achieves 68.33%/73.33% SR versus 41.67%/54.17% for InstructNav-adapted, supporting the contribution of its task-level decision workflow

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.