Multi-UAV observation improves the efficiency and coverage of three-dimensional scene acquisition, but continuous onboard Gaussian Splatting reconstruction is computationally expensive, whereas transmitting dense RGB-D streams wastes communication bandwidth. We propose RAOS-GS (Reconstruction-Aware Observation Selection for Gaussian Splatting), an air–ground collaborative framework for bandwidth-constrained multi-UAV reconstruction. Each UAV maintains a lightweight Gaussian proxy solely to estimate the reconstruction value of candidate observations. The proposed Rendering-Aware Keyframe Selection (RAKS) evaluates uncovered scene content, appearance discrepancy within reliably rendered regions, and relative viewpoint change to retain complementary, non-redundant RGB-D keyframes. The ground station then constructs per-UAV Gaussian submaps, aligns their coordinate frames through cross-UAV RGB-D registration, and jointly refines camera poses and Gaussian parameters to produce a global representation. Experiments on three Replica scenes, two OpenFly-UE virtual outdoor scenes, and three real-world multi-UAV scenes demonstrate a favorable trade-off between observation volume and reconstruction quality. RAOS-GS retains only 5.1%–7.9% of the input observations on Replica and the real-world sequences and 18.7%–24.0% on OpenFly-UE. It consistently outperforms 10% uniform temporal sampling and remains competitive with or surpasses 30% sampling despite using fewer observations. Ablation studies confirm that rendering-aware selection identifies reconstruction-relevant observations, viewpoint constraints suppress redundancy, and ground-side pose–Gaussian refinement improves final reconstruction quality. These results show that reconstruction-aware observation selection can reduce air–ground data transmission without requiring high-fidelity onboard Gaussian mapping.
Reliable drone-to-drone perception is essential for safe autonomous operation in increasingly complex low-altitude airspace. This study proposes STDG-UAV, a lightweight spatial–temporal domain-generalized network for onboard drone-to-drone detection and collision-risk perception. The framework combines spatial tiny-target feature preservation with adaptive multi-scale fusion to enhance weak structural cues of distant UAVs. A motion-guided temporal memory module further exploits adjacent-frame information to improve detection continuity under fast motion, blur, and weak visibility. To enhance robustness to unseen environments, image-level domain randomization, feature-style perturbation, and consistency learning are introduced under a source-only training setting. Experiments show that STDG-UAV achieves 95.4% mAP@0.5, 69.8% mAP@0.5:0.95, and 64.7% APS, with only 5.8 M parameters and 14.9 GFLOPs. In continuous UAV videos, the model reaches a detection continuity rate of 96.2% and reduces the maximum consecutive missed frames to two. On unseen real-flight domains, it achieves an average mAP@0.5 of 87.6% with a generalization gap of 7.8 percentage points. Onboard experiments further demonstrate real-time inference at 28.7 FPS and accurate TTC estimation with an MAE of 0.42 s. These results demonstrate the potential of STDG-UAV for robust, real-time drone-to-drone perception and collision-risk assessment.
Coordinating heterogeneous unmanned aerial vehicles (UAVs) requires allocating complementary capabilities to temporally dependent tasks. We study a dynamic allocation model in which the dependency structure is known in advance and all other task information is revealed on arrival. The proposed framework combines a graph-transformer policy trained with proximal policy optimization, action masking, and ordered commitment to support structurally admissible allocation decisions. On a 229-scene benchmark, the method achieves higher mean recognized and Oracle-relative completion than the compared implementations, with lower completion variability across three seeds than MAPPO, and remains competitive from 12 to 40 UAVs. Within a separate ablation configuration, removing masking or ordered commitment produces the largest completion decreases. These results compare the reported configurations combining learned allocation with local structural and capacity checks.
This paper presents a comparative framework for robust trajectory tracking of a Skywalker X8 flying-wing unmanned aerial vehicle (UAV) using optimized smooth sliding mode control (SSMC), cascaded proportional–integral–derivative (PID) control, and model predictive control (MPC). The controllers are evaluated on a nonlinear six-degree-of-freedom UAV model incorporating a high-fidelity aerodynamic representation derived from wind-tunnel measurements and flight-test-informed refinements. The SSMC employs a cascaded outer–inner-loop architecture with smooth nonlinear switching functions, while PID and SSMC parameters are optimized using derivative-free pattern search, and the constrained linear MPC benchmark is tuned using Bayesian optimization. Performance is assessed under nominal, stochastic-wind, and sensor-noise conditions at multiple climb rates, with additional Monte Carlo and empirical disturbance-scaling analyses. MPC provides the highest tracking precision and the lowest actuator effort under nominal and wind-disturbed conditions, but its tracking error and control effort increase most strongly under sensor noise. SSMC consistently improves upon optimized PID, particularly in tracking quality, transient behavior, and saturation, and yields the lowest tracking error and control effort under sensor noise. Empirical disturbance-scaling tests further show that PID and SSMC tolerate substantially larger wind amplification than MPC within the simulated horizon. Overall, SSMC provides a favorable compromise between tracking performance, actuator demand, and disturbance robustness.
Designing complex UAV swarm performances like drone light shows and aerial displays demands both geometric precision and expressive flexibility, yet existing approaches rely on manually crafted trajectories, requiring substantial manual effort and limiting creative flexibility. While recent large language models (LLMs) offer an opportunity to bridge natural language intent and trajectory generation, general-purpose LLMs lack the geometric reasoning needed to produce spatially consistent paths. We propose an interactive end-to-end framework that translates natural language and speech instructions into executable multi-agent flight trajectories, in which language-specified 2D patterns are mapped to 3D flight coordinates at assigned altitudes. Our approach introduces SVG path commands as a vectorized representation that is both LLM-friendly and geometrically precise, serving as a semantic–geometric bridge. To equip LLMs with geometric awareness, we construct a difficulty-tiered dataset and perform supervised fine-tuning (SFT) with LoRA, complemented by greedy nearest-neighbor target assignment and BallTree-accelerated artificial potential field (APF) deconfliction for scalable execution. Experiments across ten frontier LLMs reveal a pronounced capability cliff: the best Tier 1 accuracy ranges from 88% to 92% across the two judges, whereas Tier 2 accuracies range from 0% to 26% across models and judges. After domain enhancement, our SFT-enhanced 4B-parameter model with few-shot prompting achieves 80% on basic shapes and 22% on text outlines, giving 51% overall accuracy versus 43% for the 9B model under the tested configurations. Different LoRA ranks and training token cutoffs confound this scale comparison. The Llama comparison also shows that fine-tuning gains do not generalize across the tested model families and settings. The 2 m minimum-separation threshold is satisfied in 95–100% of randomized-seed runs for swarms up to N=50 under the tested grid-launch protocol and in all tested phased-deployment runs for the N=100 dense text formation; these are empirical results rather than formal safety guarantees. The structural validity of the recommended fine-tuned configuration reaches 100% on both tiers (versus 0–34% for the two Qwen pretrained baseline configurations), and AirSim physics-based simulation validation reports a mean formation error below 1 m for five of six tested patterns, averaged from the minimum-error frame to the end of each run, with one transient separation violation observed in a single pattern. For small and moderate swarms, the offline design cycle takes approximately 30–74 s and is dominated by one LLM inference call; the densest N=200 case additionally requires up to 48.7 s for APF convergence.
