AeroDuo: Dual-UAV VLN Framework

• AeroDuo is a dual-UAV framework that assigns global reasoning to high-altitude drones and precise navigation to low-altitude drones based on target-oriented language instructions.
• The system integrates a multimodal large language model (Pilot-LLM) with a lightweight navigation policy (MSP) and orthophoto map construction to guide collision-free navigation.
• Empirical results on the HaL-13k dataset show significant improvements in success rate and efficiency over single-UAV baselines, underscoring its practical benefits.

Searching arXiv for the named AeroDuo paper and a few closely related supporting papers mentioned in the source block. AeroDuo is a dual-UAV collaborative vision-and-language navigation framework introduced for Dual-Altitude UAV Collaborative VLN (DuAl-VLN), a task in which a high-altitude UAV performs broad environmental reasoning and a low-altitude UAV performs precise navigation, collision avoidance, and target grounding from target-oriented language instructions and visual observations (Wu et al., 21 Aug 2025). In the literature assembled here, the same label is also used more loosely for several paired aerial or aero-enabled architectures; however, its most explicit and formalized use is the 2025 framework that couples a multimodal LLM with a lightweight navigation policy, supported by the HaL-13k dataset of 13,838 collaborative high-low UAV demonstration trajectories (Wu et al., 21 Aug 2025).

1. Definition and conceptual scope

AeroDuo is defined by a division of labor across altitude. The high-altitude UAV, denoted $U_h$, is responsible for broad environmental coverage, global reasoning, and target-region estimation. The low-altitude UAV, denoted $U_l$, is responsible for precise navigation, collision avoidance, and target grounding. This separation is motivated by a central asymmetry in UAV vision-and-language navigation: high altitude provides multi-grained perspectives and wide field of view, whereas low altitude provides fine-grained perception and safer maneuvering in cluttered environments. The framework therefore constrains the low-altitude UAV’s effective motion space while still exploiting aerial mobility (Wu et al., 21 Aug 2025).

Within this formulation, AeroDuo is not a generic two-drone coordination scheme. It is specifically organized around target-oriented instructions rather than step-by-step route descriptions, and it is explicitly designed to avoid the human intervention, iterative dialogue, and overly detailed route-level guidance that prior UAV-VLN systems often require. A common misconception is that the method is simply redundant sensing with two identical agents. The architecture instead assigns distinct inference roles to the two platforms and limits communication to compact spatial outputs rather than raw sensory streams (Wu et al., 21 Aug 2025).

The task-level success criterion is also asymmetric. An episode succeeds only when $U_l$ comes within a distance threshold $d$ of the target location $p^d$ and grounds the target. Failure occurs if $U_l$ exceeds a navigation-time upper bound without reaching the target or if either UAV collides with obstacles. The collaboration is therefore “soft”: $U_h$ periodically reasons over the environment and sends guidance, while $U_l$ executes navigation between these updates (Wu et al., 21 Aug 2025).

2. Formal task structure and system architecture

At initialization, the two vehicles share horizontal position but differ in altitude: $$P^l_0 = (x^l_0, y^l_0, z^l_0), \qquad P^h_0 = (x^h_0, y^h_0, z^h_0),$$ with

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

They receive a target-oriented language instruction describing the target’s direction, visual characteristics, and surrounding environmental context. Time indices are separated as $U_l$0 for $U_l$1 and $U_l$2 for $U_l$3. At each low-altitude step, $U_l$4 captures a forward-facing RGB image $U_l$5 and an omnidirectional point cloud $U_l$6. At each high-altitude step, $U_l$7 captures a BEV-like RGB observation $U_l$8 and a LiDAR point cloud $U_l$9 covering the same field of view (Wu et al., 21 Aug 2025).

The high-altitude branch is built around Pilot-LLM. Rather than treating high-altitude observations as a temporal video only, AeroDuo constructs a global orthophoto map and a depth map: $U_l$0 The pipeline reconstructs an elevation map from accumulated point clouds, estimates camera extrinsics from UAV poses, reprojects BEV observations onto the ground plane, stitches the reprojected views into an orthographic map $U_l$1, and computes a global depth map $U_l$2. At most five historical images are stitched, after which the map is resized to a standard resolution before entering Pilot-LLM (Wu et al., 21 Aug 2025).

Pilot-LLM uses mask tokens to predict a continuous target probability map rather than explicit coordinates. The special tokens are defined as

$U_l$3

where $U_l$4 is the positional embedding at coordinate $U_l$5 and $U_l$6 is a trainable embedding. The multimodal input is

$U_l$7

and the resulting probability map is

$U_l$8

The outputs sent to $U_l$9 are the target probability map $d$0 and the global depth map $d$1 (Wu et al., 21 Aug 2025).

The low-altitude branch is the Multi-Stage Pathfinder (MSP), a three-stage policy implementing key waypoint decision, collision-free navigation, and target localization. MSP first computes a sub-goal as the centroid of the target probability distribution: $d$2 Using the altitude difference $d$3, it constructs an occupancy map from the high-altitude depth estimate,

$d$4

then applies A* search with a Manhattan-distance heuristic and erosion for safer obstacle clearance. The resulting coarse path is segmented into $d$5 uniformly spaced key waypoints $d$6. MSP then invokes an RL-based Navigator, inspired by NavRL, that uses 3D ray casting of the point cloud, an MLP controller, and PPO training in Isaac Sim. Target grounding is handled asynchronously by GroundingDINO on the latest RGB frame and the language instruction; exploration terminates when a detection exceeds a confidence threshold or when the waypoint sequence is exhausted and a new guidance request is issued (Wu et al., 21 Aug 2025).

3. HaL-13k and the learning pipeline

HaL-13k is the first dual-UAV VLN dataset designed for DuAl-VLN. It contains 13,838 collaborative dual-altitude trajectory pairs across 14 realistic outdoor scenarios from OpenUAV, with realistic UAV dynamics and multimodal sensory streams including RGB, BEV images, and LiDAR point clouds. Each trajectory pair is annotated with target-oriented language instructions that describe target orientation, visual features, and surrounding context, but do not provide step-by-step route descriptions or interactive assistance (Wu et al., 21 Aug 2025).

Trajectory construction proceeds in two stages. First, low-altitude expert paths are obtained by constructing an occupancy map from point clouds and running the A* algorithm to compute optimized paths that maintain effective exploration altitude while respecting occupancy constraints. Second, high-altitude trajectories are randomly sampled subject to visibility constraints so that the high-altitude UAV maintains visual coverage of the low-altitude route. The resulting paired trajectories are synchronized and overlapping in field of view, enabling the collaborative formulation (Wu et al., 21 Aug 2025).

The validation protocol explicitly separates environment generalization from object generalization. The unseen map set comprises two scenes absent from training and contains 175 test episodes. The unseen object set uses familiar scenes but novel object categories and also contains 175 test episodes. This split is designed to evaluate generalization across both novel layouts and unfamiliar targets (Wu et al., 21 Aug 2025).

Training is divided into pretraining and fine-tuning. Pretraining strengthens cross-modal understanding, especially for BEV images, geospatial reasoning, and depth perception. It includes referring segmentation and depth estimation on ground-view images, referring segmentation on BEV images using 850,894 image-text pairs, and depth estimation for BEV images. Fine-tuning then trains Pilot-LLM to predict target-region probability maps for navigation. The model uses Qwen2-VL as backbone, freezes the visual projector $d$7, fine-tunes the multimodal LLM parameters with LoRA, and optimizes with AdamW, a learning rate of $d$8, and a cosine scheduler. In parallel, the Navigator is trained with PPO in Isaac Sim using rewards that penalize collisions and abrupt velocity changes while rewarding progress toward the goal (Wu et al., 21 Aug 2025).

4. Empirical performance and ablation structure

Evaluation uses five metrics: Success Rate (SR), Success weighted by Path Length (SPL), Success weighted by Search Time (SST), Oracle Success Rate (OSR), and Navigation Error (NE). For DuAl-VLN, SR and SST are emphasized as the primary indicators of successful and efficient task completion. The baselines are Random, CMA, TravelUAV, and TravelUAV (L1 assistant), with all baselines retrained on HaL-13k for fair comparison (Wu et al., 21 Aug 2025).

The main results show a substantial gap between AeroDuo and the strongest single-UAV baseline.

Split	Strongest single-UAV baseline	AeroDuo
Unseen Map	TravelUAV (L1 assistant): SR 6.86%, SST 6.48%, NE 107.91 m	SR 16.57%, SST 14.63%, NE 84.31 m
Unseen Object	TravelUAV (L1 assistant): SR 5.71%, SST 5.31%, NE 140.42 m	SR 14.86%, SST 13.54%, NE 108.66 m

On unseen maps, AeroDuo achieves SR $d$9, SST $p^d$0, SPL $p^d$1, OSR $p^d$2, and NE $p^d$3 m. On unseen objects, it achieves SR $p^d$4, SST $p^d$5, SPL $p^d$6, OSR $p^d$7, and NE $p^d$8 m. The paper highlights a $p^d$9 absolute improvement in SR over the strongest single-UAV baseline on validation, depending on split (Wu et al., 21 Aug 2025).

The ablation results clarify the framework’s internal dependencies. Collision-free navigation alone yields only average SST $U_l$0 and SR $U_l$1. Adding pretraining improves these to SST $U_l$2 and SR $U_l$3. Adding global map construction raises performance sharply to SST $U_l$4 and SR $U_l$5. Replacing collision-free navigation with key waypoint decision gives SST $U_l$6 and SR $U_l$7. The full model—pretraining, global map construction, key waypoint decision, and collision-free navigation—achieves overall average SST $U_l$8 and SR $U_l$9. A second misconception is therefore that the multimodal LLM alone accounts for the gains; the empirical evidence instead attributes performance to the joint effect of orthophoto construction, probabilistic waypointing, RL navigation, and grounding (Wu et al., 21 Aug 2025).

5. Broader uses of the label and adjacent paired architectures

In the literature assembled here, “AeroDuo” is not a universally standardized cross-domain term. It is also used interpretively for architectures built around paired aerial roles, paired density layers, or paired vehicles. This suggests a broader design trope: a two-part system in which one element provides global reach, structural support, or coarse deceleration, while the other provides local service, precision, or fine preservation.

One such use appears in 5G networking, where a UAV-mounted distributed unit acts as an aerial DU and a terrestrial central unit remains on the ground, with the CU-DU interface complying with 3GPP F1AP and the aerial node providing integrated access and backhaul. The prototype is implemented with OpenAirInterface, a USRP B200 mini, a custom power amplifier, and a Quectel RM500Q-GL 5G module, and it demonstrates coverage extension and a reported $U_h$0 Mbps throughput for a user that is otherwise out of coverage (Mundlamuri et al., 2023).

A second use appears in planetary mission design, where a Photon cruise stage releases one or more small drag-modulation aerocapture vehicles. In that architecture, aerocapture replaces $U_h$1 to $U_h$2 km/s of orbit-insertion $U_h$3, enabling low circular $U_h$4 km orbits at Mars and Venus with small periapsis-raise maneuvers of $U_h$5 m/s and $U_h$6 m/s, respectively. The same study states that if each small satellite weighs $U_h$7 kg with the aerocapture system mass included, five satellites can be carried by a single Photon, making two-spacecraft “duo” missions a special case of a broader multi-orbiter pattern (Girija, 2023).

A third use is aerodynamic rather than robotic or orbital: a reusable booster employing aerosurfaces operates in two distinct modes, powered rocket ascent with fins and unpowered gliding descent with deployed wings. The model-scale design targets an expected maximum altitude of $U_h$8 m, uses a NACA 2414 airfoil, and reports best-glide estimates including $U_h$9 for the small booster and $U_l$0 for the Energia-2 side booster case (Dayanov, 2023).

A fourth use is materials-based. The Tanpopo capture medium is a dual-layer, dual-density hydrophobic silica aerogel tile with a $U_l$1 surface layer and a $U_l$2 open-topped, box-shaped base layer. Sixty tiles were manufactured, 57 were flight-qualified, and the configuration was designed to balance intact dust capture with the mechanical requirements of ISS flight hardware. Its density-optics relation is expressed as

$U_l$3

In that context, the “duo” resides in staged deceleration and structural framing rather than in multiple vehicles (Tabata et al., 2015).

6. Limitations, implementation pathways, and future development

AeroDuo’s reported performance remains far from saturation. Success rates of roughly $U_l$4 to $U_l$5 on unseen maps and unseen objects indicate meaningful progress over baselines but also substantial headroom before large-scale operational deployment. The failure modes implied in the paper include instruction ambiguity, extremely cluttered or occluded environments, out-of-map or non-visible targets, and domain shifts between OpenUAV, Isaac Sim, and real platforms. The authors also identify execution efficiency and scalability as future concerns (Wu et al., 21 Aug 2025).

The most immediate future directions are faster orthophoto construction, more lightweight LLM reasoning, improved real-time performance, richer collaboration, stronger grounding for small or occluded targets, and eventual transfer from simulation to real UAV platforms under weather, lighting, and sensor noise. The paper also points toward extension from dual-UAV collaboration to larger multi-UAV swarms with more altitudes or role allocations (Wu et al., 21 Aug 2025).

In a broader aerial-systems sense, dual-UAV execution would plausibly benefit from software substrates purpose-built for multi-robot aerial systems. Aerostack2 is one such framework: it is ROS 2 based, open-source, platform-independent, plugin-oriented, and explicitly designed for multi-robot aerial systems with behavior-based mission control, and it has been validated in both simulation and real flights across heterogeneous platforms. This does not mean that AeroDuo is implemented on Aerostack2 in the cited work; it indicates that the surrounding ecosystem already contains a mature systems framework for the class of multi-agent aerial deployments to which AeroDuo belongs (Fernandez-Cortizas et al., 2023).

Taken together, the literature presents AeroDuo in two layers of meaning. In the narrow and formal sense, it is a dual-altitude UAV collaborative VLN framework centered on Pilot-LLM, MSP, and HaL-13k (Wu et al., 21 Aug 2025). In the broader interpretive sense, it names a recurring engineering pattern: a paired architecture in which aerial, orbital, aerodynamic, or material subsystems are split into complementary roles to trade off global coverage against local precision, softness against stopping power, or ascent performance against recovery capability (Mundlamuri et al., 2023).