- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
RSI-Nav tries to reduce the cost of language supervision for continuous vision-language navigation. A single shared Qwen3.5-9B handles instruction following, instruction generation, and spatial understanding. After supervised initialization, it processes offline trajectories in shards: generate instructions, execute them, filter by success and SPL, then train in both directions on accepted instructions paired with the original trajectories, with spatial-data replay mixed in. The main claimed contribution is that supervision is synchronized frequently with the evolving policy. The paper reports 68.5% and 72.2% success on R2R-CE and RxR-CE validation-unseen, and generated instructions are more useful downstream. Spatial performance stays strong overall but drops on three of four benchmarks after evolution (Tables 1 and 4).
Critical strengths
- The navigation gains over the same initialization are substantial. Table 1 shows gains of 15.6 and 17.5 percentage points in success rate on R2R-CE and RxR-CE, with improvements in SPL and RxR-CE nDTW as well. The final success rates are 3.1 and 2.9 points above the strongest listed results. These numbers support the practical value of the overall training pipeline with single-view RGB observations.
- Generated instructions are evaluated by their downstream training utility, which is the right test. Table 4 trains identical navigation models on the same trajectories annotated by different generators. RSI-Nav supervision gives 41.3% SR and 35.1% SPL, versus 36.8% and 30.4% for SRDF. That directly asks whether the generated language helps navigation. The paper also appropriately qualifies NLP-score comparisons across different observation settings.
Critical weaknesses
The evidence does not yet establish the specific synchronization/co-evolution contribution. SRDF already iterates instruction generation, navigation-based filtering, and generator/navigator refinement; LANA already jointly learns following and generation with shared encoders. RSI-Nav’s clearest distinctions are complete parameter sharing and finer-grained updates between disjoint trajectory shards. Table 3 is useful: on the same 64K candidate trajectories, the preferred shard size gets 57.2% SR versus 55.1% for the full-subset offline setting. But supervision freshness and the training schedule change together, and the paper does not report accepted-sample counts or optimizer/replay exposure. So the proposed explanation is not cleanly separated from other effects of the update schedule. Also, Tables 4–5 do not compare the generator before and after recursive evolution. Better utility than SRDF does not show that RSI-Nav’s own supervision improves progressively. The system is effective overall, but the evidence for the specific advance over existing data-flywheel methods is still limited.
Questions for authors
- Does an evolving generator beat a frozen cold-start generator when using the same trajectory shards, current-policy filtering, and comparable optimization/replay budgets? What are the accepted-sample counts and training budgets for the settings in Table 3?
- Which checkpoint produces the instructions evaluated in Table 4? How do the cold-start and evolved generators compare on a fixed trajectory set, especially in downstream training utility?
- Please specify the recursive trajectory pool’s sources, size, and scene splits, the numerical value of $ \tau_{\mathrm{SPL}} $, and the per-shard optimization and spatial-replay schedule.
5 comments
Q2. Instruction-Generation Checkpoints and Downstream Training Utility
We appreciate the reviewer’s close reading of Tables 4–5 and the question about which checkpoint produced the evaluated instructions. We also thank you for recognizing the value of evaluating generated language through downstream navigation training. Your request to compare our own generator before and after evolution helped us examine the progression of supervision quality more directly.
Clarification of Table 4
The instructions evaluated in Table 4 are generated by the same checkpoint used in the third row of Table 2 and the last row of Table 5. This checkpoint is obtained by fine-tuning Qwen3.5-9B jointly on VL2A and VA2L data, without spatial-understanding supervision or recursive evolution.
Controlled Comparison Across Checkpoints
Following the reviewer’s suggestion, we compare four saved generators—Cold Start, 64K Evolving, 512K Evolving, and Final—under consistent evaluation protocols.
For instruction-generation quality, NLP metrics are evaluated on the same trajectory set as Table 4.
To assess downstream training utility, each generator independently annotates the same randomly sampled 512K unlabeled trajectories, followed by execution-based filtering with identical criteria (SR = 1 and SPL ≥ 0.75). Each resulting dataset contains sufficient accepted samples to construct a matched training set of 512K navigation samples. We then fine-tune Qwen3.5-9B from the same initialization, using identical data volumes and optimization configurations.
| Generator | BLEU-4 ↑ | CIDEr ↑ | SPICE ↑ | Downstream SR ↑ | Downstream SPL ↑ |
|---|---|---|---|---|---|
| Cold Start | 28.4 | 50.2 | 20.5 | 36.8 | 30.5 |
| 64K Evolving | 28.6 | 50.1 | 21.0 | 39.2 | 32.9 |
| 512K Evolving | 29.3 | 50.8 | 21.9 | 41.5 | 35.2 |
| Final | 29.9 | 48.9 | 22.8 | 43.8 | 37.6 |
Evolution of Supervision Quality
The results show that conventional NLP metrics improve modestly and are not strictly monotonic. These metrics primarily measure similarity to human instructions and do not necessarily reflect their effectiveness as navigation supervision.
Downstream training utility improves consistently across all four checkpoints under the same trajectory pool, filtering criterion, training-data volume, and navigator initialization. Downstream SR increases from 36.8% to 43.8%, and SPL from 30.5% to 37.6%.
These results demonstrate that instructions generated by successive RSI-Nav checkpoints become increasingly useful for downstream navigation learning, providing additional evidence for the language–action co-evolution proposed in Sec. 3.2.
The reviewer’s question was helpful in directing our attention to this within-model comparison. Thank you for identifying where the original tables left the progression of supervision quality unclear. We will include these results in the revised manuscript.
Q3. Trajectory Pool, Filtering Threshold, and Training Schedule
Thank you for identifying the implementation details that needed a clearer explanation. We appreciate the care with which the reviewer examined the trajectory sources, filtering rule, and training schedule. These questions helped us recognize where the manuscript relied on details that should have been stated explicitly. We provide the requested information below.
Trajectory Pool and Scene Splits
Our recursive evolution uses approximately 2.37M unlabeled continuous shortest-path trajectories derived from HM3D training scenes.
Instruction-generation NLP evaluation and R2R-CE Val-Unseen navigation evaluation use MP3D unseen scenes, with no scene overlap between recursive training and evaluation. The recursive trajectory pool is separate from the multi-source supervised data used for cold-start initialization, described in Sec. 3.3.
SPL Filtering Threshold
We use a fixed τ_SPL = 0.75 throughout recursive evolution. Accepted samples satisfy SR = 1 and SPL ≥ 0.75.
The numerical threshold was omitted from Eq. (4) in the original manuscript. We also provide a controlled sensitivity comparison of 0.1, 0.75, and 1.0 in our response to reviewer ywxj.
Optimization and Spatial Replay
As explained in our response to Q1, each candidate batch processes 256 trajectories per GPU, and accepted VL2A/VA2L training samples accumulate across candidate batches in a persistent global queue.
Once the queue contains at least 24,576 training samples, we take 24,576 samples in total across VL2A and VA2L and combine them with 8,192 spatial-replay samples for one SFT update. Unused samples remain in the queue for subsequent updates.
Each SFT update uses:
- 32,768 training samples in total.
- One training epoch.
- A global batch size of 128.
- A constant learning rate.
- 256 optimizer steps.
- 8,192 spatial-replay samples drawn using a fixed seed.
The updated checkpoint is then used for subsequent instruction generation and execution-based verification.
We appreciate the reviewer’s effort to examine the details behind the reported results. Your questions helped us make the connection between the trajectory pool, acceptance rule, and actual training schedule more explicit. We will clarify these details in Secs. 3.2–3.3 of the revised manuscript.
Q1. Isolating the Contribution of Policy-Synchronized Instruction Generation
We sincerely thank the reviewer for the careful examination of our method and experiments. We appreciate your recognition of the navigation gains and the downstream utility of the generated instructions. Your observation about the interaction between supervision freshness and the training schedule is particularly helpful, and your suggestion of a frozen cold-start generator gave us a concrete way to investigate this question. We agree that the original Table 3 alone cannot fully distinguish these effects.
Training Schedule of Our Recursive Self-Improvement
Controlling the Training Budget. In Table 3, candidates/GPU denotes trajectories processed per generation and verification batch. Accepted VL2A/VA2L training samples accumulate across batches in a global queue. Each SFT update consumes 24,576 training samples plus 8,192 spatial-replay samples. Unused samples are retained. The 32,768-sample budget and optimizer configuration per update are fixed, while candidates needed to fill the queue depend on filtering outcomes. We will clarify this in the revised paper.
With 64 GPUs, a global batch size of 128, and one training epoch, each update corresponds to 256 optimizer steps. The same optimization configuration and replay allocation apply to Frozen and Evolving.
Frozen vs. Evolving Generator
Following the reviewer's suggestion, we compare Frozen and Evolving under the same 256-candidate-per-GPU setting as Table 3. Both use the same 64K trajectory pool, cold-start initialization, filtering criteria (SR = 1 and SPL ≥ 0.75), and queue-triggered SFT mechanism.
Navigation execution and filtering always use the current policy. The only difference is instruction generation. Frozen retains the cold-start generator throughout the experiment. Evolving uses the latest updated model to generate instructions for subsequent candidates.
Under the same 256-candidate configuration, Evolving outperforms Frozen by 1.6 SR points and 2.0 SPL points. Frozen provides only modest improvements over the 1024-candidate offline setting.
This comparison suggests that more frequent updates alone do not fully explain the improvement in Table 3. Refreshing instruction generation with the evolving policy provides an additional benefit, supporting the policy-synchronized bootstrapping mechanism described in Sec. 3.2.
Accepted-Sample Statistics
To address the reviewer's question about accepted-sample counts, we also recover the acceptance statistics for the original Table 3 settings. Each setting processes 65,536 candidate trajectories.
These settings yield relatively similar acceptance rates, suggesting that the observed performance differences cannot be simply explained by modest differences in the amount of accepted supervision.
We are grateful for the specificity of the reviewer's questions, which helped us present the comparison and training procedure more clearly. We will incorporate the frozen-generator comparison and the corresponding acceptance and training-budget statistics into the revised manuscript. We hope these results and clarifications address the concern about the contribution of policy synchronization.