acceptodds
Under review as a conference paper at ICLR 2027

RSI-Nav: Recursive Self-Improvement with Proactive Preflection for Training-Free Object-Goal Navigation

Abstract

Training-Free Object-Goal Navigation (TF-OGN) requires embodied agents to explore and locate target objects in unseen environments without any task-specific training. To this end, recent methods leverage foundation models. \bchange{However, many of the compared TF-OGN methods do not explicitly maintain memory conditioned on outcomes across episodes, limiting systematic reuse of completed experience. In this paper, we propose RSI-Nav, a TF-OGN framework that \rsiupdate{performs recursive self-improvement (RSI) by using completed navigation outcomes to update the rule memory guiding future decisions. Specifically, we build an agentic rule memory by extracting actionable knowledge from past trajectories. We retrieve rules by combining trajectory-derived semantic credit and historical success through UCB-based utility and exploration. In addition, we introduce a memory-guided preflection module that forecasts potential outcomes before action, reducing inefficient exploration. Extensive experiments show that our method outperforms existing training-free baselines, achieving a 10.1% relative improvement in success rate on MP3D.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Object-Goal Navigation, Self-Evolving, Embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yXGAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RSI-Nav, a training-free Object-Goal Navigation system that combines retrospective rule memory with prospective “preflection.” After each episode, an LLM converts the trajectory, target category, and binary success outcome into natural-language navigation rules; semantic credit, historical success, and a UCB term are then used to retrieve rules in later episodes. Before each frontier decision, the LLM predicts observations and possible failures for up to three candidates and selects a waypoint using the retrieved rules. The rule memory persists across episodes within each benchmark dataset. RSI-Nav reports 67.3/33.9 SR/SPL on HM3D and 49.0/19.1 on MP3D, with module, retrieval, backbone, and cost analyses.

Critical strengths
  1. The same-backbone gains are consistent across two datasets and both standard metrics. Relative to reproduced ASCENT with Qwen2.5-7B, RSI-Nav improves HM3D by 1.8 SR/1.1 SPL points and MP3D by 3.9 SR/2.9 SPL points; the trend remains with Qwen3-8B.

  2. The system is efficient and meaningfully decomposed. The paper separately studies prospective and retrospective components and several retrieval strategies, while average LLM calls increase only from 8.7 for ASCENT to 9.6 for RSI-Nav.

Critical weaknesses
  1. RSI-Nav consumes validation outcomes across episodes, but the baselines are not shown to operate under the same protocol. Episode tt can use information derived from earlier validation outcomes, whereas the quoted baselines in Table 1 apparently cannot. This changes the task's information regime and makes the SOTA comparison potentially apples-to-oranges (Section 4.1.4, p. 8; Appendix A, p. 14; Table 1, p. 5).

  2. The core self-improvement mechanism is not reproducible from the current description. Rule attribution, statistic updates, category filtering, deduplication, invalid-rule criteria, and the treatment of rewritten rules are central to the method rather than peripheral implementation details (Section 3.4, pp. 5–7; Appendix A, pp. 12–14).

Questions for authors
  1. Does the Habitat evaluation policy permit the binary outcome from one validation episode to influence later validation episodes? For each baseline in Tables 1 and 3, was the method given exactly the same feedback, persistent state, official episode order, and reset schedule (Section 4.1.4, p. 8; Tables 1 and 3, pp. 5 and 9)?

  2. How exactly are prp_r, NrN_r, and SrS_r updated for existing rules? How are Followed_Rule_Indices, conflicting rules, target-category filtering, duplicate or invalid rules, memory capacity, and statistics for rewritten rules handled (Section 3.4, pp. 5–7; Appendix A, pp. 12–14)?

  3. Why does Table 1 list VLFM in both the learning-based and training-free groups, and what distinguishes those two configurations? Please also clarify the protocol compatibility and provenance of all quoted results (Table 1, p. 5).

1 comment

We sincerely thank reviewer for your comments. To address your questions, we (1) clarify the feedback and reset protocols used in Tables 1 and 3, (2) explain how rules are attributed, updated, and maintained, and (3) distinguish the two VLFM configurations in Table 1 and document the sources of the baseline results.

Response to W1 & Q1. What feedback does each method receive?

Habitat-Lab resets the environment and agent each episode, passes observations to agent.act, and collects metrics for scoring. Feeding the previous episode’s success flag to the agent is an extension of this interface. In our online cross-episode adaptation setting (Tables 1 and 3), each dataset starts with empty memory. After each episode in the official order, its binary outcome updates the memory available to later episodes.

Baseline protocols. Table 1 quotes published baseline results. For these rows, we can verify episode order, feedback, and reset schedules only where the sources document them. In Table 3, we ran ASCENT* and RSI-Nav on the same validation episodes in the same order, with the same simulator, perception stack, and language backbone. ASCENT* resets its episodic state and receives no previous episode’s outcome. Full RSI-Nav carries updated memory. We will label this protocol difference in Table 3.

Results with no feedback from earlier validation episodes. On the full benchmarks with Qwen3-8B, preflection alone (Table 2, setting 2) achieves 66.5 / 33.5 on HM3D and 47.4 / 18.4 on MP3D (SR / SPL, %). ASCENT* achieves 65.3 / 33.0 and 45.3 / 16.2. These gains use no cross-episode rules.

For frozen transfer, we collected 1,000 training episodes across 80 HM3D scenes at each checkpoint and evaluated all four checkpoints on the same 200 tasks across 20 held-out scenes

HM3D training episodesRulesHM3D validation SR (%)HM3D validation SPL (%)
0066.0035.18
25022766.5035.54
50044468.5034.82
1,00088569.0035.98

The final snapshot succeeded on 69.00, versus 66.00 with empty memory.

Response to W2 & Q2. When is a rule updated?

1. Attribution and usage. At each frontier decision, the planner retrieves up to five rules from the current target category. Followed_Rule_Indices refers to their 1-based positions. We collect valid cited rule IDs across the episode. If there are no valid citations anywhere in the episode, we use all rule IDs shown. Each attributed rule increments NrN_r once per episode.

2. Outcome and semantic credit. In the manuscript, prp_r denotes the fraction of attributed episodes that succeeded. The implementation adjusts each binary outcome using the rule’s stored semantic credit and variance, accumulates these adjusted rewards, and divides their sum by NrN_r for the UCB reward mean.

Semantic credit SrS_r updates when reflection first proposes a rule or proposes the same rule again. We compute signed changes in goal similarity over up to three support steps. Softmax weighting (temperature 0.30) favors gains on success and drops on failure. Stored credit starts at zero and updates to 80% of the previous value plus 20% of the new weighted credit. Reuse alone updates the count and cumulative reward and preserves the stored semantic credit.

3. Rule maintenance. We accept positive or negative proposals with nonempty text. A reflection labeled lucky adds no rule. Within a target category, a proposal with identical text and type updates the existing record and retains its history. Rewording creates a new record. UCB ranks coexisting conflicting rules and the planner selects a frontier using current observations. All rules are retained, with no capacity limit or eviction.

Response to Q3. Which VLFM results are quoted in Table 1?

The two VLFM rows use the same frontier-selection method but different low-level controllers. The original VLFM uses a pretrained PointNav policy to reach each selected waypoint, so we list it in the learning-based group. ApexNav reports a variant, VLFM*, that replaces PointNav with a shortest-path planner. We list this variant in the training-free group because its navigation controller requires no training.

Table 1 configurationLow-level navigationHM3D SR / SPL (%)MP3D SR / SPL (%)Source
VLFM (original)Pretrained PointNav policy52.5 / 30.436.4 / 17.5VLFM, Table I
VLFM*Shortest-path planner50.9 / 23.632.5 / 15.9ApexNav, Table I

Table 1 compares published baseline results on the HM3D-v0.1 and MP3D validation splits using success rate (SR) and success weighted by path length (SPL). Each baseline score is quoted for the configuration evaluated in its source paper.

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

This paper presents RSI-Nav, a training-free Object Goal Navigation framework that combines prospective preflection with cross-episode retrospective rule memory. The method distills completed navigation episodes and their outcomes into reusable rules, which are retrieved to guide decisions in subsequent episodes. The reported performance is competitive, particularly on MP3D, where RSI-Nav achieves 49.0% SR and 19.1% SPL. The idea of constructing reusable navigation rules online at test time is interesting and potentially valuable for continually improving embodied agents

Critical strengths

-Strong navigation performance, particularly on MP3D. RSI-Nav reports competitive results on both HM3D-v1 and MP3D, with especially clear improvements on MP3D over the baselines included in the paper. The ablations also suggest that prospective preflection remains effective even without cross-episode retrospective updates.

  • Interesting test-time rule construction framework. Rather than only storing raw observations or trajectories, RSI-Nav retrospectively converts completed episodes and their outcomes into reusable navigation rules. The combination of outcome-conditioned rule construction, utility-aware retrieval, and prospective decision-making is an interesting direction for continual improvement without parameter updates.
Critical weaknesses
  • The retrospective component introduces test-time ordering and evaluation-fairness concerns. Rules are accumulated sequentially from completed evaluation episodes, so later episodes have access to more experience than earlier ones, and the resulting performance may depend on the official episode order. The paper should report results over multiple shuffled episode orders with mean and standard deviation. It should also compare against a cleaner setting in which the rule memory is constructed beforehand from the training split and then frozen during evaluation. This would separate experience transfer from test-time adaptation and ordering effects.

  • Real-robot validation is missing. Because the paper emphasizes recursive improvement from embodied experience, physical-robot experiments would be valuable for determining whether the generated rules remain useful under perception noise, actuation errors, environmental changes, and imperfect episode-level success signals.

  • Several important references and comparisons are missing. In particular, TrajRAG is a direct competitor because it also continually accumulates experience from past navigation episodes and retrieves that experience during subsequent navigation. The distinction is that TrajRAG stores geometric-semantic trajectory experience, whereas RSI-Nav abstracts completed episodes and their outcomes into reusable rules. The related work and experiments should also discuss CogNav, ReMemNav, etc

TrajRAG, CVPR 2026 CogNav, ICCV 2025 SG-Nav, NeurIPS 2024 ApexNav, IEEE RA-L 2025

ReMemNav arXiv 2026 ConsistNav arXiv 2026 STRIVE arXiv 2026

Questions for authors
  • can the retrospective component be constructed from training set, for fairer comparisons? is it sensitive to the ordering of test episodes? what happens if you randomize or change the ordering of the episodes
  • if the sequence is ordered in a manner that does not construct useful rules afterwards, does it provide negative impacts on the performance? what if the agent immdiately tested on MP3D after HM3D without erasing its memory?
2 comments

We sincerely thank reviewer for your comments. To thoroughly address your concerns regarding (1) evaluation protocol and task-order sensitivity, (2) cross-dataset experience reuse and suboptimal rule behavior, (3) real-robot applicability under real-world noise, and (4) discussion and comparison with recent works, we conducted additional navigation episodes in Habitat. Below we respond to each point in detail.

Response to W1 & Q1: Training-Split Memory and Sensitivity to Episode Order

Response: We completely agree with your methodological point and executed both experimental protocols:

1. Multiple Shuffled Task Orders on HM3D val (200 tasks × 3 intra-category orders): Starting from an empty memory with online rule generation and UCB updates enabled, we ran the exact same 200 HM3D val tasks (20 scenes, 10 episodes per scene) under three distinct intra-category task permutations:

Episode Order ConditionSR (%) ↑SPL (%) ↑
Official Order67.00%35.50%
Shuffled Order (seed=2026)68.00%35.94%
Shuffled Order (seed=2027)66.50%35.94%
Avg.67.17% ± 0.76%35.80% ± 0.25%

Across all three orders, SR varies within a narrow 1.5% and SPL is nearly invariant (35.50% to 35.94%, std = 0.25%).

2. Pre-Building Memory on HM3D train and Strictly Frozen Evaluation on HM3D val: To strictly separate offline experience transfer from online test-time adaptation, we collected experience across 80 scenes in HM3D train (completely disjoint from the 20 scenes in HM3D val) and saved snapshots at 0, 250, 500, and 1,000 episodes. Each snapshot was then evaluated on the 200 HM3D val tasks in strictly frozen, read-only mode:

Memory Setting & Experience SourceTrain EpisodesDistilled RulesUpdating ProtocolSR (%) ↑SPL (%) ↑
Empty Memory (Preflection only)00Empty66.00%35.18%
Empty Memory(Continous Updating)00-~176Updating67.17% ± 0.76%35.80% ± 0.25%
Frozen Snapshot @ 250 eps250227Frozen66.50%35.54%
Frozen Snapshot @ 500 eps500444Frozen68.50%34.82%
Frozen Snapshot @ 1,000 eps1,000885Frozen69.00%35.98%

Comparing Frozen Empty (66.0% SR), Online Adaptation from Scratch (67.17% ± 0.76%), and Frozen Train-Split Transfer (69.0%) cleanly decouples online adaptation from pre-built frozen experience transfer (+3.0%).

Response to Q2: HM3D-to-MP3D Transfer with Frozen and Online Memory

Response: We executed the exact proposed cross-dataset experiment: we retained the 1,000-episode HM3D train memory snapshot and immediately evaluated it on 200 MP3D val tasks across 10 scenes covering all 6 shared categories, under both Frozen and Online Updating target-domain protocols:

Initial memoryMemory updates on MP3DSR (%) ↑SPL (%) ↑
EmptyFrozen45.0019.01
HM3D 1,000-episode snapshotFrozen42.0018.23
EmptyOnline44.0018.89
HM3D 1,000-episode snapshotOnline44.0018.65

With memory frozen, transferring the HM3D snapshot lowered SR by 3 percentage points relative to empty memory. Enabling online updates raised the transferred agent’s SR from 42% to 44% and SPL from 18.23% to 18.65%. The rule history shows how RSI-Nav downweights suboptimal rules: the mean reward of one inherited potted-plant rule fell by approximately half as it was used on MP3D.

Response to W2: Real-Robot Applicability and Robustness to Real-World Noise

Response: RSI-Nav's modular architecture is explicitly designed to handle feedback noise:

  1. Semantic Abstraction Mitigates Perception & Odometry Drift: Unlike metric trajectory caches or dense point-cloud memories that break under wheel slip or depth noise, RSI-Nav abstracts experience into room-and-object semantic rules (e.g., "prioritize areas containing seating furniture such as armchairs when searching for a couch"). Local metric drift does not invalidate high-level semantic co-occurrence relations.

  2. Map Filtering & Bounded-Horizon Control for Actuation Noise: As detailed in Section 3.3 and Appendix A.1, our occupancy mapper applies vertical obstacle filtering ([0.61,0.88] m[0.61, 0.88]\text{ m}), a 1.5 m21.5\text{ m}^2 obstacle-area threshold, 10510^5-pixel hole filling, and size-5 morphological erosion to suppress depth sensor noise, while low-level execution is bounded by an HH-step horizon with automatic early termination and replanning if a waypoint becomes unreachable.

Response to W3: Comparison with TrajRAG, ReMemNav, and Related Methods

TrajRAG is the closest comparison because both methods reuse experience across episodes. TrajRAG trains an encoder to retrieve geometric–semantic trajectory segments and uses Qwen3-32B for planning. RSI-Nav distills episode outcomes into natural-language rules, updates their utility after use, and applies them to frontier selection with Qwen3-8B. RSI-Nav achieves 4.8% higher SR on HM3D-v0.1 and 6.4% higher SR on MP3D.

STRIVE and ReMemNav use proprietary VLMs for different decisions. STRIVE uses Gemini to reason over objects, viewpoints, and rooms. ReMemNav uses Gemini to revise navigation decisions using current-episode observations and to verify candidate targets. ReMemNav reports relative declines of 17.1% in SR and 40.4% in SPL when switching from Gemini to the open-weight Qwen3-VL-4B. Gemini can also incur higher planning latency. Gemini with default thinking setting took 8.2× as long per planning query as Qwen3-VL-8B[1].

The table gives the additional comparisons on HM3D-v0.1 and MP3D using each paper’s reported results. ReMemNav uses up to 40 high-level decisions per episode.

MethodReasoning modelHM3D-v0.1 SR / SPL (%)MP3D SR / SPL (%)
TrajRAGQwen3-32B62.5 / 33.942.6 / 18.0
STRIVEGemini62.9 / 34.252.3 / 23.1
ReMemNavGemini60.0 / 36.849.8 / 24.3
RSI-NavQwen3-8B67.3 / 33.949.0 / 19.1

CogNav reports different episode counts. Its HM3D dataset description lists 2,000 validation episodes, while its failure analysis reports 277 failures in 1,008 episodes. For MP3D, CogNav describes approximately 1,800 episodes across 20 categories, where standard evaluation uses 2,195 episodes across 21 categories.

Our Table 1 already includes SG-Nav, which reasons over online 3D scene graphs, and ApexNav, which fuses target evidence to guide exploration. ConsistNav accumulates evidence across frames to decide when to pursue, verify, or stop at a target. We will cover these changes in the revised version.

[1] Jadelynn Dao et al. DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners? CoRL, 2026.

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

This paper proposes RSI-Nav, a training-free object-goal navigation framework that aims to improve navigation performance in unseen environments through cross-episode experience reuse and proactive preflection. The framework keeps the parameters of the vision and language foundation models frozen, enabling cross-episode adaptation through non-parametric rule memory updates without task-specific model training. Experiments on HM3D and MP3D achieve success rates of 67.3% and 49.0%, respectively, outperforming the compared training-free navigation methods. Module ablations, comparisons of rule retrieval strategies, and qualitative navigation visualizations further demonstrate that historical rule memory and proactive preflection contribute to improved navigation performance.

Critical strengths
  1. A well-motivated framework for cross-episode experience reuse. The paper addresses the limited utilization of cross-episode experience in existing training-free object-goal navigation methods, which may lead to repeated navigation errors. It introduces a unified framework that integrates retrospective rule generation, semantic credit assignment, UCB-based dynamic retrieval, and proactive preflection. Without updating the parameters of the foundation models, the proposed approach incorporates historical navigation outcomes into subsequent decisions through an external rule memory, providing a practically meaningful adaptation mechanism for embodied agents operating under limited interaction budgets.
  2. Competitive performance across multiple navigation benchmarks, supported by ablation studies. The proposed method outperforms the compared training-free navigation approaches on both HM3D and MP3D, with particularly notable improvements on MP3D. The module ablation studies demonstrate that both retrospective rule memory and proactive preflection contribute to improved navigation performance. Comparisons of rule retrieval strategies further show that combining semantic credit assignment with UCB-based selection outperforms random retrieval, FIFO retrieval, and single-signal retrieval strategies. In addition, qualitative navigation visualizations illustrate how historical experience-based rules influence frontier selection, enhancing the interpretability of the proposed approach.
Critical weaknesses
  1. Insufficient direct validation of the recursive self-improvement mechanism. The central claim of the paper is that recursive self-improvement can be achieved through cross-episode rule memory. However, the experiments primarily report aggregate navigation performance, without analyzing how the agent's performance evolves as experience accumulates. For example, the paper does not show how navigation success rates change with the number of completed episodes, nor does it systematically investigate the relationship between navigation performance and rule memory size, rule quality, or the number of successive memory updates. Therefore, while the existing results demonstrate that rule memory contributes to improved navigation performance, they do not sufficiently establish that the agent achieves recursive self-improvement through continuous experience accumulation. This limitation directly affects the paper's main conceptual contribution.
  2. Limited direct evidence for the effectiveness of proactive risk prediction and rule update mechanisms. Although the ablation studies show that proactive preflection improves navigation success rates, the paper lacks a quantitative evaluation of the accuracy of predicted failure risks, such as whether the agent can correctly identify high-risk exploration directions or whether it incorrectly excludes potentially valuable regions. Furthermore, the effectiveness of semantic credit assignment and UCB-based retrieval is mainly evaluated indirectly through aggregate SR and SPL metrics, with limited analysis of rule correctness, cross-environment transferability, and stability under different episode orders. Consequently, it remains unclear to what extent the observed improvements arise from accurate risk prediction and reliable experience selection, rather than from additional reasoning or the incorporation of historical information in general. In addition, RSI-Nav introduces extra LLM calls and computational overhead compared with ASCENT, while the trade-off between navigation improvements and inference costs is not sufficiently analyzed.
Questions for authors

Q1. How can recursive self-improvement through cross-episode experience accumulation be directly validated? The paper aims to achieve recursive self-improvement through cross-episode rule memory, but its effectiveness is primarily demonstrated through aggregate navigation performance. Could the authors provide an analysis of how navigation success rates evolve as the number of completed episodes increases? For example, comparing continuously updated memory, frozen memory, and no-memory settings using separate experience-collection and evaluation episodes would help determine whether continuous experience accumulation leads to progressive improvement, rather than merely benefiting from static rule retrieval. Q2. How accurate and reliable are the failure-risk predictions generated by proactive preflection? Could the authors provide a quantitative comparison between the predicted failure risks of candidate exploration frontiers and their actual navigation outcomes? In particular, how accurately does preflection identify unpromising exploration directions, and how frequently does it incorrectly exclude potentially valuable regions? Such evidence would help clarify whether the observed performance gains genuinely result from effective risk prediction. Q3. How sensitive is the proposed framework to episode ordering and rule memory updates? The paper updates rule memory sequentially following a fixed episode order. Since historical experiences influence subsequent decisions, different episode orders may lead to different rule sets and navigation outcomes. Could the authors evaluate the robustness of the proposed method by shuffling episode orders, repeating the experiments, and reporting performance variance? Additionally, could the authors analyze whether transferring rules across different tasks or environments consistently provides positive benefits? Q4. What is the trade-off between navigation improvements and additional computational costs? Compared with ASCENT, RSI-Nav introduces additional processes for proactive preflection and post-episode rule generation and updates. Could the authors provide a breakdown of the computational overhead associated with these two stages? Furthermore, considering the additional LLM calls, inference latency, and token consumption, do the resulting improvements in navigation performance remain practically advantageous?

2 comments

We sincerely thank reviewer for the comments. To directly address your concerns and questions regarding (1) direct verification of experience accumulation under disjoint collection and frozen evaluation, (2) reliability of preflection risk predictions, (3) sensitivity to task ordering and cross-environment transfer, and (4) phase-wise computational overhead, we conducted additional navigation episodes in Habitat. Below we provide point-by-point responses with complete empirical tables.

Response to W1 & Q1: Direct Evidence of Recursive Self-Improvement, Experience Scaling Curve, and Frozen vs. Online vs. Empty Memory

Response: Following your suggested experimental design, we strictly separated experience collection from test-time evaluation using disjoint scene splits on HM3D:

  1. Disjoint Experience Collection (HM3D train, 1,000 episodes across 80 scenes): We ran RSI-Nav on the 80 training scenes of HM3D train for 1,000 episodes and saved memory checkpoints at 0, 250, 500, and 1,000 episodes (containing 0, 227, 444, and 885 rules, respectively).

  2. Strictly Frozen Evaluation (HM3D val, 200 episodes across 20 unseen validation scenes): Each checkpoint was loaded in read-only mode (all rules, UCB statistics, and visual memories completely frozen during evaluation) and evaluated on the exact same 200 HM3D val tasks under a unified runtime configuration.

  3. Comparison with Online Continuously Updated Memory (HM3D val, 600 episodes across 3 task orders): On the exact same 200 HM3D val tasks, we also evaluated RSI-Nav starting from an empty memory with continuous online rule generation and UCB updates across three intra-category task orders.

Memory Setting & Experience SourceTrain EpisodesDistilled RulesUpdating ProtocolSR (%) ↑SPL (%) ↑
Empty Memory (Preflection only)00Empty66.00%35.18%
Empty Memory(Continous Updating)00-~176Updating67.17% ± 0.76%35.80% ± 0.25%
Frozen Snapshot @ 250 eps250227Frozen66.50%35.54%
Frozen Snapshot @ 500 eps500444Frozen68.50%34.82%
Frozen Snapshot @ 1,000 eps1,000885Frozen69.00%35.98%

Key Findings:

  • Monotonic Scaling with Accumulated Experience: As HM3D train experience grows from 0 → 250 → 500 → 1,000 episodes (and the rule bank expands from 0 → 227 → 444 → 885 rules), frozen validation SR increases monotonically from 66.0% → 66.5% → 68.5% → 69.0% (+3.0), and SPL improves from 35.18% → 35.98% (+0.80). (An initial standalone frozen run of the 1,000-episode snapshot also achieved 68.0% SR / 35.83% SPL.)

  • Empty vs. Online vs. Pre-Built Frozen Memory: Online updating from scratch improves over Empty Memory (67.2% vs. 66.0% SR, 35.80% vs. 35.18% SPL) as rules accumulate progressively over the 200 episodes. Initializing with the mature 1,000-episode train snapshot and keeping it frozen achieves the highest performance (69.0% SR / 35.98% SPL) because early validation episodes immediately benefit from 885 rules without cold-start exploration.

Response to W2 (Part 1) & Q2: Reliability of Preflection Risk Predictions and Rule Selection

Response: We evaluated preflection's risk-prediction reliability at both the benchmark ablation level and the trajectory log level:

  1. Path Efficiency (SPL) Directly Quantifies Low-Value Direction Filtering: Because SPL weights binary success by Lshortestmax⁡(Lactual,Lshortest)\frac{L_{\text{shortest}}}{\max(L_{\text{actual}}, L_{\text{shortest}})}, any erroneous exclusion of valid regions that forces detours or backtracking directly penalizes SPL. In Table 2, enabling Prospective Preflection alone (Setting 2 vs. Setting 1, with zero rules) improves SR by +1.8 (64.7% → 66.5%) and SPL by +1.0 (32.5% → 33.5%) on HM3D, and improves SR by +3.5 (43.9% → 47.4%) and SPL by +2.6 (15.8% → 18.4%) on MP3D. Adding rule memory on top of preflection (Setting 4 vs. Setting 2) further improves SR/SPL to 67.3% / 33.9% on HM3D and 49.0% / 19.1% on MP3D.

  2. Low Rate of Search-Misdirection Failures: In Appendix E (Figure 5), across all failed episodes, "Never Saw Target" (exhausting the 500-step budget before ever observing the target, which upper-bounds persistent exploration misdirection) accounts for only 12.3% of failures on HM3D (4.0% of all 2,000 episodes) and 10.9% of failures on MP3D (5.6% of all 2,195 episodes), whereas 2D detector False Positives (51.4% HM3D, 35.0% MP3D) and geometric/stair Infeasibility (36.3% HM3D, 43.7% MP3D) dominate.

Response to Q3: Sensitivity to Task Order and Cross-Task / Cross-Environment Rule Transfer

Response: We ran 1,400 episodes across two dedicated experiments to evaluate (A) task-order sensitivity and (B) cross-dataset transfer (HM3D train → MP3D val).

  1. Task-Order Sensitivity on HM3D val (600 episodes = 200 tasks × 3 intra-category orders):

Starting from an empty memory with online rule generation and UCB updates enabled, we evaluated the exact same 200 HM3D val tasks under three distinct intra-category episode permutations (official order, random shuffle seed=2026, and random shuffle seed=2027):

Episode Order ConditionSR (%) ↑SPL (%) ↑
Official Order67.00%35.50%
Shuffled Order (seed=2026)68.00%35.94%
Shuffled Order (seed=2027)66.50%35.94%
Mean ± Sample Std67.17% ± 0.76%35.80% ± 0.25%

Across all three orders, SR varies within a narrow 1.5% and SPL is nearly invariant (35.50% to 35.94%, std = 0.25%).

  1. Cross-Dataset Experience Transfer from HM3D train to MP3D val:

HM3D and MP3D differ in the number of rooms an agent must consider when choosing where to search. For RSI-Nav, freezing memory preserves both the search rules and the scores they accumulated on HM3D. We therefore expect some inherited rules to retain a higher priority than their usefulness on MP3D warrants.

Initial memoryMemory updates on MP3DSR (%) ↑SPL (%) ↑
EmptyFrozen45.0019.01
HM3D@1,000-eps snapshotFrozen42.0018.23
EmptyOnline44.0018.89
HM3D@1,000-eps snapshotOnline44.0018.65

With memory frozen, transferring the HM3D snapshot lowered SR by 3% relative to empty memory. Enabling online updates raised SR and SPL**.** It shows RSI-Nav can handle the noises in rules automaticlly

Response to W2 (Part 2) & Q4: Phase-Wise Computational Cost Breakdown vs. Navigation Gains

Response: We instrumented our inference client across all 3 × 200 episodes of the HM3D order-sensitivity suite to record the exact per-episode LLM calls, wall-clock request latency, and total tokens (prompt + completion) for each stage:

Execution PhaseFrequencyAvg. LLM Calls / EpisodeAvg. Latency (s) / EpisodeAvg. Tokens / EpisodeShare of Total Tokens
1. Prospective Planning & PreflectionPer frontier decision3.417.99 s4,530.2076.40%
2. Intra-Episode Step SummarizationPer frontier decision3.42.71 s875.214.80%
3. Post-Episode Rule Distillation & UpdateOnce after episode ends0.862.03 s520.48.80%
Total LLM-related—7.6522.73 s5,925.80100.00%
Non-LLM Steps(Detections, Segmentations, etc.)——108.11s——

Relationship to Performance Gains over ASCENT:

  1. Zero Extra Calls During Intra-Episode Preflection: Prospective preflection and frontier selection are executed within the same single LLM call per frontier decision(ASCENT also need to call LLM here), avoiding separate reflection and action queries.

  2. Minimal Post-Episode Overhead (~8.8% tokens): Retrospective rule distillation requires only 0.86 LLM calls (2.03 s, 520.4 tokens) per episode (less than 1.0 because episodes flagged as lucky or trivial terminate early), matching the ~10% overall call increase reported on the full benchmark in Appendix B Table 5 (9.6 calls / 45.9 s / 10.5k tokens for RSI-Nav vs. 8.7 calls / 43.4 s / 9.3k tokens for ASCENT, and 3.2× faster than MSGNav at 24.8 calls / 151.4 s). Under the exact same LLM backbones (Table 3), this ~10% overhead yields +1.8 to +2.0 pp SR on HM3D and +3.7 to +3.9 pp SR / +2.9 pp SPL on MP3D over reproduced ASCENT.

  3. Step Summarization and Post-Episode Rule Generation Account for 2.41% of Episode Time. Across the 3 × 200 HM3D runs, an episode took 196.83 s on average. Intra-episode summarization and post-episode rule generation together took 4.74 s per episode (2.41%), of which post-episode rule generation accounted for 2.03 s (1.03%).

Comments on the paper

1 comment

We sincerely thank the reviewers for their valuable feedback and constructive comments. We would appreciate it if the reviewers could let us know whether our responses and revisions have adequately addressed their concerns. We would be happy to provide further clarification if any concerns remain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.