Dual-Altitude UAV Collaborative VLN

• The paper demonstrates that splitting aerial VLN tasks into high-altitude global mapping and low-altitude precise navigation improves success rates by 9.71% over single-UAV methods.
• DuAl-VLN reformulates the problem by employing dual UAVs that exchange minimal map-based guidance, balancing coarse environmental reasoning with fine obstacle avoidance.
• The AeroDuo architecture integrates Pilot-LLM for global target probability mapping with a Multi-Stage Pathfinder validated on the HaL-13k dataset, highlighting robust performance across diverse urban environments.

Searching arXiv for the cited aerial VLN and multi-UAV papers to ground the article in the latest relevant literature. Search 1: arXiv (Wu et al., 21 Aug 2025) AeroDuo and Dual-Altitude UAV Collaborative VLN. Dual-Altitude UAV Collaborative Vision-and-Language Navigation (DuAl-VLN) is an aerial embodied AI task in which two UAVs at different altitudes collaborate to solve target-oriented vision-language navigation. In the formulation introduced with AeroDuo, the high-altitude UAV performs broad environmental reasoning and coarse target search, while the low-altitude UAV performs precise navigation, obstacle avoidance, and final target grounding. The task is motivated by the observation that single-UAV aerial VLN must simultaneously handle long-horizon trajectories, a large 3D motion space, and conflicting perceptual requirements: broad spatial coverage is favored by higher viewpoints, whereas precise grounding and safe local maneuvering are favored by lower viewpoints (Wu et al., 21 Aug 2025).

1. Emergence from aerial VLN

DuAl-VLN arose from the broader development of aerial VLN, which extends embodied language-guided navigation from ground agents to UAVs. AerialVLN established an outdoor UAV benchmark in 25 city-level environments with 8,446 routes/paths, 25,338 instructions, 661.8 m average path length, 204 actions per path on average, and 9.7 referenced objects per instruction on average, while emphasizing that aerial navigation is harder than ground VLN because agents must reason about flying height, complex spatial relationships, long trajectories, and 3D obstacle avoidance (Liu et al., 2023). The benchmark also exposed a large human-model gap: on Test Unseen, CMA achieved SR 1.6, whereas Human achieved SR 73.5 (Liu et al., 2023).

Subsequent single-UAV systems approached the problem with different architectural biases. UAV-VLN proposed a modular pipeline in which a fine-tuned TinyLlama-1.1B parses instructions into sub-goals, Grounding DINO performs open-vocabulary grounding, and a ROS2 task planner converts sub-goals into executable actions, but the formulation remains explicitly single-UAV and does not define dual-altitude collaboration or multi-agent coordination (Saxena et al., 30 Apr 2025). OpenVLN addressed data scarcity and long-horizon planning through RL-based VLM fine-tuning, a value model, and waypoint-level dense rewards, again in a single-UAV setting (Lin et al., 9 Nov 2025). ViSA-enhanced Aerial VLN replaced text-centric scene-graph reasoning with image-plane verification and structured visual prompting, but likewise remained a single-UAV framework despite its bird’s-eye-view reasoning design (Tong et al., 9 Mar 2026).

Against this background, DuAl-VLN reformulates aerial VLN from a single embodied policy into a collaborative, altitude-specialized problem. The central claim is not simply that UAV altitude matters, but that the perceptual and control burdens of aerial VLN are better split across two complementary agents: a high-altitude UAV for global reasoning and a low-altitude UAV for local execution (Wu et al., 21 Aug 2025).

2. Task formulation and operational semantics

In DuAl-VLN, the low-altitude UAV $U_l$ and high-altitude UAV $U_h$ begin from horizontally co-located positions with different heights:

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$

with

$$x^l_0 = x^h_0,\qquad y^l_0 = y^h_0,\qquad z^h_0 > z^l_0.$$

The system receives a target-oriented natural language instruction describing the target’s direction, characteristics, and surrounding environmental context (Wu et al., 21 Aug 2025).

The two agents operate at different temporal granularities. The low-altitude UAV uses step index $t$, while the high-altitude UAV uses step index $\tau$. This asymmetry reflects the division of labor: the high-altitude agent need not replan as frequently as the low-altitude executor (Wu et al., 21 Aug 2025).

Their observations are modality-specific. At low-level step $t$, $U_l$ observes a forward-facing RGB image $I^l_t$ and an omnidirectional point cloud $V^l_t$. At high-level step $U_h$0, $U_h$1 observes a BEV image $U_h$2 and a LiDAR point cloud map $U_h$3, with the paper stating that these modalities cover the same field of view for the high UAV (Wu et al., 21 Aug 2025). The episode succeeds when the low-altitude UAV comes within a distance threshold of the target location $U_h$4; in evaluation, the threshold used for SR and OSR is 20 m. Failure occurs if the low UAV exceeds the time limit without reaching the target or if the UAVs collide with obstacles (Wu et al., 21 Aug 2025).

A defining design choice is the communication interface. The two UAVs do not share dense feature tensors or raw observation histories. Instead, the high-altitude UAV provides compact guidance in the form of a target probability map $U_h$5 and a global depth map $U_h$6, from which the low-altitude UAV extracts subgoals and plans a path (Wu et al., 21 Aug 2025). This minimal-information interface is intended to preserve efficiency while retaining collaboration.

This suggests that DuAl-VLN is best understood as a decomposition of aerial VLN into two coupled subproblems: coarse target-region inference under global spatial context, and collision-aware local navigation with final object confirmation.

3. HaL-13k dataset

To support DuAl-VLN, AeroDuo introduced HaL-13k, described as the first dual-UAV VLN dataset with synchronized high-low-altitude trajectories. The dataset contains 13,838 collaborative trajectory pairs across 14 scenarios, built on the OpenUAV platform (Wu et al., 21 Aug 2025).

Its demonstrations are generated in two stages. First, low-altitude trajectories are produced by constructing an occupancy map from point clouds and running A* path planning under occupancy constraints, with the aim of keeping the low UAV at an effective exploration altitude. Second, the high-altitude trajectory is generated by randomly sampling flight paths under strict visibility constraints, so that the high UAV maintains full visual coverage of the low UAV route. This pairing is intended to ensure synchronized high/low trajectories, visual overlap, and a realistic collaborative relation between broad observation and local execution (Wu et al., 21 Aug 2025).

The instructions are intentionally target-oriented rather than route-detailed. They describe target orientation, target visual features, and surrounding environmental context, while avoiding highly detailed route descriptions and real-time human assistance (Wu et al., 21 Aug 2025). This linguistic design is important because it distinguishes DuAl-VLN from benchmarks centered on long imperative route descriptions; the task is framed more as language-guided target search with navigation than as literal route transcription.

HaL-13k includes paired high- and low-altitude trajectories, multimodal sensor streams, and target-oriented language instructions. The validation protocol includes two explicit generalization settings, each with 175 episodes: an Unseen Map Set, sampled from 2 scenes absent from training, and an Unseen Object Set, where scenes are familiar but object categories are unseen during training (Wu et al., 21 Aug 2025). These splits test both environment generalization and target/object generalization.

Some dataset characteristics remain unspecified in the available description, including the exact train/validation/test counts beyond the 175 + 175 validation subsets, the exact number of object categories, instruction-length statistics, and altitude distributions (Wu et al., 21 Aug 2025).

4. AeroDuo architecture

AeroDuo is the collaborative framework proposed for DuAl-VLN. It couples a high-altitude multimodal LLM, Pilot-LLM, with a low-altitude Multi-Stage Pathfinder (MSP) (Wu et al., 21 Aug 2025).

The high-altitude module first constructs a global orthographic representation from accumulated BEV observations, trajectory history, and point clouds:

$U_h$7

The orthophoto map $U_h$8, projected trajectory history, and instruction $U_h$9 are encoded and concatenated. Rather than asking the MLLM to emit direct coordinates, Pilot-LLM predicts a dense target likelihood over the map using learnable spatial mask tokens. The mask-token embedding at location $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$0 is

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$1

and the final target probability map is produced as

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$2

The paper argues that probability maps are preferable to direct coordinate generation because the target for a UAV should be a feasible region rather than a single point, and because MLLMs are weak at precise geospatial coordinate prediction (Wu et al., 21 Aug 2025).

The low-altitude MSP then converts this coarse guidance into executable motion through three stages. In Key Waypoint Decision (KWD), it computes a subgoal from the centroid of the probability map:

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$3

Using the high-altitude depth map and the altitude difference $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$4, it builds an occupancy map

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$5

runs A* with a Manhattan-distance heuristic and erosion, and segments the approximate route into key waypoints

$$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$6

In Collision-Free Navigation (CFN), the low UAV uses local point cloud $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$7, current subgoal, and ego state to predict the next velocity $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$8 with an RL-based controller trained by PPO in Isaac Sim. In Target Localization, it continuously applies GroundingDINO to its observations and stops when the detected target’s confidence exceeds a threshold, though the threshold itself is not specified (Wu et al., 21 Aug 2025).

Pilot-LLM is built from the visual projector $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$9 and LLM backbone of Qwen2-VL. The visual projector is frozen, the MLLM is fine-tuned with LoRA, optimization uses AdamW with a cosine scheduler and initial learning rate $$x^l_0 = x^h_0,\qquad y^l_0 = y^h_0,\qquad z^h_0 > z^l_0.$$0, and the prediction heads use binary cross-entropy for mask prediction and MSE loss for depth estimation (Wu et al., 21 Aug 2025). Its training includes pretraining on RefCOCO referring segmentation and depth estimation with labels from Depth Anything v2, plus BEV referring segmentation with 850,894 image-text pairs and BEV depth estimation. The low-altitude controller runs at 10 Hz, with each low-level step equal to 0.1 s (Wu et al., 21 Aug 2025).

A salient architectural feature is semantic-motion asymmetry. The high-altitude UAV reasons over a stitched global map and emits compact spatial guidance, while the low-altitude UAV performs local planning, collision avoidance, and target confirmation. This suggests that AeroDuo is less a symmetric multi-agent policy than a deliberately imbalanced collaboration between a global semantic planner and a local geometric executor.

5. Evaluation, results, and ablations

DuAl-VLN is evaluated with five metrics: Success Rate (SR), SPL, Success weighted by search time (SST),

$$x^l_0 = x^h_0,\qquad y^l_0 = y^h_0,\qquad z^h_0 > z^l_0.$$1

Oracle Success Rate (OSR), and Navigation Error (NE). The paper states that SST and SR are the main metrics (Wu et al., 21 Aug 2025).

All baselines are retrained on HaL-13k for fairness. On the Unseen Map split, AeroDuo achieves SST 14.63, SR 16.57, SPL 13.86, OSR 28.57, and NE 84.31. The strongest reported single-UAV comparator, TravelUAV (L1 assistant), achieves SST 6.48, SR 6.86, SPL 5.89, OSR 17.14, and NE 107.91; TravelUAV without the L1 assistant achieves SST 0.57, SR 0.57, SPL 0.57, OSR 1.14, and NE 152.20; CMA and Random are effectively near zero on success (Wu et al., 21 Aug 2025). On the Unseen Object split, AeroDuo reaches SST 13.54, SR 14.86, SPL 13.35, OSR 19.43, and NE 108.66, compared with TravelUAV (L1 assistant) at SST 5.31, SR 5.71, SPL 5.05, OSR 10.29, and NE 140.42 (Wu et al., 21 Aug 2025). The abstract summarizes the gain as an absolute 9.71% improvement in success rate over existing single-UAV methods (Wu et al., 21 Aug 2025).

The ablation study isolates four components: Pretrain, Global Map Construction (GMC), Key Waypoint Decision (KWD), and Collision-Free Navigation (CFN). With CFN only, average SR is 1.43. Adding Pretrain but still without GMC or KWD raises average SR to 2.29. Adding GMC yields the largest increase, bringing average SR to 14.29. Using Pretrain + GMC + KWD but removing CFN gives average SR 8.86. The full model reaches average SST 14.08 and average SR 15.71, with SR 16.57 on unseen maps and 14.86 on unseen objects (Wu et al., 21 Aug 2025). The paper explicitly notes that MLLM pretraining plus GMC gives an 11.41% increase in SST (Wu et al., 21 Aug 2025).

These results identify the main functional dependencies of DuAl-VLN. Local obstacle avoidance alone is insufficient; BEV/geospatial pretraining helps but is not enough on its own; global orthographic mapping is the largest single driver of performance; and waypoint-level decomposition still requires a local collision-free controller. The lower results on unseen objects than on unseen maps suggest that object generalization remains difficult.

6. Position in the literature, misconceptions, and open directions

DuAl-VLN sits at the intersection of aerial VLN and collaborative UAV autonomy, but it should not be conflated with either prior single-UAV VLN or generic multi-UAV planning. Single-UAV aerial VLN systems such as UAV-VLN, OpenVLN, and ViSA address semantic decomposition, open-vocabulary grounding, long-horizon planning, or image-plane verification, yet none formulates a multi-agent state, cross-altitude perception fusion, or a joint high/low-altitude objective (Saxena et al., 30 Apr 2025, Lin et al., 9 Nov 2025, Tong et al., 9 Mar 2026). Conversely, multi-UAV frameworks such as the HAPS-based hierarchical control architecture for integrated terrestrial and non-terrestrial networks and the MultiUAV-Plat benchmark provide collaboration, strategic-local role separation, or hidden-validation multi-agent task planning, but they do not study image-grounded language navigation, landmark-based aerial reasoning, or dual-altitude visual grounding (Yan et al., 6 Jun 2025, Zhang et al., 30 Jun 2026).

A common misconception is that DuAl-VLN is simply “aerial VLN with two drones.” The AeroDuo formulation is more specific. The high-altitude UAV is not a redundant copy of the low-altitude navigator; it is a map-level reasoner that predicts a target probability distribution and global depth support, while the low-altitude UAV is the only agent required to reach the final goal region and ground the target object (Wu et al., 21 Aug 2025). Another misconception is that collaboration in DuAl-VLN requires heavy inter-agent bandwidth. In the reported system, the agents exchange only minimal coordinate information, operationalized through compact map-based guidance rather than dense feature sharing (Wu et al., 21 Aug 2025).

The current formulation also has clear limits. The reported SR remains 16.57% on unseen maps and 14.86% on unseen objects, and the conclusion identifies execution efficiency and scalability as future directions (Wu et al., 21 Aug 2025). The paper does not report robustness to communication latency, more than two UAVs, or explicit bandwidth sweeps. It also does not study alternative altitude configurations or explicit ablations that remove the high-altitude UAV altogether. This suggests that DuAl-VLN, while a substantive reformulation of aerial VLN, remains an early-stage benchmark and systems perspective rather than a solved navigation regime.

Within the aerial VLN literature, its main significance is architectural. AerialVLN established the difficulty of long-horizon 3D instruction following; UAV-VLN showed that LLM-based semantic decomposition plus open-vocabulary grounding is promising for a single aerial agent; OpenVLN emphasized data efficiency and long-horizon value-guided planning; ViSA showed that image-plane spatial verification can outperform text-centric pipelines; and DuAl-VLN integrates a different insight, namely that broad search and precise execution may be better assigned to distinct altitudes and distinct agents (Liu et al., 2023, Saxena et al., 30 Apr 2025, Lin et al., 9 Nov 2025, Tong et al., 9 Mar 2026, Wu et al., 21 Aug 2025). This suggests that future progress may depend as much on viewpoint allocation and inter-agent abstraction as on stronger unimodal or monolithic navigation policies.