acceptodds
Under review as a conference paper at ICLR 2027

A Recursive Self-Improving Harness for Training-Free Generalist Navigation

Abstract

Generalist embodied navigation requires agents to adapt their behavior as task objectives, spatial contexts, and environmental conditions evolve. While recent agentic navigation methods leverage pretrained vision-language models (VLMs) for flexible reasoning, they remain largely task-specific and lack mechanisms for reusing experience across tasks and environments. We introduce RING-Nav, a recursive self-improving (RSI) harness for training-free generalist navigation. RING-Nav maintains a contextualized navigation state (CNS) that unifies task progress, spatial evidence, episodic memory, and tool feedback, enabling a frozen VLM to dynamically orchestrate shared navigation tools. To support continuous improvement across episodes, an RSI experience graph (RSI-EG) records contextualized tool transitions and their outcomes, retrieves relevant past experience, and guides subsequent tool decisions without updating model parameters. RING-Nav thus unifies spatially grounded tool orchestration (SGTO) with experience-driven self-improvement, enabling a single navigation agent to adapt across diverse tasks and environments. Extensive experiments demonstrate state-of-the-art performance on both individual navigation tasks and heterogeneous multi-stage instructions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual Navigation, Physical Agent, Harness

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

This paper introduces RING-Nav, a training-free method capable of solving diverse navigation tasks in a unified framework. RING-Nav integrates task progress, episodic memory, and spatial evidence into CNS, which is then applied to tool decision-making and execution in SGTO. RSI-EG is proposed to store cross-episode tool executions and can be used in subsequent episodes through relevant experience retrieval.

Critical strengths
  1. Improving a unified navigation system without repeatedly fine-tuning models on diverse tasks is a practical direction.
  2. The empirical results are promising. The gains in R2R-CE and OmniNavBench suggest that the framework may offer practical benefits beyond simply deploying a fixed underlying vision-language model.
Critical weaknesses
  1. The central claim of RSI is insufficiently substantiated.

The experience graph evolves primarily through the accumulation of cross-episode records, while the procedures of retrieval, utility estimation, and decision fusion remain fixed. Improving task performance with a growing memory does not mean that the system is recursively improving itself. The paper does not demonstrate that the agent or the RSI-EG is becoming better (e.g., better estimation, better retrieval, or generating useful improvements) over iterations.

Moreover, as the central aspect of the paper statement, the authors should provide related work on RSI.

  1. The difference from conventional experience-based memory is unclear.

The current mechanism appears closely related to retrieval-augmented decision-making. For example, Expel [1] leverages experience accumulated in previous tasks to improve performance without fine-tuning. The authors should clarify the difference between RSI and experience retrieval methods.

  1. Concern about unfair evaluation.

RING-Nav maintains cross-episode tool execution history to inform subsequent navigation decisions. However, the applied benchmarks typically evaluate episodes independently, without allowing information leakage from other test episodes. Since test episodes may be conditioned in identical environments, inter-episode information sharing during evaluation risks introducing information leakage—thereby compromising evaluation fairness. For instance, in object navigation tasks, if the target object (e.g., a chair) has already been observed in a prior episode, the agent may bypass essential exploration behaviors, artificially inflating performance metrics. To ensure fair comparison, the authors should evaluate on lifelong (e.g., goat [2]) or continual learning (e.g., C-Nav [3]) benchmarks, which inherently allow the use of information from previous episodes.

  1. Missing many details.

The definitions of active subtask, current progress, and episodic memory in lines 208–209 are not specified. The weight values in the formulas of section 3.3 are not provided. The inputs and outputs of each tool are not explained. The harness prompt is not provided. The absence of these details critically undermines the method’s reproducibility.

[1] Expel: Llm agents are experiential learners (AAAI2024)

[2] Goat-bench: A benchmark for multi-modal lifelong navigation (CVPR2024)

[3] C-nav: Towards self-evolving continual object navigation in open world (NeurIPS2026)

Questions for authors
  1. What is the precise definition of the recursive self-improving in RING-Nav?
  2. Does the system become more effective at improving itself over successive iterations?
  3. Does the order of episodes have a significant impact on the evaluation results?
6 comments
AuthorsOpenReview ↗

Response to Reviewer GohK (1/5)

We are encouraged by the reviewer’s recognition of the practical value of training-free unified navigation and positive assessment of our promising results on R2R-CE and OmniNavBench. We address all the concerns point by point through clarifications of RSI and evaluation protocols, additional cumulative-improvement, GOAT-Bench, and episode-order experiments, and expanded RSI related work and implementation details.

Q1

Q1 What is the precise definition of the recursive self-improving in RING-Nav?

We thank the reviewer for requesting a clearer RSI definition and supporting evidence. For Q1, we define harness-level RSI as cumulative improvement arising from feedback between an agent’s evolving experience and decision-making. This interpretation is closely related to SpecForge [1], accepted to NeurIPS 2026, which refines external behavioral specifications through execution feedback while keeping model weights and the search procedure fixed.

In RING-Nav, the current experience graph guides tool decisions; the resulting episodes and execution outcomes update the graph, which then guides decisions and experience acquisition in the next episode. Thus, each updated harness helps generate the experience used for its subsequent update.

We have added this definition to the Introduction of the revised manuscript.

[1] SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specifications (NeurIPS 2026).

W1 and Q2

W1 The central claim of RSI is insufficiently substantiated. The experience graph evolves primarily through the accumulation of cross-episode records, while the procedures of retrieval, utility estimation, and decision fusion remain fixed. Improving task performance with a growing memory does not mean that the system is recursively improving itself. The paper does not demonstrate that the agent or the RSI-EG is becoming better (e.g., better estimation, better retrieval, or generating useful improvements) over iterations.

Moreover, as the central aspect of the paper statement, the authors should provide related work on RSI.

Q2 Does the system become more effective at improving itself over successive iterations?

We thank the reviewer for these constructive comments. We respond in four parts:

  1. Adaptive decision-making: Regarding “the procedures of retrieval, utility estimation, and decision fusion remain fixed,” we clarify how historical utility and fusion weights adapt to the evolving RSI-EG and current context.
  2. Recursive feedback: Regarding “Improving task performance with a growing memory does not mean...recursively improving itself,” we explain how the updated harness generates experience for subsequent updates, as defined in Q1.
  3. Evidence of improvement: Regarding “does not demonstrate that the agent or the RSI-EG is becoming better...over iterations” and Q2, we report additional SR/SPL experiments across successive experience-collection rounds.
  4. RSI related work: Regarding “the authors should provide related work on RSI,” we summarize the added literature and its relationship to our formulation.

1. The tool-selection decision in manuscript Eq. (7) is

=\underset{(u_j,\xi)\in\mathcal{A}_t}{\arg\max} \left[ \left(1-\lambda_{ij}^{(n)}(x_t^{j,\xi})\right)L_t(j,\xi) +\lambda_{ij}^{(n)}(x_t^{j,\xi})Q_{ij}^{(n)}(x_t^{j,\xi}) \right].$$ First, the agent directly uses retrieved experience before assigning candidate scores $L_t(j,\xi)$. As the graph evolves, changing evidence can alter its reasoning and assessment of candidate tools. Second, both $Q_{ij}^{(n)}$ and $\lambda_{ij}^{(n)}$ are adaptive: Eqs. (5)–(6) compute historical utility from relevant execution outcomes and the fusion weight from their effective support in the current navigation context. Graph updates can therefore change both a candidate’s estimated utility and the influence of historical evidence on its selection. **2.** As clarified in Q1, recursion is the feedback between the evolving experience graph and the policy generating its updates: the graph guides tool decisions, the resulting executions and outcomes update it, and the updated graph guides decisions and experience acquisition in the next episode. Thus, later experience is generated by the updated harness rather than an unchanged initial policy. We empirically evaluate whether this process produces useful improvements below.
AuthorsOpenReview ↗

Response to Reviewer GohK (2/5)

3. Because multiple navigation tools can successfully complete the same subgoal, tool selection generally lacks a unique ground-truth label. We therefore assess the navigation policy produced by successive RSI-EG updates through task success (SR) and success-weighted path efficiency (SPL). We track both metrics over 0–600 development episodes on VLN-CE, OGN, and OmniNavBench, averaging OGN results across HM3D and MP3D.

We also evaluate context-conditioned historical-utility guidance Qij(n)Q_{ij}^{(n)} through the w/o Qij(n)Q_{ij}^{(n)} ablation in Table 1. Although resembling conventional experience-based memory, this guidance addresses a specific challenge of embodied navigation. Whereas digital agents often receive structured states and tool feedback, embodied agents face perceptual errors and motion-induced observation changes that can degrade visual, geometric, and execution-feedback inputs, leading to decisions without sufficient evidence. RSI-EG supplements VLM reasoning with historical utility estimated through rule-based experience retrieval, providing a harness-level constraint intended to reduce unreliable judgments and unsupported reasoning.

For the full method (with Qij(n)Q_{ij}^{(n)}), increasing experience from 0 to 600 episodes raises SR/SPL from 36.0%/32.2% to 62.0%/52.6% on VLN-CE, from 44.3%/19.6% to 57.8%/25.7% on OGN, and from 12.5%/8.6% to 28.8%/21.2% on OmniNavBench. Starting from the same initial results, w/o Qij(n)Q_{ij}^{(n)} reaches only 40.0%/32.7%, 48.8%/19.9%, and 18.8%/12.3%, respectively, with smaller gains and partial declines in later rounds. At 600 episodes, the full method exceeds this variant by an average of 13.7% in SR and 11.5% in SPL. We have added the corresponding performance curves to Fig. 4 of the revised manuscript.

These results support cumulative improvement in the navigation policy through successive RSI-EG updates and the contribution of historical-utility guidance alongside experience-informed reasoning. They also support our design motivation: constraining unreliable VLM decisions under perceptual uncertainty helps translate accumulated experience into larger, more sustained navigation gains.

Table 1. Navigation performance across successive episodes. The full method (with Qij(n)Q_{ij}^{(n)}) combines experience-informed VLM reasoning with context-conditioned historical value. The w/o Qij(n)Q_{ij}^{(n)} variant retains experience-informed agent reasoning but removes this Qij(n)Q_{ij}^{(n)}. Column headings indicate accumulated episodes; all SR and SPL values are percentages.

ModeTaskMetric0120240360480600
with Qij(n)Q_{ij}^{(n)}VLN-CESR36.053.057.060.062.062.0
SPL32.247.249.851.652.552.6
OGNSR44.353.155.456.957.657.8
SPL19.622.423.124.625.825.7
OmniNavBenchSR12.522.525.027.528.828.8
SPL8.614.916.517.621.221.2
w/o Qij(n)Q_{ij}^{(n)}VLN-CESR36.039.041.043.043.040.0
SPL32.234.033.635.635.232.7
OGNSR44.346.047.849.249.248.8
SPL19.620.120.721.321.319.9
OmniNavBenchSR12.516.317.520.020.018.8
SPL8.610.211.813.212.812.3
AuthorsOpenReview ↗

Response to Reviewer GohK (3/5)

4. We thank the reviewer for this constructive suggestion and have added related work on RSI to the revised manuscript:

Recursive Self-Improvement (RSI) broadly concerns systems using their own capabilities and feedback to improve themselves, with definitions varying in what is improved and how recursion is realized. Mechanism-level approaches explore self-modifying weights and the optimization of programs or algorithms that drive subsequent improvement [irie2022selfreferential, zelikman2024stop, ishibashi2025selfdeveloping, zhang2026hyperagents, liu2026evox]. Agent-level approaches enable cumulative improvement through execution-logic modification, successor evaluation, and evolutionary search [yin2025godel, zhang2026darwin, wang2026huxley], as well as experience sharing [weng2026gea] and memory-based adaptation [ouyang2026reasoningbank]. Self-generated rewards and interaction data also support iterative model improvement [yuan2024selfrewarding, wu2025metarewarding, zhang2026rsir]. Consistent with our harness-level view of RSI, SpecForge iteratively refines external behavioral specifications using execution feedback while keeping model weights and the search procedure fixed [cheng2026specforge]. However, embodied RSI must account for spatial constraints and context-dependent tool effectiveness [liu2026mobiagent], which RING-Nav addresses through geometry-aware experience retrieval and adaptive tool decision fusion.

W2

W2 The difference from conventional experience-based memory is unclear. The current mechanism appears closely related to retrieval-augmented decision-making. For example, Expel [1] leverages experience accumulated in previous tasks to improve performance without fine-tuning. The authors should clarify the difference between RSI and experience retrieval methods.

Our RSI recursively connects experience updates with subsequent experience acquisition. In ExpeL’s reported protocol [2], experiences collected through ReAct and Reflexion support insight extraction and retrieval-augmented evaluation; the resulting cross-task knowledge is not fed back into another collection round. In RING-Nav, the experience graph guides tool decisions, their execution outcomes update the graph, and the updated graph guides subsequent decisions and experience acquisition. Thus, each updated harness generates experience for its own subsequent update, establishing the harness-level RSI defined in Q1. For fair comparison with zero-shot methods, RSI-EG is frozen during evaluation on VLN-CE, OGN, and OmniNavBench, with no test interactions written back.

[2] ExpeL: LLM Agents Are Experiential Learners (AAAI 2024).

Reviewer_jz3JOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses generalist embodied navigation, where existing agentic navigation methods built on pretrained VLMs remain largely task-specific and lack a mechanism for reusing experience across tasks and environments. RING-Nav is a training-free recursive self-improving (RSI) harness around a frozen VLM. It maintains a Contextualized Navigation State (CNS) that unifies task progress, spatial evidence, episodic memory, and tool feedback, so the frozen VLM can dynamically orchestrate a shared set of navigation tools (spatially grounded tool orchestration, SGTO). An RSI experience graph (RSI-EG) records contextualized tool transitions and their outcomes, retrieves relevant past experience, and guides later tool decisions without updating model parameters.

Critical strengths
  1. Timely problem and a cleanly decomposed framework design. The work tackles a hard and timely problem—training-free, cross-task and cross-environment experience reuse for embodied agents—and the architecture separates three concerns with explicit boundaries: state representation (CNS), spatially grounded tool orchestration (SGTO), and episodic experience accumulation (RSI-EG). The frozen VLM never receives parameter updates, tools are exposed through a unified MCP interface, and experience is stored as contextualized transitions rather than task-specific prompts. This decomposition is what lets one agent operate across both single-task navigation and heterogeneous multi-stage instructions, and it gives the community a reusable scaffold rather than a single tuned pipeline.
  2. Broad evaluation and traceable component attribution. Experiments cover three complementary regimes—instruction-following (R2R-CE), object navigation (HM3D, MP3D), and heterogeneous multi-stage navigation (OmniNavBench). Table 4's ablation isolates each component: removing CNS and RSI-EG drops SR from 24.0 to 12.70 on R2R-CE, while RSI-EG alone raises OmniNavBench SR from 12.50 to 28.75 and SPL from 8.41 to 21.24. The backbone ablation (GPT5.5 versus Qwen2.5-VL-7B) shows the harness is not tightly coupled to one VLM, which strengthens the generality claim.
Critical weaknesses
  1. Inference cost and efficiency are not reported. As a closed-loop harness that repeatedly invokes a VLM and dispatches navigation tools, RING-Nav's practicality depends on per-episode token cost, wall-clock time, and number of tool calls, none of which are reported. The paper compares SR and SPL against training-free baselines without a cost–performance trade-off, so a reader cannot tell whether the gains come from better orchestration or simply from heavier VLM invocation. This matters because training-free methods are typically judged by their inference budget as well as their accuracy.
  2. Experience-graph scaling, retrieval robustness, and negative transfer are not discussed. RSI-EG helps most on the heterogeneous OmniNavBench subset (SR 12.50 → 28.75) but only modestly on single-task benchmarks (R2R-CE 49.2 → 53.3); the paper does not explain this asymmetry. It also leaves open how the graph grows with episodes, how retrieval similarity is defined, whether stale or negatively correlated experience can be surfaced later, and how the system behaves over long task streams. These are open engineering questions rather than fatal flaws, but they limit how far the "recursive self-improving" claim should be generalized.
Questions for authors
  1. Could the authors report, per benchmark and per backbone (GPT4o and GPT5.5), the average per-episode token consumption (input and output separately), average wall-clock duration, and average number of tool calls? Please also provide a cost–performance comparison against the strongest training-free baselines in Tables 1–3, so reviewers can assess the inference budget behind the reported SR and SPL.
  2. How does the RSI-EG grow in size (nodes, edges, stored transitions) as the number of episodes increases? What similarity measure and top-k rule are used to retrieve past experience, and how are task, visual, geometry, and history features normalized before matching?
4 comments
AuthorsOpenReview ↗

Response to Reviewer jz3J (1/3)

We thank the reviewer for recognizing the importance of training-free experience reuse and for appreciating our clearly decomposed framework, broad evaluation, and component ablations. We address the reviewer’s concerns point by point below, with additional cumulative-improvement experiments, efficiency and failure analyses, and implementation details.

W1 and Q1

W1 Inference cost and efficiency are not reported. As a closed-loop harness that repeatedly invokes a VLM and dispatches navigation tools, RING-Nav's practicality depends on per-episode token cost, wall-clock time, and number of tool calls, none of which are reported. The paper compares SR and SPL against training-free baselines without a cost–performance trade-off, so a reader cannot tell whether the gains come from better orchestration or simply from heavier VLM invocation. This matters because training-free methods are typically judged by their inference budget as well as their accuracy.

Q1 Could the authors report, per benchmark and per backbone (GPT4o and GPT5.5), the average per-episode token consumption (input and output separately), average wall-clock duration, and average number of tool calls? Please also provide a cost-performance comparison against the strongest training-free baselines in Tables 1–3, so reviewers can assess the inference budget behind the reported SR and SPL.

We thank the reviewer for this constructive suggestion and have added efficiency experiments following the requested benchmarks and backbones. We report per-episode input/output tokens, VLM calls, and wall-clock duration across three benchmark with GPT4o/GPT5.5 backbone, measured on an Intel Core i9-14900KF workstation with one NVIDIA RTX 5090 GPU (32 GB VRAM). We compare Open-Nav on VLN-CE and ASCENT on ObjectNav using matched GPT-4o and GPT-5.5 backbones. EvoNav’s official implementation was unavailable for efficiency measurements. On OmniNavBench, the baselines use trained navigation policies; OmniNav directly predicts continuous waypoints and arrival/orientation outputs rather than autoregressive text, so its output-token count is marked “–”.

RING-Nav uses fewer VLM calls across all three tasks. On VLN-CE, with matched GPT-4o and GPT-5.5 backbones, it reduces total input/output tokens by 40.9–47.8%, VLM calls by 84.8–87.0%, and runtime by 30.7–35.6% relative to Open-Nav. On OGN, it uses more tokens than ASCENT but reduces calls by 51.0–52.7% and runtime by 14.2–16.3%, while supporting heterogeneous navigation tasks. On OmniNavBench, it requires 13.4/14.7 calls versus OmniNav’s 98.5 and reduces input tokens by 63.9–68.3%. Runtime is 16.1–26.9% longer, including remote API communication and service waiting, whereas OmniNav runs locally. We have added these results to Fig. 5 of the revised manuscript.

Table 1. Mean inference cost and efficiency per episode. Duration is measured in seconds. OmniNav uses its original local model; unavailable output-token usage is denoted by –.

TaskMethodModelInput tokensOutput tokensVLM callsDuration (s)
VLN-CEOpen-NavGPT-4o74,6207,153.0365.60294.97
GPT-5.578,77123,579.0066.93369.45
RING-NavGPT-4o43,6374,680.008.52204.27
GPT-5.547,8635,583.0010.15237.85
OGNASCENTGPT-4o2,441315.106.37215.14
GPT-5.52,5121,219.277.70230.60
RING-NavGPT-4o13,0341,726.003.12184.60
GPT-5.515,7201,986.003.64192.91
OmniNavBenchOmniNavLocal model211,927–98.53109.14
RING-NavGPT-4o67,1697,150.0013.40126.72
GPT-5.576,4018,085.0014.70138.45
AuthorsOpenReview ↗

Response to Reviewer jz3J (2/3)

W2 and Q2

W2 Experience-graph scaling, retrieval robustness, and negative transfer are not discussed. RSI-EG helps most on the heterogeneous OmniNavBench subset (SR 12.50 → 28.75) but only modestly on single-task benchmarks (R2R-CE 49.2 → 53.3); the paper does not explain this asymmetry. It also leaves open how the graph grows with episodes, how retrieval similarity is defined, whether stale or negatively correlated experience can be surfaced later, and how the system behaves over long task streams. These are open engineering questions rather than fatal flaws, but they limit how far the "recursive self-improving" claim should be generalized.

Q2 How does the RSI-EG grow in size (nodes, edges, stored transitions) as the number of episodes increases? What similarity measure and top-k rule are used to retrieve past experience, and how are task, visual, geometry, and history features normalized before matching?

We thank the reviewer for these constructive questions. We address them in three parts: (1) experience scaling, retrieval robustness, and negative transfer; (2) differences in performance gains across tasks; and (3) additional implementation details.

(1) Experience scaling, retrieval robustness, and negative transfer. We evaluate SR/SPL over 0–600 development episodes on VLN-CE, OGN, and OmniNavBench, averaging OGN results across HM3D and MP3D. Each updated graph guides the next collection round and is frozen during held-out evaluation. To evaluate the retrieval robustness, we designed the ablation output on Qij(n)Q_{ij}^{(n)}. In manuscript Eq. (7) historical utility Qij(n)Q_{ij}^{(n)} explicitly contributes to tool selection through the following fusion rule:

=\underset{(u_j,\xi)\in\mathcal{A}_t}{\arg\max} \left[ \left(1-\lambda_{ij}^{(n)}(x_t^{j,\xi})\right)L_t(j,\xi) +\lambda_{ij}^{(n)}(x_t^{j,\xi})Q_{ij}^{(n)}(x_t^{j,\xi}) \right].$$ The w/o $Q_{ij}^{(n)}$ variant retains experience-informed VLM reasoning but removes explicit historical-utility guidance. As shown in Table 2, the full method gains an average of 18.6%/13.0% in SR/SPL and exceeds w/o $Q_{ij}^{(n)}$ by 13.7%/11.5% at 600 episodes. Its performance generally improves before saturating, whereas w/o $Q_{ij}^{(n)}$ loses some gains in later rounds. These results support cumulative improvement and the contribution of utility guidance within the tested range. The corresponding curves are included in Fig. 4 of the revised manuscript. This work focuses on using retrieved experience to improve navigation performance. However, aggregate SR/SPL results do not isolate negative transfer from individual retrieved records. We will further investigate when experience misleads decisions to improve retrieval reliability and navigation performance. **Table 2.** Navigation performance across successive episodes. The full method (with $Q_{ij}^{(n)}$) combines experience-informed VLM reasoning with context-conditioned historical value. The w/o $Q_{ij}^{(n)}$ variant retains experience-informed agent reasoning but removes this $Q_{ij}^{(n)}$. Column headings indicate accumulated episodes; all SR and SPL values are percentages. | **Mode** | **Task** | **Metric** | **0** | **120** | **240** | **360** | **480** | **600** | | ----------------------- | ---------------- | ---------- | ----- | ------- | ------- | -------- | -------- | -------- | | **with $Q_{ij}^{(n)}$** | **VLN-CE** | **SR** | 36.0 | 53.0 | 57.0 | 60.0 | **62.0** | **62.0** | | | | **SPL** | 32.2 | 47.2 | 49.8 | 51.6 | 52.5 | **52.6** | | | **OGN** | **SR** | 44.3 | 53.1 | 55.4 | 56.9 | 57.6 | **57.8** | | | | **SPL** | 19.6 | 22.4 | 23.1 | 24.6 | **25.8** | 25.7 | | | **OmniNavBench** | **SR** | 12.5 | 22.5 | 25.0 | 27.5 | **28.8** | **28.8** | | | | **SPL** | 8.6 | 14.9 | 16.5 | 17.6 | **21.2** | **21.2** | | **w/o $Q_{ij}^{(n)}$** | **VLN-CE** | **SR** | 36.0 | 39.0 | 41.0 | **43.0** | **43.0** | 40.0 | | | | **SPL** | 32.2 | 34.0 | 33.6 | **35.6** | 35.2 | 32.7 | | | **OGN** | **SR** | 44.3 | 46.0 | 47.8 | **49.2** | **49.2** | 48.8 | | | | **SPL** | 19.6 | 20.1 | 20.7 | **21.3** | **21.3** | 19.9 | | | **OmniNavBench** | **SR** | 12.5 | 16.3 | 17.5 | **20.0** | **20.0** | 18.8 | | | | **SPL** | 8.6 | 10.2 | 11.8 | **13.2** | 12.8 | 12.3 |
AuthorsOpenReview ↗

Response to Reviewer jz3J (3/3)

(2) Differences in gains across tasks. We first clarify that 49.2% → 53.3% reflects the contribution of CNS on HM3D. In the component ablation, adding RSI-EG increases SR from 53.3% to 67.0% on HM3D, from 34.9% to 48.5% on MP3D, and from 37.0% to 62.0% on R2R-CE. Its benefits on individual tasks are therefore also substantial. OmniNavBench’s larger relative gain, from 12.50% to 28.75%, likely reflects its greater demand for heterogeneous tool coordination and its lower starting SR. OGN remains more dependent on perception and exploration, whereas RSI-EG improves tool decisions without updating the underlying perception modules. We have added this explanation to the component-ablation analysis in the revised manuscript.

(3) Additional implementation details. The revised manuscript provides these details at the following locations:

  1. Graph size, record growth, retention rules, and resource costs: Appendix A.6, Graph organization and growth, Eq. (8); Appendix A.1, Retrieval Time (Table 9) and Storage Overhead.
  2. Similarity measure and top-K retrieval: Sec. 3.3, Contextual experience retrieval, Eqs. (3)–(4).
  3. Task, visual, geometric, and history representations: Appendix A.4, Cross-episode experience construction.
  4. Retrieval and fusion hyperparameters: Appendix A.6, Hyperparameter settings, Table 14.
  5. Feature normalization and distance rules: Appendix A.6, Feature normalization and distance computation, Table 15.

Dear Reviewer,

We understand your busy schedule and wanted to check whether the revisions and details provided have addressed your concerns. Please let us know if any issues remain or if you need further clarification. Your feedback is invaluable in helping us improve our work.

Thank you for your time and effort in reviewing our paper.

Best regards,

Submission 927 authors

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

This paper studies generalist navigation across heterogeneous objectives and spatial contexts. RING-Nav combines a contextualized navigation state (CNS) with a cross-episode experience graph (RSI-EG) to guide tool selection by a frozen VLM. Experiments report improvements on instruction-following, object-goal, and heterogeneous multi-stage navigation.

Critical strengths
  • Coordinating heterogeneous navigation behaviors and reusing experience are practically relevant challenges.

  • The framework incorporates spatial evidence, local geometry, and execution feedback into tool selection and experience retrieval.

  • Experiments cover several navigation benchmarks, multiple VLM backbones, and component ablations.

Critical weaknesses
  • Limited methodological novelty. Beyond AgentVLN’s within-episode skill selection, RING-Nav introduces cross-episode experience reuse. However, experience storage, similarity-based retrieval, and experience-informed decisions are already common in agent research. Applying them to navigation with visual and geometric features does not sufficiently establish novelty. With fixed models, tools, and decision rules, the method largely resembles retrieval-augmented decision-making. The authors should identify navigation-specific challenges requiring new mechanisms and compare against standard experience-memory methods under matched budgets.

  • Unclear contribution of the graph structure. RSI-EG associates experience records with directed tool transitions, effectively indexing them by previous-tool–candidate-tool pairs. An ordinary experience database could provide similar functionality, and no multi-step graph reasoning or other graph-dependent mechanism is demonstrated. Table 4 establishes the usefulness of the complete experience module, but not its graph representation. A comparison with a flat experience store using identical records, features, and transition conditioning is needed.

  • Insufficient evidence for recursive self-improvement. The experience graph is constructed from development episodes and frozen before testing. Tables 1–4 therefore demonstrate the benefit of a fixed memory, rather than cumulative improvement through recursive interaction. The RSI claim exceeds the evidence presented. Performance curves across successive experience-accumulation rounds, compared with one-time memory construction under equal interaction budgets, would better support this claim.

  • Limited deployment evidence and efficiency analysis. Heterogeneous navigation is evaluated only on the MP3D–AlienGo subset, without cross-robot or real-world validation. Decision latency, model-call costs, retrieval time, and storage growth are also unreported. Given the 28.75% complete-task success rate on OmniNavBench, the authors should analyze major failure sources and whether retrieved experience sometimes misleads decisions.

Questions for authors
  • Could the authors qualify their state-of-the-art claims? MP3D SPL and OmniNavBench SII are not the best reported results; metric-specific advantages and trade-offs should be stated accurately.
6 comments
AuthorsOpenReview ↗

Response to Reviewer EHff (1/5)

We thank the reviewer for recognizing the importance of coordinating heterogeneous navigation behaviors and RSI experience, and for appreciating our harness framework and extensive evaluation. We address all concerns through additional cumulative-improvement and cross-embodiment experiments, efficiency and failure analyses, and clarifications of our RSI formulation.

W1

W1 Limited methodological novelty. Beyond AgentVLN’s within-episode skill selection, RING-Nav introduces cross-episode experience reuse. However, experience storage, similarity-based retrieval, and experience-informed decisions are already common in agent research. Applying them to navigation with visual and geometric features does not sufficiently establish novelty. With fixed models, tools, and decision rules, the method largely resembles retrieval-augmented decision-making. The authors should identify navigation-specific challenges requiring new mechanisms and compare against standard experience-memory methods under matched budgets.

Heterogeneous navigation requires coordination beyond isolated capabilities. Instruction following, object search, and human interaction have different goals, constraints, and completion criteria. Each also presents distinct challenges: spatial grounding and localization uncertainty in VLN [1], incomplete scene context and false-positive target detections in ObjectNav [2], and tracking moving people under occlusion in human following [3]. Combining them within one heterogeneous instruction requires appropriate tool transitions and reliable progress assessment as the environment and task stage change [4]. Heterogeneous navigation therefore compounds task-specific perception and execution difficulties with the need for consistent coordination across goals.

RING-Nav extends task-specific agentic navigation to heterogeneous tool orchestration. Task-specific frameworks typically lack unified coordination of mixed navigation goals within one instruction [4]. RING-Nav addresses this through a unified decision context: CNS integrating the active subtask, spatial evidence, and execution feedback for feasible tool selection and completion checks. RSI-EG reuses tool-transition experience under comparable task, visual, geometric, and execution-history conditions. Historical utility is adaptively fused with current VLM reasoning to guide tool choices; execution outcomes update the graph, forming a recursive feedback loop that shapes subsequent decisions and experience acquisition.

Our RSI recursively connects experience updates with subsequent experience acquisition. Conventional experience retrieval uses stored experience to guide decisions, whereas RING-Nav additionally uses the updated graph to generate new experience that drives its subsequent updates. For example, ExpeL [5] collected experience supports knowledge extraction and retrieval-augmented evaluation. In RING-Nav, RSI denotes cumulative improvement through a feedback loop: the experience graph guides tool decisions, their execution outcomes update the graph, and the updated graph guides subsequent decisions and experience acquisition.

[1] AgentVLN: Towards Agentic Vision-and-Language Navigation (ECCV 2026).

[2] SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation (NeurIPS 2024).

[3] TrackVLA: Embodied Visual Tracking in the Wild (CoRL 2025).

[4] Beyond Isolation: A Unified Benchmark for General-Purpose Navigation (RSS 2026).

[5] ExpeL: LLM Agents Are Experiential Learners (AAAI 2024).

W2

W2 Unclear contribution of the graph structure. RSI-EG associates experience records with directed tool transitions, effectively indexing them by previous-tool–candidate-tool pairs. An ordinary experience database could provide similar functionality, and no multi-step graph reasoning or other graph-dependent mechanism is demonstrated. Table 4 establishes the usefulness of the complete experience module, but not its graph representation. A comparison with a flat experience store using identical records, features, and transition conditioning is needed.

We agree that a flat experience database with identical records, features, transition conditioning, and retrieval rules could provide equivalent functionality. We adopt a graph because it naturally expresses tools as nodes and their executed transitions as directed edges, with experience records associated with these transitions. This organizational choice aligns with established practices in robotics, such as the computational graph used in ROS [6] to represent software components and their communication relationships. In RSI-EG, graphs organize experiences. Our contribution lies in experience-guided decision-making and recursive update mechanisms, while the graph serves as the data structure through which we implement these concepts.

[6] Robot Operating System 2: Design, Architecture, and Uses in the Wild (Science Robotics 2022)

AuthorsOpenReview ↗

Response to Reviewer EHff (2/5)

W3

W3 Insufficient evidence for recursive self-improvement. The experience graph is constructed from development episodes and frozen before testing. Tables 1–4 therefore demonstrate the benefit of a fixed memory, rather than cumulative improvement through recursive interaction. The RSI claim exceeds the evidence presented. Performance curves across successive experience-accumulation rounds, compared with one-time memory construction under equal interaction budgets, would better support this claim.

We thank the reviewer for this constructive suggestion and the opportunity to further validate RSI-EG’s benefits. We evaluate RING-Nav over 0–600 development episodes on VLN-CE, OGN, and OmniNavBench, averaging OGN results across HM3D and MP3D. Each updated RSI-EG guides the next collection round and is frozen during held-out evaluation to prevent test-data leakage. Historical utility Qij(n)Q_{ij}^{(n)} explicitly contributes to tool selection through the following fusion rule:

=\underset{(u_j,\xi)\in\mathcal{A}_t}{\arg\max} \left[ \left(1-\lambda_{ij}^{(n)}(x_t^{j,\xi})\right)L_t(j,\xi) +\lambda_{ij}^{(n)}(x_t^{j,\xi})Q_{ij}^{(n)}(x_t^{j,\xi}) \right].$$ The w/o $Q_{ij}^{(n)}$ ablation removes this explicit utility guidance while retaining experience-informed VLM reasoning. **Table 1.** Navigation performance across accumulated development episodes, with and without $Q_{ij}^{(n)}$. SR and SPL are reported as percentages. | **Mode** | **Task** | **Metric** | **0** | **120** | **240** | **360** | **480** | **600** | | ----------------------- | ---------------- | ---------- | ----- | ------- | ------- | -------- | -------- | -------- | | **with $Q_{ij}^{(n)}$** | **VLN-CE** | **SR** | 36.0 | 53.0 | 57.0 | 60.0 | **62.0** | **62.0** | | | | **SPL** | 32.2 | 47.2 | 49.8 | 51.6 | 52.5 | **52.6** | | | **OGN** | **SR** | 44.3 | 53.1 | 55.4 | 56.9 | 57.6 | **57.8** | | | | **SPL** | 19.6 | 22.4 | 23.1 | 24.6 | **25.8** | 25.7 | | | **OmniNavBench** | **SR** | 12.5 | 22.5 | 25.0 | 27.5 | **28.8** | **28.8** | | | | **SPL** | 8.6 | 14.9 | 16.5 | 17.6 | **21.2** | **21.2** | | **w/o $Q_{ij}^{(n)}$** | **VLN-CE** | **SR** | 36.0 | 39.0 | 41.0 | **43.0** | **43.0** | 40.0 | | | | **SPL** | 32.2 | 34.0 | 33.6 | **35.6** | 35.2 | 32.7 | | | **OGN** | **SR** | 44.3 | 46.0 | 47.8 | **49.2** | **49.2** | 48.8 | | | | **SPL** | 19.6 | 20.1 | 20.7 | **21.3** | **21.3** | 19.9 | | | **OmniNavBench** | **SR** | 12.5 | 16.3 | 17.5 | **20.0** | **20.0** | 18.8 | | | | **SPL** | 8.6 | 10.2 | 11.8 | **13.2** | 12.8 | 12.3 | As shown in Table 1, the full method gains an average of 18.6%/13.0% in SR/SPL from 0 to 600 episodes and outperforms w/o $Q_{ij}^{(n)}$ by 13.7%/11.5% at 600 episodes. Its performance improves before saturating, while w/o $Q_{ij}^{(n)}$ loses some gains in later rounds. These results support cumulative RSI-EG improvement and the benefit of historical-utility guidance, which is designed to reduce unreliable VLM decisions under perceptual uncertainty. The corresponding performance curves are shown in Fig. 4 of the revised manuscript.
AuthorsOpenReview ↗

Response to Reviewer EHff (3/5)

W4

W4 Limited deployment evidence and efficiency analysis. Heterogeneous navigation is evaluated only on the MP3D–AlienGo subset, without cross-robot or real-world validation. Decision latency, model-call costs, retrieval time, and storage growth are also unreported. Given the 28.75% complete-task success rate on OmniNavBench, the authors should analyze major failure sources and whether retrieved experience sometimes misleads decisions.

We thank the reviewer for this constructive suggestion. We address the concern in four parts: (1) evaluation across three robot types, (2) inference cost and efficiency, (3) retrieval time and storage, and (4) failure analysis on OmniNavBench.

(1) Evaluation across robot types. We extend the original quadrupedal evaluation to include wheeled and humanoid robots in OmniNavBench. As shown in Table 2, RING-Nav achieves SR/SPL of 36.25%/24.39%, 26.25%/22.55%, and 28.75%/21.24% on wheeled, humanoid, and quadrupedal robots, respectively. It achieves the highest SR and SPL among the compared methods for all three robot types, exceeding OmniNav's SR by 35.00%, 23.75%, and 5.00%, respectively. These results provide additional evidence of its applicability across different simulated robot embodiments. We have updated Table 4 of the revised manuscript with these results.

Table 2. Performance comparison across three robot types in OmniNavBench. Higher is better for all metrics except SII. Bold indicates the best result for each robot type.

MethodRobotSR ↑SPL ↑OSR ↑CSR ↑SGC ↑HFS ↑HFR ↑SII ↓
MTU3DWheeled0.000.002.500.0025.432.3325.971.66
Humanoid0.000.007.500.0025.370.0021.418.13
Quadruped0.000.0010.002.5028.170.0014.198.52
PoliFormerWheeled0.000.001.250.0023.432.3325.152.17
Humanoid0.000.005.000.0023.050.0020.997.44
Quadruped0.000.0017.502.5024.2713.959.804.11
Uni-NaVidWheeled2.501.203.750.0035.3239.5355.878.77
Humanoid0.000.008.751.2533.2213.9539.9714.43
Quadruped3.751.8510.000.0029.454.6522.6613.70
OmniNavWheeled1.251.165.003.7534.1239.5361.8113.65
Humanoid2.501.503.752.5026.8118.6034.8820.69
Quadruped23.7520.4128.7510.0043.9116.2828.7719.08
RING-NavWheeled36.2524.3941.2516.2566.7456.1572.018.68
Humanoid26.2522.5532.5010.0043.2153.8069.417.42
Quadruped28.7521.2433.7513.7557.3353.4869.669.83

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers

Dear all,

We thank all reviewers for their constructive feedback and for recognizing the practical importance of unified navigation (EHff, jz3J, GohK), our clearly decomposed harness framework design (EHff, jz3J), the broad evaluation and component ablations (EHff, jz3J), and the promising gains on R2R-CE and OmniNavBench (jz3J, GohK). We have addressed all concerns raised by the reviewers. In response to this feedback, we summarize our main clarifications, additional experiments, and implementation details below:

  1. (EHff, GohK) RSI definition and methodological novelty. We clarify definition of the recursive self-improving in RING-Nav, heterogeneous navigation challenges, and differences from retrieval-augmented decision-making.
  2. (EHff, jz3J, GohK) Cumulative improvement. We add the experiments results with performance curve for various navigation tasks across consecutive episodes of experience accumulation, as well as an ablation experiment demonstrating the guiding role of historical utility.
  3. (GohK) Evaluation fairness and episode order. We add GOAT-Bench evaluation, and test three random episode orders with limited performance variation.
  4. (EHff) Cross-robot performance on OmniNavBench. We extend OmniNavBench evaluation to wheeled, humanoid, and quadrupedal robots.
  5. (EHff, jz3J) Efficiency and resource costs. We report input/output tokens, VLM calls, and runtime across benchmarks and backbones, alongside retrieval latency and storage measurements.
  6. (EHff, jz3J, GohK) Additional clarifications and analyses. We add OmniNavBench failure analysis and RSI-related work, clarify the graph’s organizational role, and explain asymmetry gains between tasks. We also discuss future negative-transfer analysis and revise broad state-of-the-art claims to reflect metric-specific trade-offs.
  7. (jz3J, GohK) Implementation details. We expand the appendix with graph-growth rules, retrieval and normalization details, hyperparameters, task-state and memory definitions, tool interfaces, and the harness prompt.

We have updated the manuscript, with revisions highlighted in blue throughout the main text and appendix. We hope these additions address the reviewers’ concerns and clarify our contributions. Detailed point-by-point responses follow below. We welcome further suggestions and appreciate the opportunity to improve our work.

Best regards,

Submission 927 authors

Dear Reviewers,

Please review the authors' rebuttal and the other reviewers' comments, and participate in the discussion before it closes on November 18 (AoE). Kindly indicate whether your concerns have been adequately addressed and raise any remaining questions or concerns.

Thank you for your time and contributions.

Best regards,

Area Chair

Dear Reviewers,

A quick reminder that the discussion period ends on November 18 (AoE). If you haven't already, please review the authors' rebuttal and provide a brief follow-up indicating whether your concerns have been adequately addressed or if further clarification is needed.

Your timely feedback would be greatly appreciated and will help facilitate the final decision.

Thank you for your time and contributions!

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.