AirVLA: Aerial Vision-Language-Action Systems

• AirVLA is a research umbrella for aerial systems that fuse vision, language, and action to perform UAV navigation, tracking, and manipulation in 3D environments.
• Its architectures range from end-to-end control to decoupled safety layers and continuous control chunks, addressing the unique dynamics and constraints of flight.
• Benchmarks and datasets like AIR-VLA and AerialVLA quantify performance in precision, task completion, and cooperative behaviors, highlighting both progress and current limitations.

Searching arXiv for AirVLA-related papers and variants to ground the article. AirVLA denotes a class of aerial vision-language-action systems that bind visual perception, language grounding, and action generation for UAVs operating in 3D environments. In the current literature, the label is not fully standardized: it is used as a generic category for aerial VLA agents, as the name of a specific physics-guided transfer system, and in close association with distinct artifacts such as the AIR-VLA benchmark and the AerialVLA navigation model. Across these usages, the common objective is closed-loop mapping from natural-language intent and aerial observations to embodied flight, manipulation, tracking, or handover behavior (Zeng et al., 29 May 2026, Sun et al., 29 Jan 2026, Xu et al., 15 Mar 2026, Tucker et al., 26 Mar 2026).

1. Terminology and scope

The term “AirVLA” is used in several non-identical ways across papers. Some works use it as a broad designation for aerial VLA systems, whereas others assign it to a particular method or benchmark. This terminological heterogeneity is itself a salient feature of the field.

Usage in the literature	Referent	Representative source
AirVLA as a class	Aerial vision-language-action systems for UAVs	(Zeng et al., 29 May 2026)
AIR-VLA	Benchmark, dataset, and testbed for aerial manipulation	(Sun et al., 29 Jan 2026)
AerialVLA	Minimalist end-to-end UAV navigation model; official name	(Xu et al., 15 Mar 2026)
AirVLA	Physics-guided transfer system for aerial manipulation and navigation	(Tucker et al., 26 Mar 2026)
AirVLA realized as DroneVLA	Fetch-and-handover aerial manipulation system	(Mehboob et al., 20 Jan 2026)
UAV-VLA	Distinct system, not an alias of AirVLA	(Sautenkov et al., 9 Jan 2025)

Several clarifications follow directly from the source texts. “AerialVLA” is the official name of the model introduced in “AerialVLA: A Vision-Language-Action Model for UAV Navigation via Minimalist End-to-End Control,” and “there is no distinct technical variant called AirVLA” in that work (Xu et al., 15 Mar 2026). By contrast, “AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation” defines the first VLA benchmark specifically tailored for aerial manipulation rather than a single deployed policy (Sun et al., 29 Jan 2026). “UAV-VLA: Vision-Language-Action System for Large Scale Aerial Mission Generation” explicitly states that UAV-VLA is a distinct system and “is not presented as an alias of, nor equivalent to, ‘AirVLA’” (Sautenkov et al., 9 Jan 2025).

This suggests that AirVLA is best understood as an umbrella research area with several concrete instantiations, rather than as a single canonical architecture.

2. Core architectural patterns

Despite the terminological variation, the systems grouped under AirVLA share a consistent problem formulation: a UAV receives language-conditioned intent, fuses it with visual observations and state, and produces either direct low-level actions or intermediate navigation/manipulation intents. The main design divide is between end-to-end policies and decoupled architectures.

“AerialVLA” adopts a minimalist end-to-end design. It consumes a vertically mosaicked front-plus-down composite image, a fuzzy directional hint derived solely from onboard sensors, and a target description, then autoregressively emits three numerical action tokens and an intrinsic landing signal. Its control space is a continuous 3-DoF action vector with de-quantized outputs

$$\Delta x \in [0,5]\ \text{m},\quad \Delta z \in [-5,5]\ \text{m},\quad \Delta \psi \in [-\pi,\pi],$$

together with a special LAND token (Xu et al., 15 Mar 2026). In contrast, “DroneVLA” constrains the VLA action to binary gripper decisions,

$a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$

while delegating navigation, 3D localization, planning, and servoing to deterministic modules such as Grounding DINO, RealSense depth mapping, human-aware A*, and MediaPipe-based handover control (Mehboob et al., 20 Jan 2026).

A second pattern is chunked continuous control. “UAV-Track VLA” outputs 25-step 4-DoF displacement chunks aligned with the target relative pose

$P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$

using a $\pi_{0.5}$-style flow-matching action expert augmented by temporal compression and a spatial-aware auxiliary grounding head (Zhang et al., 2 Apr 2026). “RaceVLA” similarly regresses a continuous 4D control vector

$a_t=[V_x,V_y,V_z,\omega]$

directly from FPV frames and language, emphasizing high-frequency reactive control for racing scenarios (Serpiva et al., 4 Mar 2025).

A third pattern inserts an explicit safety or feasibility layer between semantic reasoning and motor execution. “VLA-AN” predicts 3D waypoint intents and desired yaw from multimodal inputs, then applies a lightweight real-time action module with geometric safety correction rather than a stochastic generative action expert (Wu et al., 17 Dec 2025). “AIR-VLA+” formalizes the same separation more aggressively for aerial manipulation by splitting the action field into UAV movement and arm manipulation components, with cascaded dual-action decoders and an asymmetric feature-level Mixture of Experts so that movement can observe manipulator intent without destabilizing it through backpropagation (Sun et al., 11 Jun 2026).

These design choices reflect the core aerial difficulty: language grounding transfers relatively well, but flight dynamics, safety envelopes, payload changes, and camera-induced viewpoint shifts make action generation substantially more brittle than in fixed-base or 2D mobile settings.

3. Navigation, tracking, and aerial VLN

Navigation-oriented AirVLA systems span pure end-to-end control, explicit waypoint prediction, and hybrid language-grounded visual navigation. On the TravelUAV benchmark, AerialVLA reports seen-environment performance of NE $65.88$ m, SR $47.96\%$, OSR $57.69\%$, and SPL $38.54$, and on unseen maps reports NE $67.42$ m, SR $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$0, OSR $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$1, and SPL $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$2. Relative to LongFly on unseen maps, its SR increases from $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$3 to $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$4, approximately $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$5, while removing dependencies on dense oracle guidance and external object detectors (Xu et al., 15 Mar 2026).

For onboard deployment, “VLA-AN” targets resource-constrained aerial platforms. Its pipeline combines a ViT-based visual encoder, projector, LLM reasoning core, temporal comparator, and a safety-aware action module. The paper reports a high-fidelity 3D-GS dataset with 100K+ trajectories and 1M+ multimodal samples, a progressive three-stage training curriculum, up to $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$6 throughput improvement on Jetson Orin NX, robust $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$7–$a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$8 Hz closed-loop inference onboard, and a maximum single-task success rate of $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$9 (Wu et al., 17 Dec 2025).

Tracking systems expose a related but distinct AirVLA regime: persistent language-conditioned target maintenance in dynamic urban scenes. “UAV-Track VLA” constructs a CARLA-based benchmark with $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$0 frames, $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$1 tasks, and $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$2 diverse objects, then introduces temporal compression and a dual-branch decoder to decouple spatial grounding from action generation. In long-distance pedestrian tracking on seen maps, it reports ATF $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$3 and SR $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$4, while reducing single-step inference latency by $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$5 to $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$6 s relative to the original $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$7 (Zhang et al., 2 Apr 2026).

High-speed FPV control appears in “RaceVLA,” which adapts OpenVLA-7B to racing-drone flight. It reports motion generalization $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$8 versus $P_t=[\Delta x,\Delta y,\Delta z,\Delta \psi]^\top,$9 for OpenVLA and semantic generalization $\pi_{0.5}$0 versus $\pi_{0.5}$1, with average velocity $\pi_{0.5}$2 m/s and maximum speed $\pi_{0.5}$3 m/s in multi-lap tests (Serpiva et al., 4 Mar 2025). The architectural lesson is that aerial VLAs often benefit from control spaces expressed directly in velocity and yaw-rate primitives, rather than in manipulation-oriented delta-action parameterizations.

Aerial VLN papers adjacent to AirVLA make the same issue explicit from a navigation-centric perspective. “Aerial Vision-and-Language Navigation with Grid-based View Selection and Map Construction” reformulates aerial VLN as grid-based view selection with per-view vertical offsets, achieving on AerialVLN-S val-seen NE $\pi_{0.5}$4 m and SR $\pi_{0.5}$5, compared with $\pi_{0.5}$6 SR for LAG, while reporting vertical prediction accuracy of $\pi_{0.5}$7 exact/relaxed on selected views (Zhao et al., 14 Mar 2025). “ViSA-Enhanced Aerial VLN” instead keeps reasoning on the image plane through Set-of-Mark prompting and three-stage verification, surpassing the fully trained FlightGPT baseline on CityNav Test-Unseen with NE $\pi_{0.5}$8, SR $\pi_{0.5}$9, OSR $a_t=[V_x,V_y,V_z,\omega]$0, and SPL $a_t=[V_x,V_y,V_z,\omega]$1, corresponding to a $a_t=[V_x,V_y,V_z,\omega]$2 improvement in SR and a $a_t=[V_x,V_y,V_z,\omega]$3 improvement in SPL over FlightGPT (Tong et al., 9 Mar 2026). These systems are not always branded as AirVLA, but they occupy the same design space of language-grounded aerial action under 3D mobility constraints.

4. Aerial manipulation, handover, and physics-aware control

Aerial manipulation is the domain in which AirVLA diverges most sharply from ground VLA assumptions. “AIR-VLA” formalizes this gap as a benchmark built in NVIDIA Isaac Sim for a quadrotor equipped with a 7-DoF Franka Panda arm and a parallel gripper. The benchmark provides $a_t=[V_x,V_y,V_z,\omega]$4 manually teleoperated demonstrations, four task suites—Base Manipulation, Object and Spatial Understanding, Semantic Understanding, and Long-Horizon Planning—an average episode length of approximately $a_t=[V_x,V_y,V_z,\omega]$5 steps at $a_t=[V_x,V_y,V_z,\omega]$6 Hz, and long-horizon tasks exceeding $a_t=[V_x,V_y,V_z,\omega]$7 steps (Sun et al., 29 Jan 2026).

Its evaluation protocol makes the aerial-specific failure modes explicit through the weighted score

$a_t=[V_x,V_y,V_z,\omega]$8

where positioning accuracy, manipulator efficacy, environmental disturbance safety, and task progression are scored separately (Sun et al., 29 Jan 2026). Baseline results show that large pretrained VLAs transfer better than conventional policies: $a_t=[V_x,V_y,V_z,\omega]$9 achieves the highest overall weighted total, approximately $65.88$0, whereas $65.88$1-FAST, ACT, and Diffusion Policy are markedly weaker (Sun et al., 29 Jan 2026).

“AIR-VLA+” is a direct architectural response to these coupled aerial-manipulation dynamics. On the standardized AIR-VLA benchmark it reports an overall average score of $65.88$2, with $65.88$3, $65.88$4, $65.88$5, and $65.88$6. Relative to the single-head $65.88$7 policy, the overall task completion score improves by $65.88$8, and the Base Manipulation and Long-Horizon suites show reported success-rate leaps of $65.88$9 and $47.96\%$0, respectively (Sun et al., 11 Jun 2026). The central mechanism is decoupling: movement and manipulation are generated by separate cascaded decoders, movement observes detached manipulator intent, and a feature-level MoE lets different movement experts specialize in approach, grasp-hover, and move-to-target stages (Sun et al., 11 Jun 2026).

A second manipulation line studies transfer from existing foundation VLAs rather than designing aerial-specific policies from scratch. In “$47.96\%$1, But Make It Fly,” AirVLA fine-tunes the manipulation-pretrained $47.96\%$2 model to aerial pick-and-place and navigation, then modifies the sampler at inference time through Payload-Aware Guidance. The guided velocity field is

$47.96\%$3

with the correction term derived from the gradient of a payload-aware objective over the denoised action chunk (Tucker et al., 26 Mar 2026). Across a cumulative $47.96\%$4 real-world experiments, synthetic 3DGS navigation data raises hover success to $47.96\%$5 where teleoperation-only data with RTC attains $47.96\%$6; Payload-Aware Guidance increases real-world place success from $47.96\%$7 to $47.96\%$8; and the long-horizon compositional task reaches $47.96\%$9 place success (Tucker et al., 26 Mar 2026).

“DroneVLA” extends AirVLA to human-facing fetch-and-handover. It integrates a lightweight VLA policy inspired by TinyVLA, open-vocabulary grounding, RGB-D localization, human-aware A* planning, and hybrid PBVS/IBVS servoing. The system completes a three-leg mission—home→object, object→human, human→home—and reports localization and navigation errors of $57.69\%$0 m max, $57.69\%$1 m mean Euclidean, and $57.69\%$2 m RMSE in a $57.69\%$3 m $57.69\%$4 $57.69\%$5 m indoor lab, with all $57.69\%$6 end-to-end experiments reported successful (Mehboob et al., 20 Jan 2026). In this configuration, AirVLA is not a monolithic policy but an integrated embodied system in which language reasoning, grounding, planning, and human-centric control are tightly composed.

5. Benchmarks, datasets, and metrics

A distinctive feature of AirVLA research is the proliferation of task-specific evaluation ecosystems rather than a single benchmark. AIR-VLA is the most explicit benchmarking effort for aerial manipulation: besides the VLA score above, it defines VLM-planning metrics for process planning, spatial navigation understanding, object grounding, and skill selection, using

$57.69\%$7

with equal weights $57.69\%$8 across components (Sun et al., 29 Jan 2026).

Navigation work uses more conventional embodied-navigation metrics but extends them to 3D mobility. The grid-based Aerial VLN benchmark reports NE, SR, OSR, SDTW, and nDTW, plus exact and relaxed vertical-decision accuracy, with success defined as stopping within $57.69\%$9 m of the ground-truth destination (Zhao et al., 14 Mar 2025). ViSA-enhanced CityNav evaluation similarly relies on NE, SR, OSR, and SPL, again with a $38.54$0 m success threshold (Tong et al., 9 Mar 2026). TravelUAV uses NE, SR, OSR, and SPL for direct end-to-end UAV control (Xu et al., 15 Mar 2026).

Tracking work introduces closed-loop persistence measures rather than endpoint metrics. UAV-Track VLA defines success rate

$38.54$1

and average tracked frames

$38.54$2

with fatal failure triggered when the target remains out of valid distance range or FOV for more than $38.54$3 consecutive frames in the reported experiments (Zhang et al., 2 Apr 2026).

Cooperation studies push the evaluation problem further. “Can Aerial VLA Models Cooperate?” introduces CARLA-Air, a single-process CARLA–AirSim environment for air–ground interaction, and defines Tracking Success Rate, Landing Success Rate, Cooperative Conversion Rate

$38.54$4

Cooperation Gain, Recovery Success Rate, Re-acquisition Time, Decision Frequency, Effective Coordination Latency, and simulation-timestamp alignment error

$38.54$5

CARLA-Air reports $38.54$6 ms by construction, compared with mean $38.54$7 ms and max $38.54$8 ms in a bridged two-process runtime, together with a $38.54$9 reduction in cooperation-metric standard deviation from $67.42$0 to $67.42$1 (Zeng et al., 29 May 2026).

The evaluation picture that emerges is highly fragmented but technically revealing. AirVLA is assessed not only by task completion, but by hover stability, manipulation precision, payload robustness, partner-state coordination, and temporal consistency under long horizons.

6. Limitations, open problems, and research direction

The most persistent limitation is the dynamics gap between pretrained VLA priors and aerial embodiment. “$67.42$2, But Make It Fly” states the issue directly: visual representations transfer effectively, but “the specific control dynamics required for flight do not” (Tucker et al., 26 Mar 2026). AIR-VLA benchmark results corroborate this diagnosis: models are notably better at coarse UAV mobility than precise arm control, and performance degrades under dense distractors, long-horizon subtask chaining, and removal of fixed third-person views (Sun et al., 29 Jan 2026).

Perception remains a second bottleneck. DroneVLA identifies sensitivity to lighting, clutter, and depth noise, as well as language ambiguities and the limitations of a 1-DoF gripper (Mehboob et al., 20 Jan 2026). ViSA notes dependence on landmark priors, limited active 3D perception, and deployment latency induced by an online VLM API (Tong et al., 9 Mar 2026). The grid-based aerial VLN work reports a generalization gap between seen and unseen scenes and notes that a 2D BEV map may be insufficient for highly cluttered 3D environments (Zhao et al., 14 Mar 2025).

Onboard execution remains uneven across the literature. VLA-AN demonstrates $67.42$3–$67.42$4 Hz onboard closed-loop inference on Jetson Orin NX and explicitly targets further increases toward approximately $67.42$5 Hz for highly dynamic scenes (Wu et al., 17 Dec 2025). UAV-Track VLA reduces latency to $67.42$6 s per step but is still validated in CARLA rather than in real flight (Zhang et al., 2 Apr 2026). RaceVLA runs at approximately $67.42$7 Hz with server-side RTX 4090 inference, which the paper itself identifies as a constraint for faster maneuvers (Serpiva et al., 4 Mar 2025).

A further misconception exposed by recent work is that single-UAV competence automatically yields cooperative competence. CARLA-Air shows the opposite. In cooperative moving-platform landing, AerialVLA attains TSR $67.42$8 but only LSR $67.42$9 under independent execution, and naive bidirectional interaction reduces LSR to $a_t=\pi(I_t,L)\in \{\text{Open},\text{Close}\},$00. Across the tested text-based cue interfaces, the study concludes that zero-shot cooperative AirVLA requires explicit partner-state grounding, low-latency action coordination, and team-level objective alignment beyond the current paradigm (Zeng et al., 29 May 2026).

The field is therefore converging on a hybrid view. End-to-end aerial VLA is demonstrably viable; pretrained multimodal representations are useful; and language-grounded aerial behavior now spans navigation, tracking, racing, fetch-and-handover, and aerial manipulation. At the same time, the strongest current results rely on aerial-specific interventions: dual-view perception, temporal compression, view-selection grids, image-plane verification, physics-guided sampling, movement/manipulation decoupling, human-aware planning, and explicit safety correction. A plausible implication is that “AirVLA” will continue to denote not one architecture but a family of aerially adapted VLA systems whose common core is multimodal grounding, and whose principal innovations lie in reconciling that grounding with the dynamics, safety constraints, and long-horizon structure of flight.