- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes AlignFly, a landmark-alignment reinforcement fine-tuning method for UAV vision-language navigation. An automated pipeline extracts ordered landmark phrases from instructions, scores their visibility in expert post-action observations using a VLM judge, and assigns landmarks to trajectory stages through monotonic dynamic programming. Starting from a supervised navigation policy, AlignFly samples candidate action chunks from the same expert state, executes them independently in simulation, and applies GRPO using landmark-evidence rewards together with distance, format, and stopping rewards.
Experiments train on 9,500 trajectories from one OpenFly scene and evaluate on 231 seen-scene trajectories and 231 trajectories from another scene. AlignFly reports success rates of 44.64% and 32.85%, compared with 41.13% and 28.26% for AirVLN-R1. Reward ablations, an additional experiment applying the procedure to OpenFly-Agent, and a manual audit of 100 annotated trajectories provide supporting evidence.
Critical strengths
-
I liked the improvements supported by the reward ablations. In Table 2, removing the landmark-alignment reward reduces SR from 44.64% to 39.94% on Test-Seen and from 32.85% to 24.94% on Test-Unseen, while keeping the other rewards and training settings unchanged. This supports the usefulness of landmark feedback. Applying the complete RFT procedure to OpenFly-Agent also improves SR by 5.96 and 7.32 percentage points, suggesting that its benefits are not limited to the authors’ own SFT initialization.
-
The automated annotation pipeline is another useful contribution. It provides stage-level landmark supervision from existing instructions and expert trajectories. The manual check in Appendix D supports its quality: 85 of 100 sampled trajectories agree with the human target-landmark assignments at every decision state, and another nine differ at only one state. These annotations are used during training, so the navigation policy does not require annotated target landmarks or expert future observations at inference.
-
I also liked that the landmark reward evaluates the visual outcomes of executed actions. Candidate action chunks are executed independently from the same expert state, allowing their resulting observations to be compared under the same initial conditions. This provides semantic feedback for viewpoint changes and landmark transitions that a reward based only on distance to the final goal may miss.
Critical weaknesses
-
My main concern is that the experiments do not isolate the benefit of ordered stage alignment. Removing the landmark reward also removes the semantic feedback from the VLM judge. This supports the usefulness of landmark feedback, but does not show how much the proposed alignment procedure contributes. I would like to see comparisons with simpler alternatives, such as rewarding instruction landmarks without considering their order, or selecting a landmark independently at each state without the monotonic constraint. This is particularly relevant because landmark-focused reinforcement rewards were already explored in Landmark-RxR (He et al., 2021). The automated annotations and VLM-based scoring are useful differences, but their added value needs clearer experimental support. The LE analysis uses the same judge as training and, as the authors acknowledge, provides diagnostic evidence.
-
I also could not reconcile some reported success rates with the evaluation setup. Section 4.1 states that each test split contains 231 trajectories and that results are averaged over three runs. If SR is computed as the equally weighted fraction of successful episodes, its three-run average should equal for an integer . The reported values of 44.64% and 32.85% cannot be obtained under this calculation, even after rounding to two decimal places. There may be an additional averaging step or a different denominator, but this needs clarification.
The paper also reports no variability estimates and does not clarify whether the seeds vary training, inference, or both. Together with evaluation on only one training scene and one held-out scene, this makes it difficult to assess how reliable and general the improvements are.
Questions for authors
-
Could you clarify how SR is calculated in Tables 1–3? Please provide the number of evaluated episodes and successful episodes in each run for AlignFly and AirVLN-R1, together with the averaging procedure. With 231 episodes per run and three equally weighted runs, the closest possible averages around 44.64% are 44.59% and 44.73%. Is there an additional averaging step or a different denominator?
-
To better understand the contribution of ordered stage alignment, could you compare AlignFly with two simpler alternatives: a landmark reward that ignores stage order, and independent landmark selection at each state without the monotonic constraint? Keeping the initialization, judge, and other RFT settings unchanged would help show how much the proposed alignment procedure contributes.
-
Do the three seeds vary training, inference, or both? Please report the variability across runs or confidence intervals for the main performance differences. Also, which scenes were used, how were they selected, and was a separate validation split used to select checkpoints and hyperparameters?
We thank the reviewer for identifying the inconsistency in the reported success rates and for requesting clearer experimental controls and variability estimates. To address these concerns, we corrected the evaluation denominator to include all scheduled episodes and updated Tables 1–3. We now report means and sample standard deviations alongside individual run results and success counts (Appendix I, Tables 7–10). We have also added controlled comparisons of target-landmark assignment strategies (Section 4.5, Table 4). Our responses clarify the scope of evaluation seeds, scene selection, data splits, checkpoint selection, and hyperparameter tuning. We address each question below.
1.The original evaluation script discarded episodes terminated due to unparseable predicted action sequences, which reduced the denominator and introduced statistical bias. Under the revised evaluation protocol, all scheduled episodes are included in the denominator; episodes terminated early by parsing errors are counted as failures. Each split contains 231 episodes per run. The overall success rate across the three replicated runs is calculated as the total number of successes divided by 693 (231 × 3). The corrected success rates are 44.59% and 32.76% for AlignFly on Test-Seen and Test-Unseen, respectively, and 41.13% and 28.14% for AirVLN-R1. We have updated Tables 1–3 and report individual per-run results in Appendix I.
2.We added a controlled comparison in Section 4.5 and Table 4. Independent visual matching selects the highest-scoring instruction landmark at each expert state without a sequence constraint. Uniform segmentation divides the expert trajectory into approximately equal-length intervals by decision step and assigns landmarks in instruction order. Both variants and monotonic DP share the SFT checkpoint, VLM judge, reward formula, other reward terms, and GRPO settings. Monotonic DP achieves mean SRs of 44.59% and 32.76% on Test-Seen and Test-Unseen. Its gains over independent matching are 5.34 and 9.52 percentage points, and its gains over uniform segmentation are 3.61 and 8.95 points, computed from the unrounded rates. These comparisons support using both visual evidence and instruction order to assign stage targets.
3.The three random seeds only vary action sampling during evaluation for fixed checkpoints. In the revised manuscript, Tables 1–4 report the mean and sample standard deviation for all four metrics, and Appendix I lists the individual results for each run. For example, AlignFly achieves an SR of 44.59 ± 0.75% on Test-Seen and 32.76 ± 0.66% on Test-Unseen. These standard deviations quantify the variability originating from evaluation sampling. We use the OpenFly dataset [1], which has also been adopted for training and evaluation by Xu et al. [2] and Ma et al. [3].The OpenFly-S subset used by Xu et al. contains both airsim_18 and airsim_16. In our work, we train on airsim_18 and reserve 231 trajectories from this scene for Test-Seen. airsim_16 is held out from training, and 231 trajectories sampled from it are used as Test-Unseen. This scene comparison evaluates the model’s generalization to one held-out layout under the same simulator engine. The training data is split 85%/15% into training and validation partitions. The initial learning rates and LoRA ranks for both stages are selected based on the validation loss on this fixed validation split. We initialize RFT with the final SFT checkpoint and evaluate the final RFT checkpoint.
References
[1] Gao et al. OpenFly: A Comprehensive Platform for Aerial Vision-Language Navigation. ICLR, 2026.
[2] Xu et al. Aerial Vision-Language Navigation With a Unified Framework for Spatial, Temporal and Embodied Reasoning. IEEE Transactions on Circuits and Systems for Video Technology, 2026.
[3] Ma et al. From Semantic Grounding to Decision Optimization: A Unified Framework for Long-Horizon UAV Vision-Language Navigation. ACM Multimedia, 2026.