acceptodds
Under review as a conference paper at ICLR 2027

AirGroundVLN: A Large-Scale Benchmark for Goal-Oriented Air-Ground Collaborative Vision-and-Language Navigation

Abstract

Goal-oriented Vision-and-Language Navigation (VLN) requires agents to locate and reach targets described in natural language without prescribed routes. Air–ground collaboration is valuable for tasks requiring both wide-area search and fine-grained localization. However, systematic study of goal-oriented air–ground collaborative VLN remains limited by the lack of large-scale, diverse benchmarks and two core challenges: 1) substantial differences between aerial and ground views, together with useful observations becoming unavailable as navigation proceeds, make it difficult to maintain spatially consistent context across platforms and over time; and 2) asymmetric spatial observability makes ground perception locally detailed but spatially limited and aerial perception broad but locally coarse, limiting the reliability of single-platform planning. To address these limitations, we introduce AirGroundVLN, a benchmark containing 10,281 navigation episodes and 955 target instances across 19 Unreal Engine environments, with seen/unseen splits and an aerial-visibility protocol for systematic evaluation. Alongside the benchmark, we propose AG-CoNAV, a trainable reference framework comprising two key components: Spatiotemporally Anchored Collaborative Memory (SACM) and Aerial-Guided Regional-to-Local Planning (AGRLP). SACM maintains and retrieves spatially consistent historical context across aerial and ground observations. Meanwhile, AGRLP combines regional aerial guidance with fine-grained ground navigation. Extensive experiments demonstrate the effectiveness of AG-CoNAV and establish AirGroundVLN as a comprehensive benchmark for future exploration.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision and Language Navigation, Embodied AI, Air-ground collaboration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2e2eOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce AirGroundVLN, a large-scale benchmark for vision-and-language navigation (VLN) with collaborative UAV and UGV agents. Compared with prior aerial VLN benchmarks, AirGroundVLN focuses on air-ground collaboration, providing 10,281 navigation episodes across 19 Unreal Engine environments. This paper also proposes AG-CoNAV, a framework that incorporates Spatiotemporally Anchored Collaborative Memory (SACM) and Aerial-Guided Regional-to-Local Planning (AGRLP) with a Qwen VLM backbone. In the experiments, AG-CoNAV consistently outperforms both training-free and training-based baselines.

Critical strengths
  1. The problem setting with heterogeneous UAV-UGV collaboration for goal-oriented VLN is interesting. Since most prior work focuses on a single UAV agent, this setting opens a new avenue for studying collaboration under asymmetric perception and mobility constraints.

  2. The proposed method demonstrates strong empirical performance. The ablation results show that both SACM and AGRLP contribute to the final performance, with SACM providing a relatively large improvement.

Critical weaknesses
  1. I think the major weakness lies in the limited fundamental distinction from AeroDuo. AeroDuo also considers a collaborative navigation setting with two UAVs operating from complementary viewpoints, whereas AirGroundVLN appears to replace one of the UAVs with a UGV while retaining a similar hierarchical collaboration structure. The problem formulation is quite close, and the dataset scale is also comparable.

  2. It is also unclear whether the proposed method is specifically tailored to AirGroundVLN or whether it generalizes to other collaborative VLN settings. Evaluating AG-CoNAV on the AeroDuo benchmark would be valuable. This would also help clarify whether SACM and AGRLP provide benefits for collaborative navigation more broadly, rather than primarily exploiting the particular structure of AirGroundVLN.

Questions for authors

I have no critical questions but please address weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2e2e

Dear Reviewer 2e2e,

We sincerely thank you for your thoughtful evaluation and constructive feedback. We appreciate your positive assessment of the heterogeneous UAV–UGV collaboration setting and the contributions of SACM and AGRLP. We provide detailed responses to your concerns below.


W1. Clarifying the Distinction Between AirGroundVLN and AeroDuo

Thank you for highlighting the need to clarify the distinction between AirGroundVLN and AeroDuo.

(1) Benchmark and Task Formulation

(i) AeroDuo studies homogeneous collaboration between two UAVs operating at different altitudes and exploiting complementary aerial viewpoints. In contrast, AirGroundVLN studies heterogeneous air–ground collaboration, where aerial observations guide a UGV toward a language-specified, ground-reachable target. This is not merely a platform substitution, as the UGV operates under ground-specific navigation constraints and has a fundamentally different action space from a UAV.

(ii) AirGroundVLN focuses on ground-reachable targets and includes 955 target instances across six categories: humans, vehicles, urban infrastructure, daily-life services, industrial infrastructure, and other objects.

(iii) AirGroundVLN further introduces an aerial-visibility protocol that distinguishes aerial-visible and aerial-occluded targets, enabling an explicit evaluation of navigation under different aerial visibility conditions. For evaluation, AirGroundVLN contains 962 test-seen and 848 test-unseen episodes (1,810 in total), compared with 350 validation episodes in AeroDuo.

(2) Method Design

Although both methods adopt a hierarchical collaboration structure, their coordination mechanisms and planning objectives are fundamentally different.

(i) AeroDuo uses a high-altitude Pilot-LLM to predict target-region probability maps. As described in Sec. 4.2 of AeroDuo, the high-altitude UAV computes its flight direction and step length from language instructions and compass data. AG-CoNAV jointly predicts UAV and UGV waypoint distributions through a shared memory-augmented multimodal model.

(ii) AG-CoNAV combines collaborative memory and planning. SACM retrieves observations across views and time using spatial anchors, dispersion, and timestamps. The retrieved context informs platform-specific waypoint predictions, and AGRLP combines regional aerial guidance with local UGV planning under ground traversability constraints.


W2. Cross-Benchmark Evaluation on AeroDuo

Thank you for your valuable suggestion. To this end, we adapt AG-CoNAV to AeroDuo's two-UAV collaborative navigation setting and evaluate it on the HaL-13k benchmark using its official Unseen Map and Unseen Object splits. Specifically, we retain Qwen3.5-2B as the VLM backbone and replace the UGV with a low-altitude UAV to adapt AG-CoNAV to the two-UAV setting. For SACM, we store one memory entry every five frames, using the corresponding images and camera poses provided in action.json. The waypoint predictions are supervised using the ground-truth waypoints provided in gt_waypoints.json. For evaluation, we follow AeroDuo's original success criterion, using a 20 m success threshold instead of the 5 m threshold adopted in AirGroundVLN.

The results are reported below.

MethodUnseen Map SR(%)↑\uparrowUnseen Map OSR(%)↑\uparrowUnseen Object SR(%)↑\uparrowUnseen Object OSR(%)↑\uparrow
AeroDuo16.5728.5714.8619.43
AG-CoNAV (adapted)21.1431.4317.1420.57

Compared with AeroDuo, the adapted AG-CoNAV achieves consistent improvements on both the Unseen Map and Unseen Object splits. These results provide cross-benchmark evidence for the effectiveness of integrating collaborative memory and coordinated planning in multi-agent navigation.

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

This paper introduces AirGroundVLN, a benchmark for goal-oriented air–ground collaborative Vision-and-Language Navigation (VLN), and a trainable framework AG-CoNAV. Most existing VLN studies focus on a single platform, while heterogeneous air–ground collaboration remains less explored. AirGroundVLN contains 10,281 navigation episodes and 955 target instances across 19 environments. It provides seen and unseen environment splits to evaluate generalization, alongside aerial-visible and aerial-occluded subsets to assess navigation when targets are hidden from overhead views. The proposed AG-CoNAV integrates aerial and ground information through Spatiotemporally Anchored Collaborative Memory (SACM) and Aerial-Guided Regional-to-Local Planning (AGRLP). The UAV maintains a fused BEV map, while SACM compresses observations from both platforms into memory tokens associated with spatial anchors, dispersion, and timestamps. Relevant historical features are retrieved to augment waypoint queries derived from current observations and the instruction. The predicted aerial and ground waypoints are then passed to regional and local planners for execution.

Critical strengths
  1. AirGroundVLN provides a substantially larger collection of air–ground navigation episodes than existing benchmarks. Its environment and aerial-visibility splits support systematic evaluation of generalization and complementary perception, making it a useful resource for the community.

  2. SACM provides a coherent mechanism for preserving and retrieving information across aerial and ground views. Learned queries compress observations into spatially anchored memory tokens, allowing historical context to support subsequent waypoint prediction. AG-CoNAV achieves higher success rates than the evaluated training-free and ground-only baselines, and the ablations report positive contributions from both SACM and AGRLP.

Critical weaknesses
  1. The contribution of modules and their individual design choices are not sufficiently established. For AGRLP, it produces relatively small improvements in success rate under both seen and unseen settings in the ablation studies. For SACM, the authors argue that spatial anchors alone are insufficient and emphasize dispersion as an important addition. However, there is no direct support for the claimed necessity of dispersion.

  2. Several baselines are constructed by the authors, while the coverage of existing VLN approaches remains limited. A broader set of established methods under matched observation, training, and execution settings would help support AirGroundVLN as an effective evaluation benchmark and clarify the gains attributable to the design of AG-CoNAV.

Questions for authors
  1. Could you provide a controlled SACM ablation or other evidence to support the claimed necessity of dispersion in this module design.

  2. Could you provide a more complete reproducibility checklist, including training configurations, memory and waypoint parameters, map construction and other necessities for reproducing your work.

  3. Could you report results for additional existing VLN methods on AirGroundVLN or apply AG-CoNAV to existing VLN benchmarks to clarify the gains attributable to AG-CoNAV.

  4. Could you clarify the specific contribution of AGRLP in AG-CoNAV.

4 comments
AuthorsOpenReview ↗

Response to Reviewer QtAx (Part 1/2)

Dear Reviewer QtAx,

Thank you for taking the time to review our manuscript and provide constructive feedback. We appreciate your positive assessment of AirGroundVLN as a benchmark and the role of SACM in cross-view memory modeling. We have carefully considered your feedback and provide our detailed responses below.


Q1 & W1. Isolating the Contribution of Dispersion in SACM

The ablation study in Table 4 of the main manuscript evaluates the dispersion-aware spatial prior as a whole, but does not isolate the individual contribution of dispersion. To provide a more fine-grained ablation analysis, we introduce an Anchor-only variant that retains the same spatial anchors and temporal retrieval mechanism as the full SACM, but removes token-specific dispersion from the spatial retrieval score.

SACM variantSeen SR (%)↑\uparrowSeen OSR (%)↑\uparrowUnseen SR (%)↑\uparrowUnseen OSR (%)↑\uparrow
Anchor-only29.7332.5419.2222.64
Anchor + Dispersion (full SACM)32.6435.6520.0524.06

Compared with the Anchor-only variant, incorporating dispersion improves both SR and OSR on the seen and unseen splits. These consistent improvements support the effectiveness of modeling not only the spatial locations of compressed visual evidence but also its spatial extent, which is particularly important for aerial and ground observations with different spatial coverage.


Q2. Implementation Details for Reproducibility

We agree that providing more comprehensive implementation details is important for reproducibility. We clarify the key configurations as follows:

Training Configuration: AG-CoNAV is built upon the pretrained Qwen3.5-2B vision-language backbone. We keep the pretrained backbone weights frozen and train LoRA adapters (r=8r=8) together with the SACM modules using BF16 precision. We adopt the AdamW optimizer with an initial learning rate of 2×10−52\times10^{-5}, β=(0.9,0.999)\beta=(0.9, 0.999), and ϵ=10−8\epsilon=10^{-8}, together with a cosine learning-rate decay schedule. The per-device batch size is set to 1, with one optimizer update performed per completed episode. For waypoint supervision, the coordinate-level loss weight is set to λc=10.0\lambda_{c}=10.0, and the aerial loss weight is set to λa=0.5\lambda_{a}=0.5.

Memory Configuration: We maintain a FIFO memory bank that stores up to M=16M=16 historical air–ground entries. Each shared air–ground memory feature has a shape of 64×25664\times256.

Waypoint and Navigation Parameters: Each waypoint is represented as a spatial probability distribution. During inference, we identify the region containing the top-pp cumulative probability mass and compute its probability-weighted spatial centroid, with p=0.8p=0.8. The maximum navigation budget is set to Tmax⁡=200T_{\max}=200 steps. Once a target candidate is confirmed, the UGV enters the final approach stage, for which we set the stopping radius to dstop=1.5,md_{\mathrm{stop}}=1.5\\,\mathrm{m} and the maximum number of approach steps to Tapp=20T_{\mathrm{app}}=20. All methods use the same target verification and final approach procedures to ensure a fair comparison.

Map and Observation Configuration: The UGV constructs a local voxel map from six surrounding depth observations captured at 60° yaw intervals, with a voxel size of 0.5 m. During navigation, the UGV uses a front-facing camera for its observations. After executing each step, it performs a 360° sweep over six camera orientations for target verification. The aerial orthophoto is constructed by integrating multi-frame BEV depth observations, then resized and padded to 784×784784\times784 for input to the VLM.

We further confirm that, upon acceptance, we will publicly release our code and trained model checkpoints to facilitate reproducibility and future research.

AuthorsOpenReview ↗

Response to Reviewer QtAx (Part 2/2)

Q3 & W2. Expanded Baseline and Cross-Benchmark Evaluation

Thank you for this constructive suggestion. Air–ground collaborative VLN remains an emerging research direction, with few dedicated benchmarks and methods. Addressing this gap motivates the introduction of AirGroundVLN. To the best of our knowledge, AGC-VLN [1] (arXiv, 2026) is the only method among the published works and arXiv preprints we surveyed that specifically targets air–ground collaborative VLN and provides publicly available code. We also adapted AeroDuo [2] (ACM MM, 2025), an outdoor two-UAV collaborative navigation method, to AirGroundVLN to broaden our baseline comparison.

Following your suggestion, we further evaluate an adapted version of EMOS [3] (ICLR 2025), a heterogeneous multi-robot collaboration framework, on the same AirGroundVLN test splits. The results are reported below.

  • Expanded Baseline Evaluation on AirGroundVLN:
MethodSeen SR (%)↑\uparrowSeen OSR (%)↑\uparrowUnseen SR (%)↑\uparrowUnseen OSR (%)↑\uparrow
AGC-VLN13.4116.6310.5012.38
AeroDuo17.4622.4515.4519.69
EMOS12.8913.626.847.43
AG-CoNAV32.6435.6520.0524.06
  • Cross-Benchmark Evaluation on AeroDuo:

In addition to evaluating existing methods on AirGroundVLN, we further assess the applicability of AG-CoNAV to other collaborative navigation settings by adapting the framework to AeroDuo's two-UAV collaborative navigation task. We evaluate the adapted framework on the HaL-13k (AeroDuo) benchmark using its official Unseen Map and Unseen Object splits. Specifically, we retain Qwen3.5-2B as the VLM backbone and replace the UGV with a low-altitude UAV. For evaluation, we follow AeroDuo's original 20 m success threshold, rather than the 5 m threshold used in AirGroundVLN.

The results are reported below.

MethodUnseen Map SR(%)↑\uparrowUnseen Map OSR(%)↑\uparrowUnseen Object SR(%)↑\uparrowUnseen Object OSR(%)↑\uparrow
AeroDuo16.5728.5714.8619.43
AG-CoNAV (adapted)21.1431.4317.1420.57

Compared with AeroDuo, the adapted AG-CoNAV achieves consistent improvements on both the Unseen Map and Unseen Object splits, providing additional evidence for the applicability of the AG-CoNAV framework to multi-agent collaborative navigation. Together, these experiments broaden the evaluation of AG-CoNAV through both additional baselines on AirGroundVLN and cross-benchmark testing on AeroDuo.

References

[1] S. Zhang et al., “Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps,” arXiv, 2026.

[2] R. Wu et al., "AeroDuo: Aerial Duo for UAV-based Vision and Language Navigation," ACM Multimedia (ACM MM), 2025.

[3] D. Zhang et al., “EMOS: Embodiment-aware Heterogeneous Multi-robot Operating System with LLM Agents,” ICLR, 2025.


Q4. Role and Contribution of AGRLP

Thank you for the opportunity to clarify AGRLP’s role. The UAV’s bird’s-eye-view (BEV) map provides broad regional context, but traversability can be difficult to infer beneath tree canopies or overhangs. The UGV’s local depth map resolves nearby obstacles in greater detail, but its spatial coverage is constrained by the ground-level viewpoint and environmental occlusions. These complementary limitations motivate combining aerial guidance with ground-level planning.

(1) Method. AGRLP uses the aerial BEV map to establish a coarse route toward the predicted UGV waypoint. For example, when the waypoint lies beyond a wall, the aerial BEV map can reveal an opening and indicate which side to navigate around. Guided by this regional information, the UGV uses its local depth map to select a reachable subgoal along the coarse route and plan a local trajectory through observed free space. Thus, aerial guidance provides global direction, while ground-level planning ensures local traversability and helps avoid blocked regions.

(2) Performance. The ablation in Table 4 shows that adding AGRLP improves SR on both the seen and unseen splits, supporting the complementary role of aerial-guided regional routing and ground-level local planning.

Reviewer_QtAxOpenReview ↗

Response to the authors

Thank you for the additional experiments and detailed clarifications. The authors have addressed all my concerns, and I will maintain my score of 3.

Dear Reviewer QtAx,

We are pleased to learn that our additional experiments and clarifications have addressed your concerns. Your valuable comments have helped us further clarify our research methodology and improve the manuscript.

Once again, we thank you for the time and effort you have devoted to reviewing our work.

Best regards,

The Authors

Reviewer_bh8UOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces AirGroundVLN, a benchmark with 10,281 episodes and 955 target instances across 19 simulated environments, where a UAV helps a UGV locate and reach a language-specified target without prescribed routes. Its trainable framework, AG-CoNAV, associates historical aerial-ground observations through spatiotemporally anchored collaborative memory and combines aerial regional guidance with local ground planning.

Critical strengths
  1. The benchmark expands air-ground evaluation and covers environment generalization and top-down target occlusion.
  2. Memory preserves spatial centers and coverage, while planning combines aerial context with ground traversability.
  3. SACM improves seen and unseen SR by 5.71 and 2.12 percentage points, respectively, supporting the module's overall effectiveness.
Critical weaknesses
  1. The connection between trajectory construction and the UAV's role in gathering complementary information for ground navigation therefore remains unclear. The authors should clarify whether UAV demonstration trajectories select observation positions useful to the UGV, such as viewpoints revealing an intersection occluded from the ground.
  2. The relationship between SACM and prior UAV spatial memories [1,2,3,4] is insufficiently discussed, maybe can be discussed in related works.
  3. The authors should clarify how aerial and ground observations are aligned in a shared coordinate frame and whether SACM uses simulator-provided or estimated poses, depth, and camera parameters to define the input assumptions of its spatiotemporal anchoring.

[1] APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal Navigation. CVPRv 2026 [2] Memory-Augmented Scene Understanding and Exploration for Open-World Aerial Object-Goal Navigation. CVPR 2026. [3] Openfly: A comprehensive platform for aerial vision-language navigation. ICLR 2026. [4] GeoNav: Empowering MLLMs with Explicit Geospatial Reasoning Abilities for Language-Goal Aerial Navigation.Pattern Recognition.

Questions for authors
  1. Have you evaluated models of different sizes, such as 7B models? Would scaling up the model lead to better performance, particularly on unseen tasks, where the current 2B model shows only a marginal improvement over the baselines?
  2. Have you conducted any ablation studies on the hyperparameters?
  3. whether the target instances in the test environments are entirely distinct from those in the training environments, or whether there is any overlap
2 comments
AuthorsOpenReview ↗

Response to Reviewer bh8U (Part 1/2)

Dear Reviewer bh8U,

Thank you for your careful review of our manuscript and constructive feedback. We appreciate your positive assessment of AirGroundVLN's benchmark design, the integration of spatial memory and ground-aware planning, and the effectiveness of SACM. We have carefully considered your comments and provide our detailed responses below.


W1. Are the Observation Positions Along UAV Demonstration Trajectories Useful to the UGV?

In AirGroundVLN, UAV demonstration viewpoints provide aerial observations beyond the UGV's current field of view. We quantify this complementary spatial coverage through a voxel-level analysis using simulator-provided depth and geometric information.

  • For each UGV frame, we use its depth observations to reconstruct a local 3D voxel map within a radius of RR around the UGV. We classify each voxel as visible, occupied, or unknown. We similarly reconstruct the aerial BEV observations on the same voxel grid and classify the aerial voxels into the same three states.
  • We measure the percentage of voxels that are unknown in the UGV’s local map but observed from the air.

$ C(R) = \frac{\left|\mathcal{V}_{g}^{\mathrm{unknown}}(R) \cap \mathcal{V}_{a}^{\mathrm{known}}\right|}{\left|\mathcal{V}_{g}^{\mathrm{unknown}}(R)\right|} \times 100\% $

Here, Vgunknown(R)\mathcal{V}_{g}^{\mathrm{unknown}}(R) is the set of ground-unknown voxels within radius RR, and Vaknown\mathcal{V}_{a}^{\mathrm{known}} includes both visible and occupied aerial voxels.

Radius RR (m)Ground-Unknown Voxels Covered by Aerial Observations (%)
2081.86%
3080.48%
5078.05%

Across the three evaluated radii, observations along the UAV demonstration trajectories cover 78.05%–81.86% of the voxels that remain unknown from the ground. These results indicate that the observation positions along the UAV demonstration trajectories in AirGroundVLN are useful for ground navigation, as the aerial observations collected at these positions cover a substantial portion of the space that remains unknown from the UGV's local observations.


W2. Discussion of the Relationship Between SACM and Prior UAV Spatial Memory Methods

The four studies explore different ways to use memory for single-UAV navigation.

APEX [1] maintains dynamic 3D voxel maps to guide UAV exploration and navigation. OctMem-Agent [2] aggregates observations into an adaptive 3D octree and uses instruction-guided queries for target localization and exploration. OpenFly-Agent [3] selects informative keyframes and compresses their visual tokens to retain observation history. GeoNav [4] combines a global schematic cognitive map with a local hierarchical scene graph for coarse-to-fine navigation and target localization.

For air–ground collaboration, SACM represents UAV and UGV observations as compressed memory tokens in a common ground-plane coordinate frame. Each token is associated with a spatial anchor, spatial dispersion, and a timestamp. SACM retrieves tokens based on spatial proximity, token-specific dispersion, and temporal recency, enabling both agents to draw on observations from either platform when making navigation decisions.

We will incorporate this discussion into the Related Work section of the revised manuscript.


W3. Coordinate Alignment and Input Assumptions of SACM

Thank you for this valuable suggestion. We clarify SACM's input assumptions and coordinate alignment as follows.

(1) Input Assumptions. In our simulation setup, SACM uses depth observations and relative UAV–UGV poses provided by the simulator, rather than estimated ones. The camera intrinsics and camera-to-platform extrinsics are fixed and specified in the simulator configuration. This setup is consistent with prior simulation-based navigation studies [1,2,3].

(2) Coordinate Alignment. Using these inputs, we first back-project aerial and ground image patches into 3D using their corresponding depth observations and camera intrinsics. We then transform the resulting 3D patch locations into a shared coordinate frame using platform poses and camera-to-platform extrinsics.

We will clarify these input assumptions and alignment steps in Sec. 5.1 (Experimental Setup).

References

[1] X. Wang et al., “Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology,” ICLR, 2025.

[2] R. Wu et al., “AeroDuo: Aerial Duo for UAV-based Vision and Language Navigation,” ACM Multimedia (ACM MM), 2025.

[3] D. Zhang et al., “APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal Navigation,” CVPR, 2026.

AuthorsOpenReview ↗

Response to Reviewer bh8U (Part 2/2)

Q1. Model Scaling and Unseen-Environment Performance

Thank you for this constructive suggestion. We select Qwen3.5-4B (as the Qwen3.5 family does not provide publicly available 7B weights) as a larger backbone and compare it with the original Qwen3.5-2B variant on the same unseen test split to examine the effect of model scaling. The results are shown below.

MethodSeen SR (%)↑\uparrowSeen OSR (%)↑\uparrowUnseen SR (%)↑\uparrowUnseen OSR (%)↑\uparrow
AG-CoNAV-2B32.6435.6520.0524.06
AG-CoNAV-4B35.0337.2124.6529.13

Compared with the 2B variant, scaling up to 4B improves both SR and OSR on the seen and unseen splits, indicating that a larger backbone can further improve AG-CoNAV's navigation performance.


Q2: Hyperparameter Sensitivity Analysis

To assess the sensitivity of AG-CoNAV to key hyperparameters, we conduct a sensitivity analysis of the spatial tolerance (\rho) in SACM's spatiotemporal memory retrieval (Eq. 3). We evaluate (ρ∈0,5,10\rho\in{0,5,10}), with (ρ=5\rho=5) as the default setting. Setting (ρ=0\rho=0) removes the fixed tolerance term, leaving only token-specific dispersion in the denominator. All other hyperparameters are kept at their default values.

Spatial tolerance ρ\rhoSeen SR (%)↑\uparrowSeen OSR (%)↑\uparrowUnseen SR (%)↑\uparrowUnseen OSR (%)↑\uparrowOverall SR (%)↑\uparrowOverall OSR (%)↑\uparrow
030.3533.4721.1125.2426.02‾\underline{26.02}29.61
5 (default)32.6435.6520.05‾\underline{20.05}24.0626.7430.22
1031.70‾\underline{31.70}35.14‾\underline{35.14}19.5824.29‾\underline{24.29}26.02‾\underline{26.02}30.06‾\underline{30.06}

The results show that the default setting (ρ=5\rho=5) achieves the highest overall SR and OSR, as well as the best performance on the seen split. Although (ρ=0\rho=0) yields better results on the unseen split, the default setting provides a favorable balance across seen and unseen environments.


Q3: Target-Instance Overlap Between Training and Test Splits

We clarify that AirGroundVLN adopts an environment-level seen/unseen split to evaluate different aspects of navigation generalization.

  1. For the test-unseen split, there is no overlap with the training set in either environments or target instances. Specifically, the test-unseen split contains 848 episodes from two held-out environments. This setting is designed to evaluate the model's ability to generalize to previously unseen environments and target instances.

  2. For the test-seen split, the evaluation episodes are collected from the 17 environments used during training, and there is overlap in target instances between this split and the training set. However, the evaluation episodes differ in their initial agent positions, distances to the targets, and relative target directions. Thus, this split evaluates the model's ability to generalize to different start–goal configurations within familiar environments, requiring spatial reasoning and navigation planning under varying conditions.

Therefore, the two splits assess complementary aspects of generalization: test-seen evaluates navigation under new episode configurations in familiar environments, whereas test-unseen evaluates generalization to held-out environments and target instances.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.