acceptodds
Under review as a conference paper at ICLR 2027

Breaking Success-Path Collapse: Advancing Aerial Vision-and-Language Navigation via Active Environment Shaping

Abstract

Simulation-based trajectory synthesis has emerged as a promising route towards scalable training data for aerial vision-and-language navigation. However, existing generators, whether search-based planning or reinforcement learning algorithms, share a failure mode that we identify and term Success-Path Collapse. Specifically, trajectories converge onto a narrow set of near-optimal routes. This behavioral redundancy, masked by high task success, silently propagates into downstream training data and compromises the robustness of learned navigation policies. In this work, we propose a novel perspective to actively restructure the environment to elicit multimodal trajectory generation, while remaining fully compatible with existing data synthesis pipelines. We instantiate this perspective with Trajectory-Aware Environment Perturbation (TAEP), which diagnoses collapse structures from successful trajectories, including biased junction exits, over-visited hotspots, and redundant route clusters, and places targeted virtual obstacles that naturally turn the pursuit of successful trajectories into continual route discovery. Experiments in AirSim urban scenes show that TAEP achieves the strongest overall trajectory diversity across a suite of diagnostic metrics, outperforming vanilla PPO, intrinsic-reward exploration, and random perturbation. TAEP-generated demonstrations translate these gains into improved closed-loop VLA navigation, more than doubling vanilla PPO's success rate on the longest tasks and improving mean success rate by 6.5 percentage points over the strongest evaluated baseline. Code and data are available at https://anonymous.4open.science/r/TAEP-ICLR-22E1/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Aerial Vision-and-Language Navigation, Trajectory Diversity, Environment Shaping, Synthetic Demonstration Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YWrYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors present Trajectory-Aware Environment Perturbation (TAEP), a framework for improving trajectory diversity in aerial vision-and-language navigation. The core idea is to mitigate the success path collapse, where a navigation policy repeatedly generates similar successful routes, by modifying the environment. In experiments on two scenes, AirSim16 and AirSim23, TAEP achieves the best overall trajectory diversity.

Critical strengths

Environment perturbation is a practically meaningful idea for navigation data generation. TAEP changes traversal conditions so that alternative routes emerge under the original navigation objective, which is well aligned with applications in the real world.

Critical weaknesses
  1. The experimental scale is limited. The evaluation is restricted to only two AirSim scenes, and many of the main conclusions rely on AirSim16. In particular, downstream VLA evaluation is conducted only on AirSim16, while AirSim23 is evaluated only for long-distance trajectory diversity. Since the proposed method directly depends on scene topology, broader evaluation across diverse environments is needed.

  2. The improvement in trajectory diversity over Random and other baselines is not always clear. In Table 1, TAEP does not consistently outperform the baselines on individual metrics, and some differences are quite small. The main conclusion is largely driven by Mean Rank, while variance, confidence intervals, and statistical significance are not reported.

Questions for authors

Please address the weaknesses. 1) Why were only AirSim16 and AirSim23 evaluated? 2) Are the improvements in Table 1 statistically significant?

Reviewer_Gnz6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the problem of Success-Path Collapse in aerial vision and language navigation (VLN), where navigation policies achieve high success rates but collapse into similar trajectories. This introduces bias into the training data distribution and limits its diversity.
The authors propose Trajectory-Aware Environment Perturbation (TAEP), which analyzes successful trajectories and introduces virtual obstacles at frequently visited locations, dominant junction exits, and overlapping route segments to encourage exploration of alternative paths. The generated trajectories dataset are used to finetune a VLA model. Experiments are conducted in 2 AirSim environments for trajectory diversity evaluation and 1 environment for downstream VLA evaluation.

Critical strengths

The paper identifies a problem in navigation data generation, where high success rates do not necessarily mean sufficient trajectory diversity. The proposed approach is straightforward and could improve the diversity of automatically generated navigation demonstrations.

The authors evaluate multiple trajectory diversity metrics and downstream VLA navigation performance. The results demonstrate improvements over vanilla PPO and random environment perturbations. This provides evidence that the diverse trajectories that TAEP generated can benefit downstream VLN training.

Critical weaknesses
  1. Modifying environments by adding more obstacles to encourage exploration and improve policy robustness has been extensively studied in previous work, such as PAIRED [1] and ACCEL [2]. The proposed method mainly combines several straightforward obstacle placement heuristics to generate more diverse data. I do not find this sufficiently novel as a technical contribution.

  2. The evaluation is too limited. Trajectory diversity is evaluated in only two AirSim scenes, downstream VLA evaluation uses only 1 scene. This is insufficient to demonstrate the generalizability of the proposed method. I suspect that the effectiveness of TAEP may be enlarged by the limited diversity of the original training distribution. When training a VLA model on a large-scale dataset containing diverse scenes, the model may already acquire the behavioral diversity that TAEP aims to introduce and the benefit of TAEP may diminish substantially.

  3. The baseline comparisons are insufficient. There are many existing approaches for generating diverse navigation trajectories, but the authors mainly compare against vanilla PPO, MBIE-EB, and random obstacles. More comparisons are needed.

  4. The analysis of downstream improvements feels incomplete. Specifically, there are no ablation studies to show how much each individual perturbation module actually contributes to the VLA performance. Additionally, because action discretization varies across the demonstration datasets, making it really hard to tell if the performance gains are genuinely driven by trajectory diversity or just differences in the discretization.

  5. The downstream VLN task is simplified. It uses fixed coordinate-based instructions and explicit pose information. This make the experiments less convincing.


References:

[1] Dennis, M., Jaques, N., Vinitsky, E., Bayen, A., Russell, S., Critch, A., & Levine, S. (2020). Emergent complexity and zero-shot transfer via unsupervised environment design. Advances in Neural Information Processing Systems, 33. https://arxiv.org/abs/2012.02096

[2] Parker-Holder, J., Jiang, M., Dennis, M., Samvelyan, M., Foerster, J., Grefenstette, E., & Rocktäschel, T. (2022). Evolving curricula with regret-based environment design. Proceedings of the 39th International Conference on Machine Learning, 162, 17473–17498. https://proceedings.mlr.press/v162/parker-holder22a.html

Questions for authors
  1. What is the key technical novelty of TAEP compared with existing environment-design and obstacle-perturbation methods? Could the authors clarify why the proposed combination of heuristic obstacle-placement strategies represents a significant methodological contribution?
  2. Why is downstream VLA evaluation limited to one environment? Do the improvements generalize to previously unseen AirSim scenes with different layouts and obstacle distributions?

While the paper tackles a relevant problem and shows some promising downstream results, the method itself feels like a pretty standard application of existing environment-perturbation techniques. My main concern is the weak evaluation: the narrow experimental setup and lack of strong baseline comparisons make it hard to tell if this approach is genuinely significant or generalizable. Given the limited technical novelty and the weak empirical backing, I don't think it's ready for acceptance in its current state.

Reviewer_TL5LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies Success-Path Collapse in aerial navigation, where successful trajectory generators repeatedly follow similar routes. The authors propose TAEP, which identifies frequently used junctions and routes and introduces targeted virtual obstacles to encourage alternative paths. Experiments in AirSim show improved trajectory diversity and downstream VLA navigation, with average success rate increasing from 44.44% to 75.74%.

Critical strengths

1. Interesting problem and approach. The paper highlights an important limitation of successful trajectory generation. Reshaping the environment to encourage route diversity is a sensible alternative to relying solely on exploration rewards.

2. Meaningful downstream improvements. The experiments include comparisons with PPO, intrinsic-reward exploration, and matched random perturbations. TAEP improves average SR by 6.48 points over random perturbation, providing useful evidence for diagnosis-guided obstacle placement.

Critical weaknesses

1. Limited evaluation of language-guided navigation. The downstream experiments use a fixed instruction template with target coordinates. While this helps isolate trajectory diversity, it remains unclear whether the benefits extend to richer natural-language navigation.

2. Limited downstream generalization. Although trajectory diversity is evaluated in two AirSim scenes, downstream VLA experiments use only AirSim16 and one backbone, with 180 fixed start-goal pairs. Additional settings would help establish how broadly the improvements hold.

3. The perturbation intensity selection could be clearer. Table 1 reports the intensity with the best Mean Rank, apparently selected using the same test trajectories used for evaluation. Selecting the intensity on a separate validation set would make the comparison more convincing.

Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.