AirStar: Diverse UAV, RIS, and AO Systems

Updated 6 July 2026

AirStar is a polysemous term representing distinct systems across domains, including UAV-centric autonomy, 3D wireless STAR-RIS, and airborne adaptive optics guide stars.
In UAV embodied intelligence, AirStar integrates multimodal interaction, modular LLM planning, and sensor fusion to enable autonomous navigation and intelligent assistance.
In wireless and optical applications, AirStar systems optimize geometry-sensitive propagation and precise beamforming or positioning to enhance communication and adaptive optics performance.

Searching arXiv for relevant papers on “AirStar” and closely related usages to ground the article. AirStar is a designation used in several arXiv research streams rather than a single canonical artifact. In recent literature it names a UAV-centric embodied platform for instruction-driven aerial assistance, an aerial STAR-RIS architecture in three-dimensional wireless systems, and an airborne artificial guide star for adaptive optics; in a distinct spelling, the closely related name A-STAR denotes the All-Sky Transient Astrophysics Reporter mission concept (Wang et al., 6 Jul 2025, Yang et al., 9 Dec 2025, García et al., 2018, Osborne et al., 2013). This multiplicity is technically important because each usage binds the term to a different systems problem: embodied autonomy, programmable propagation, wavefront sensing, or time-domain X-ray survey operations.

1. Semantic scope in the literature

A useful first distinction is terminological. In the cited literature, “AirStar” is not a standardized cross-domain acronym. It appears as a proper name for at least three different classes of systems, while “A-STAR” appears as a mission acronym in high-energy astrophysics. This suggests that “AirStar” is polysemous in current arXiv usage rather than a single research program (Wang et al., 6 Jul 2025, Yang et al., 9 Dec 2025, García et al., 2018, Osborne et al., 2013).

Usage	Domain	Defining description
AirStar	Embodied UAV systems	A UAV-centric embodied platform that turns a quadrotor UAV into an intelligent aerial assistant
AirStar	Wireless communications	A STAR-RIS mounted on a UAV and operating in full 3D
AirStar	Adaptive optics	An airborne artificial guide star carried by a multirotor UAV
A-STAR	Astrophysics	The All-Sky Transient Astrophysics Reporter

The principal misconception to avoid is terminological unification. A sentence such as “AirStar is a UAV system” is accurate only in some papers, while in others it denotes a programmable radio surface or a wavefront-sensing reference source. Precision therefore requires domain qualification.

2. UAV-centric embodied intelligence

In "Hi AirStar, Guide Me to the Badminton Court." (Wang et al., 6 Jul 2025), AirStar is an embodied platform that turns a quadrotor UAV into an intelligent aerial assistant. Its stated motivation is to exploit high maneuverability, full 3D mobility, and access to vantage viewpoints in environments with relatively few obstacles, with target scenarios including exploration, inspection, aerial imaging, and everyday assistance. The architecture comprises three main parts: the LLM Task Planner, a Knowledge Base, and a Tool Factory containing a Navigation Library and Skill Library. Runtime is split across a smartphone app front-end, an on-board stack for lightweight perception and control, and a base station hosting heavy LLM/VLM computation via Wi‑Fi/5G.

The interaction model is explicitly multimodal. Voice commands are issued through a smartphone app, while gesture-based fine control supports camera framing and local positioning, removing the need for a handheld controller. The sensing stack includes an RGB-D camera, LiDAR, GPS, IMU, microphone or phone audio, and a camera for gesture perception. Perception modules include human detection, gesture recognition, GroundingDINO for open-vocabulary object localization, LightTrack for real-time tracking, and SigLIP similarity scoring for viewpoint selection. Qwen2.5-VL is used for visual-language grounding and short-range target prediction, while VLM inference is used for question answering.

Navigation is divided into geospatial-aware long-distance planning and object-guided short-range control. Long-distance navigation uses a discrete geospatial landmark map keyed by GPS coordinates, occupancy maps, A* waypoint planning, and trajectory generation with Ego-Planner. Short-range navigation grounds language and the current image into target 2D points using Qwen2.5-VL, then lifts them to 3D using RGB-D and camera intrinsics and extrinsics. The paper also specifies the relevant projection relations, including the $SE(3)$ transform $X_c = R X_w + t$ and the pinhole model $x = K [R \mid t] X$ .

The LLM serves as the cognitive core. It queries the Knowledge Base using the instruction and UAV perception inputs, reduces hallucinations via retrieval from landmark descriptions, historical plans or navigation records, and up-to-date internet information, performs chain-of-thought reasoning, decomposes tasks into API calls, and replans when success detectors report failure. A canonical example is “Hi AirStar, guide me to the badminton court and check if it’s available,” which is decomposed into a sequence including landmark retrieval, Navi_to, QA, and a return to the user. This modularity is formalized in the “Tool Factory,” whose tools are registered with names, parameter schemas, and usage descriptions.

AirStar’s built-in capabilities extend beyond navigation. The system supports cross-modal question answering by navigating to a question-relevant area, sampling viewing angles around a landmark’s GPS bearing, ranking views by SigLIP similarity to landmark nouns, and invoking a VLM to answer. It also supports intelligent filming and target tracking: subject initialization can be done by language grounding or clicking, LightTrack maintains temporal continuity, and on-board occlusion-aware LiDAR adjustments keep subjects centered while avoiding obstacles. The paper emphasizes extensibility, not by end-to-end retraining, but by composition of pretrained modules and newly defined tools.

A notable limitation is evaluative rather than architectural. The paper emphasizes system design and demonstration and does not report quantitative benchmarks such as SR or SPL for VLN, tracking MOTA or IDF1, or filming user studies. It also states that the system primarily leverages pretrained components and does not report custom training or end-to-end finetuning.

3. Aerial STAR-RIS in three-dimensional wireless propagation

In "Performance Comparison of Aerial RIS and STAR-RIS in 3D Wireless Environments" (Yang et al., 9 Dec 2025), AirStar denotes an aerial STAR-RIS system, specifically a simultaneously transmitting and reflecting reconfigurable intelligent surface mounted on a UAV and operating in full 3D. The comparison target is a conventional aerial RIS, which operates solely in reflection mode and serves users in its reflective half-space, whereas STAR-RIS provides full-space coverage via concurrent transmission and reflection at each element in energy-splitting mode.

The system model consists of an $M$ -antenna BS, an $N$ -element aerial RIS or STAR-RIS on a UAV at position $(x,y,H)$ , and $K$ single-antenna users on the ground. The paper’s simulation setting uses $M=8$ , $K=4$ , $N=20$ , $X_c = R X_w + t$ 0 dB, $X_c = R X_w + t$ 1 dBm, and $X_c = R X_w + t$ 2 dBm. A key modeling distinction is geometric orientation. The horizontal RIS has surface normal $X_c = R X_w + t$ 3, while the vertical STAR-RIS has normal

$X_c = R X_w + t$ 4

where the in-plane orientation angle $X_c = R X_w + t$ 5 is a controllable degree of freedom. This angle enters the directional path-loss terms through incidence and departure geometry, and the paper treats it as central to performance.

Both BS–surface and surface–UE channels are modeled as Rician fading with directional gains. The normalized radiation intensities follow $X_c = R X_w + t$ 6 for $X_c = R X_w + t$ 7, with directivities $X_c = R X_w + t$ 8 and parameter values $X_c = R X_w + t$ 9, $x = K [R \mid t] X$ 0, and $x = K [R \mid t] X$ 1. The resulting directional path-loss terms reduce to expressions involving distance and $x = K [R \mid t] X$ 2 factors; for STAR-RIS, the magnitude uses $x = K [R \mid t] X$ 3 and the sign determines whether a user lies in the reflection side or the transmission side.

The optimization objective is sum-rate maximization. For AirStar, the problem is

$x = K [R \mid t] X$ 4

subject to the BS power constraint, the per-element amplitude-splitting constraint

$x = K [R \mid t] X$ 5

and the phase-coupling constraint

$x = K [R \mid t] X$ 6

For the aerial RIS baseline, the surface coefficients are unit-modulus reflection phases. The solution methodology is a WMMSE reformulation combined with block coordinate descent and penalty dual decomposition. The paper states that the resulting algorithm exhibits monotonic improvement and converges rapidly to a stable sum-rate for all tested orientations.

The main comparative result is geometric. At low altitude, such as $x = K [R \mid t] X$ 7– $x = K [R \mid t] X$ 8 m, AirStar significantly outperforms aerial RIS because full-space coverage allows simultaneous service to users in both transmission and reflection half-spaces, while horizontal RIS suffers from large incidence and reflection angles that reduce cascaded channel strength. At higher altitude, such as $x = K [R \mid t] X$ 9– $M$ 0 m, and near the BS, aerial RIS can surpass STAR-RIS because horizontal deployment yields more favorable angular alignment. Orientation is decisive for AirStar: for a representative placement at $M$ 1 m and $M$ 2 m, the paper reports the best sum-rate at $M$ 3, while $M$ 4 degrades performance substantially.

A plausible implication is that, in this usage, “AirStar” is best understood less as a general airborne networking platform than as a geometry-sensitive propagation-control surface. The dominant design variables are altitude, horizontal placement, and orientation, not autonomy or embodied task execution.

4. Active STAR-RIS-aided IoT NOMA

In "Aerial Active STAR-RIS-Aided IoT NOMA Networks" (Zhao et al., 5 Jan 2025), AirStar again denotes a UAV-mounted STAR-RIS, but now with active amplification and explicit coupling to IoT NOMA. The system contains one BS, one UAV carrying the active STAR-RIS, and $M$ 5 IoT devices partitioned into reflection and transmission regions. The active STAR-RIS consists of $M$ 6 tunable elements, and each element contains a reflection amplifier, a power divider, and two phase shifters. This distinguishes the framework from the passive aerial STAR-RIS comparison above.

The signal model makes the active structure explicit. The amplification gain matrix is

$M$ 7

the reflection and transmission amplitude matrices implement energy splitting, and the phase-shift matrices are diagonal with unit-modulus entries. The active elements introduce noise $M$ 8. The UAV trajectory is discretized into $M$ 9 equal-length time slots of duration $N$ 0, with fixed altitude $N$ 1 and motion constraints based on the maximum per-slot displacement $N$ 2.

The optimization problem maximizes the total sum-rate over UAV trajectory, active STAR-RIS beamforming, and BS power allocation under several coupled constraints: per-element and total active STAR-RIS power constraints, STAR-RIS feasibility constraints such as $N$ 3 and $N$ 4, NOMA/SIC ordering constraints, the BS power budget, and UAV kinematic constraints. To simplify the tri-linear coupling among amplification, splitting, and phase, the paper defines combined variables $N$ 5 and $N$ 6, then lifts them to positive semidefinite matrices $N$ 7. The resulting rank-one constraints are enforced by a penalty-based approach using

$N$ 8

The proposed solver is an alternating-optimization procedure with three subproblems: active STAR-RIS beamforming, UAV trajectory, and BS power allocation. The beamforming block uses a penalty-based method with successive convex approximation to handle rank-one structure; the trajectory and power blocks are likewise treated by successive convex optimization. To avoid exhaustive search over $N$ 9 SIC orders, the paper proposes a low-complexity decoding-order rule based on the initial UAV trajectory, assigning higher decoding priority to devices closer to the UAV.

The reported simulation setup uses $(x,y,H)$ 0 devices in an $(x,y,H)$ 1 m area, BS position $(x,y,H)$ 2 m, initial UAV position $(x,y,H)$ 3 m, final UAV position $(x,y,H)$ 4 m, $(x,y,H)$ 5 m, $(x,y,H)$ 6 m/s, $(x,y,H)$ 7 s, $(x,y,H)$ 8 dBm, active STAR-RIS total power $(x,y,H)$ 9 dBm, $K$ 0 dBm, $K$ 1 dBm, $K$ 2 dB, $K$ 3, and $K$ 4 dB. The baselines are STAR-NOMA, RIS-NOMA, ASTAR-OMA, STAR-OMA, and ASTAR-random phase.

The key result is that ASTAR-NOMA consistently outperforms all baselines in sum-rate. The AO algorithm converges within a few iterations, with the paper giving about $K$ 5 iterations when $K$ 6. Gains increase with the number of elements, BS power, flight time, and number of devices. The mechanism is described as “high-quality channel construction and power compensation”: the active STAR-RIS shapes both transmission and reflection paths while the UAV trajectory reduces distances and improves LoS, but this is traded against active noise and amplifier power constraints.

This communication-theoretic AirStar differs fundamentally from the embodied-assistant AirStar. Here the UAV is principally a mobile carrier for an active surface, and the central abstractions are lifted beamforming matrices, SIC ordering, and AO/SCA optimization, not navigation semantics or multimodal human interaction.

5. Airborne artificial guide stars for adaptive optics

In adaptive optics, AirStar denotes an airborne artificial guide star carried by a UAV. Two arXiv papers develop this idea with different emphases. "Implementation of MUAV as reference source for GLAO systems" (García et al., 2018) studies a multirotor UAV as a Ground-Layer Adaptive Optics reference source, while "Artificial guide stars for adaptive optics using unmanned aerial vehicles" (Basden et al., 2018) studies a rotary UAV beacon for astronomical and solar adaptive optics, with the UAV providing precise relative position estimates to recover true atmospheric tip–tilt.

The GLAO formulation begins from site turbulence statistics. For San Pedro Mártir, the cited LOLAS-2 ground-layer values are $K$ 7 arcsec and $K$ 8 arcsec, $K$ 9 cm and $M=8$ 0 cm, $M=8$ 1 arcmin and $M=8$ 2 arcmin, and $M=8$ 3 ms and $M=8$ 4 ms, at $M=8$ 5 nm and $M=8$ 6 nm respectively. Using Tyler’s cone-effect relation,

$M=8$ 7

the paper derives a design choice of $M=8$ 8 m for a $M=8$ 9 m telescope operating in the NIR. At that altitude, the sidereal emulation speed near zenith is $K=4$ 0 cm/s, and the positional stability requirement implied by a $K=4$ 1 arcsec WFS field of view is $K=4$ 2 cm over intervals shorter than $K=4$ 3 ms. A standard $K=4$ 4 cd LED at $K=4$ 5 m gives apparent magnitude $K=4$ 6, and the source linear size should be $K=4$ 7– $K=4$ 8 mm, based on

$K=4$ 9

The second line of work addresses the longstanding AO dependence on natural guide stars. The UAV beacon provides a stable absolute tip–tilt reference whose relative position is known from onboard sensors and downlinked to the AO system, while laser guide stars continue to provide high-order signals. The core positional requirement is stringent: to reach $N=20$ 0 arcsec tilt precision at $N=20$ 1 km, the lateral position knowledge must satisfy $N=20$ 2 mm, using $N=20$ 3. The paper proposes high-rate inertial sensing, RTK-GNSS, Kalman filtering, and low-latency telemetry, and gives the inertial drift model

$N=20$ 4

The modeled astronomical AO system uses an $N=20$ 5 m telescope, H-band evaluation, a $N=20$ 6-layer atmosphere with $N=20$ 7 cm at $N=20$ 8 nm and $N=20$ 9 m, four sodium LGSs, $X_c = R X_w + t$ 00 subapertures, a $X_c = R X_w + t$ 01 actuator DM, and a $X_c = R X_w + t$ 02 Hz AO update rate. The reported result is a Strehl ratio improvement by a factor of at least $X_c = R X_w + t$ 03 relative to LGS-only on an $X_c = R X_w + t$ 04 m class telescope in the cases studied. Performance improves with UAV altitude and with tighter position knowledge; with $X_c = R X_w + t$ 05 mm rms UAV lateral position uncertainty, meaningful improvement appears for $X_c = R X_w + t$ 06 km, while with $X_c = R X_w + t$ 07 mm rms benefit appears already for $X_c = R X_w + t$ 08 km.

A recurring misconception is that a UAV guide star is presented as a full replacement for laser guide stars. The cited papers do not make that claim. In one case the UAV source is a GLAO reference that does not sense high-altitude turbulence; in the other, it replaces the natural guide star for absolute tip–tilt while LGSs still provide high-order measurements. The limitation is therefore fundamental: finite beacon altitude leaves unsensed high-altitude tilt and focal anisoplanatism, which can be mitigated by higher altitude or multiple UAVs, but not eliminated by terminology alone.

A distinct but related usage appears in "A-STAR: The All-Sky Transient Astrophysics Reporter" (Osborne et al., 2013). Here the name is hyphenated and denotes a proposed small mission for high-cadence X-ray transient astrophysics rather than a UAV, RIS, or adaptive-optics beacon. The mission was designed to locate X-ray counterparts to ALIGO and other gravitational-wave detector sources, study low luminosity gamma-ray bursts, and discover a wide variety of transient high-energy phenomena.

The survey concept is operationally specific. A-STAR surveys the entire available sky twice per $X_c = R X_w + t$ 09 hours from a low-Earth orbit of about $X_c = R X_w + t$ 10 km, with three $X_c = R X_w + t$ 11 s dwells per orbit and a $X_c = R X_w + t$ 12 Sun-avoidance constraint. The payload comprises Owl, a coded-mask instrument covering $X_c = R X_w + t$ 13– $X_c = R X_w + t$ 14 keV, and Lobster, a wide-field focusing soft X-ray instrument covering $X_c = R X_w + t$ 15– $X_c = R X_w + t$ 16 keV. Owl uses $X_c = R X_w + t$ 17 Schottky CdTe detectors of size $X_c = R X_w + t$ 18 mm and thickness $X_c = R X_w + t$ 19 mm, with total geometrical area $X_c = R X_w + t$ 20 cm $X_c = R X_w + t$ 21, a $X_c = R X_w + t$ 22 mm Ta mask at $X_c = R X_w + t$ 23 cm, coded field of view about $X_c = R X_w + t$ 24 or about $X_c = R X_w + t$ 25 sr, and on-axis effective area about $X_c = R X_w + t$ 26 cm $X_c = R X_w + t$ 27 at $X_c = R X_w + t$ 28 keV. Lobster consists of three $X_c = R X_w + t$ 29 modules with combined field of view about $X_c = R X_w + t$ 30 deg $X_c = R X_w + t$ 31.

The triggering and localization logic is likewise explicit. Owl performs multi-band rate triggering on timescales from $X_c = R X_w + t$ 32 ms to $X_c = R X_w + t$ 33 s, with image-domain searches on $X_c = R X_w + t$ 34– $X_c = R X_w + t$ 35 s timescales. Lobster uses a two-stage image trigger, first projecting the image into two perpendicular one-dimensional histograms with a $X_c = R X_w + t$ 36 candidate threshold in both axes, then confirming a transient at higher significance in a cross-shaped region; the paper gives a false trigger probability of about $X_c = R X_w + t$ 37. Alert dissemination to the ground is designed for about $X_c = R X_w + t$ 38 minute latency.

Reported performance figures include Owl localizations of about $X_c = R X_w + t$ 39 arcmin at faint $X_c = R X_w + t$ 40 detections and about $X_c = R X_w + t$ 41 arcmin for bright sources above $X_c = R X_w + t$ 42, and Lobster localizations of less than $X_c = R X_w + t$ 43 arcmin for $X_c = R X_w + t$ 44 of GRBs, less than $X_c = R X_w + t$ 45 arcmin for $X_c = R X_w + t$ 46, and less than $X_c = R X_w + t$ 47 arcmin for $X_c = R X_w + t$ 48. The mission is expected to trigger on about $X_c = R X_w + t$ 49 GRBs per year, with additional approximate rates of $X_c = R X_w + t$ 50– $X_c = R X_w + t$ 51 gravitational-wave electromagnetic counterparts per year, more than $X_c = R X_w + t$ 52 magnetar giant or intermediate flares per year, about $X_c = R X_w + t$ 53 SN shock breakout per year, about $X_c = R X_w + t$ 54 TDEs per year, and multiple daily classes of variable sources. The mission was proposed to ESA’s 2012 Small Mission call but was not selected.

Within the broader “AirStar” naming landscape, A-STAR is best treated as orthographically adjacent rather than conceptually continuous. Its inclusion is justified only because the name is sometimes rendered by readers as “AirStar”; technically, however, it belongs to survey astrophysics and multi-messenger alerting rather than airborne robotics or wireless control surfaces.