UAVM 2025 Challenge: Benchmark Ambiguity & Trends

• UAVM 2025 Challenge is an umbrella term that ambiguously represents a spectrum of UAV benchmarks and multimodal research directions.
• It encompasses diverse tasks such as aerial-ground person re-identification, multimodal fusion in audio-visual contexts, and collaborative 3D perception.
• The challenge ecosystem emphasizes practical evaluation metrics, enhanced safety protocols, and systems-level integration for future UAV applications.

UAVM 2025 Challenge is not a standardized title in the currently cited arXiv literature. No paper in the cited set explicitly defines a competition under exactly that name; instead, the phrase aligns with a broader 2024–2026 UAV benchmark ecosystem and with a separate acronymic usage in multimodal learning. The AG-VPReID 2025 challenge paper states that it does not mention “UAVM 2025 Challenge,” and the UAV-VLN survey likewise states that it does not introduce a single new challenge, dataset, or leaderboard under that name. At the same time, “UAVM” already denotes the “Unified Audio-Visual Model” in audio-visual representation learning, which is unrelated to a specific UAV benchmark (Nguyen et al., 28 Jun 2025, Chen et al., 15 Apr 2026, Gong et al., 2022).

1. Terminology, ambiguity, and scope

The available literature supports two distinct readings of the expression. One reading treats it as a search label for the 2025 UAV challenge landscape: aerial-ground re-identification, multi-modal anti-UAV tracking, collaborative 3D perception, wireless localization, maritime computer vision, and vision-language navigation. The other reading is acronymic and technical: “UAVM” as “Unified Audio-Visual Model,” a shared-parameter audio-visual architecture rather than a UAV competition (Nguyen et al., 28 Jun 2025, Chen et al., 15 Apr 2026, Gong et al., 2022).

Reference point	Explicit statement in the literature	Relevance to the term
AG-VPReID 2025	The paper does not mention “UAVM 2025 Challenge”	Closest explicit 2025 UAV challenge paper
UAV-VLN survey	The paper does not introduce a challenge under that name	Indicates no canonical benchmark title
“UAVM” (audio-visual)	“UAVM” means “Unified Audio-Visual Model”	Separate acronymic usage

This ambiguity matters because concrete claims can be made only about explicitly documented challenges and benchmark families. A plausible implication is that “UAVM 2025 Challenge” functions less as a single formally specified competition than as a convenient umbrella for several 2025-adjacent UAV benchmarking traditions.

2. The nearest explicit 2025 challenge: AG-VPReID 2025

The clearest 2025 challenge paper in the neighboring literature is “AG-VPReID 2025: Aerial-Ground Video-based Person Re-identification Challenge Results,” described as the first large-scale video-based competition focused on high-altitude aerial-ground person re-identification. It builds on AG-ReID 2023, moving from image-based aerial-ground ReID at 20–60 m to video-based matching at 80–120 m (Nguyen et al., 28 Jun 2025).

The challenge uses the AG-VPReID dataset with 3,027 identities, 13,511 tracklets, and approximately 3.7 million frames captured from DJI UAVs, Bosch CCTV cameras, and GoPro Hero10 wearable cameras. UAV imagery is recorded at 80 m and 120 m; the dataset is explicitly multi-platform and is labeled as A/G/W for Aerial / Ground / Wearable. The task is defined in two directions: Aerial-to-Ground (A2G), where an aerial video tracklet is used as query against ground tracklets, and Ground-to-Aerial (G2A), where a ground tracklet is queried against aerial tracklets. The G2A protocol includes 1,693 distractor identities in the gallery, which makes retrieval more realistic and more difficult.

The benchmark is built around tracklets rather than still images and emphasizes six key difficulties: extreme viewpoints, low resolution, occlusions, illumination changes, clothing similarities, and pose variations. The paper also stresses temporal discontinuities, motion inconsistencies, and strong cross-platform scale disparity, with aerial body sizes ranging from 18×37 pixels to 293×542 pixels.

The evaluation metrics are Rank-1 (R1), Rank-5 (R5), Rank-10 (R10), and mean Average Precision (mAP). The challenge was hosted on Kaggle, with a maximum of ten submissions per day per team and live leaderboard updates. Four participating teams were reported: UAM with X-TFCLIP, DUT / IIAU LAB with TFCLIP-AG-VPReID, IT/UBI with ACP-VPReID, and MFHO, plus the TF-CLIP baseline.

Method	Overall R1	Overall mAP
X-TFCLIP	71.56	73.60
TFCLIP-AG-VPReID	71.89	72.70
ACP-VPReID	67.72	69.34
MFHO	64.93	67.32
TF-CLIP	63.75	66.26

The paper repeatedly identifies X-TFCLIP as the winning method. At the same time, TFCLIP-AG-VPReID has slightly higher overall Rank-1 (71.89 vs 71.56), while X-TFCLIP has the best overall mAP (73.60 vs 72.70). This suggests that the official ranking likely prioritized mAP or another overall criterion, but the paper does not explicitly state the final ranking rule. The paper also contains a direct inconsistency on external data usage: the main submission section says that all teams used exclusively the official training set without external datasets, whereas the appendix states that TFCLIP-AG-VPReID trained with AG-VPReID 2025 + G2AVReID.

In the narrower sense of a documented 2025 UAV-centered challenge paper, AG-VPReID 2025 is therefore the strongest explicit anchor. In the strict documentary sense, however, it cannot be identified as “the UAVM 2025 Challenge,” because the paper itself rejects that identification.

3. Neighboring benchmark traditions in the 2024–2025 UAV challenge ecosystem

Several benchmark papers document the technical environment in which a phrase such as “UAVM 2025 Challenge” would be interpreted. Together they define a landscape of multi-modal sensing, collaborative perception, real-world field validation, and domain-specific UAV evaluation (Ye et al., 2024, Kudyba et al., 2024, Kiefer et al., 2023, Deng et al., 2024).

Benchmark	Core tasks	Distinctive scope
UAV3D	Single-UAV and collaborative-UAV 3D detection/tracking	1,000 scenes, 5 UAVs, 5 cameras per UAV
AERPAW “Find a Rover”	UAV-assisted wireless localization	3-minute and 10-minute estimates in a twenty-acre area
MaCVi 2024	UAV maritime MOT with ReID; USV segmentation/detection/tracking	over 195 submissions
UG2+/MMUAD	UAV detection, classification, and 3D tracking	Stereo fisheye, lidar, radar; four UAV classes

“UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles” provides a benchmark designed for both 3D and collaborative 3D perception tasks with UAVs. It contains 1,000 scenes, each with 20 frames, for 20,000 total frames / samples; each sample involves 5 UAVs per frame and 5 RGB cameras per UAV, yielding 25 images per sample, 500,000 images total, and 3.3 million 3D bounding boxes. The four tasks are single-UAV 3D object detection, single-UAV object tracking, collaborative-UAV 3D object detection, and collaborative-UAV object tracking. The benchmark is simulated through CARLA-AirSim co-simulation, uses a fixed cross-shaped formation at 60 m altitude, and evaluates with mAP, NDS, mATE, mASE, mAOE for detection and AMOTA, AMOTP, MOTA, MOTP, TID, LGD for tracking. It is therefore a strong reference for any challenge interpretation centered on collaborative 3D UAV perception.

“A UAV-assisted Wireless Localization Challenge on AERPAW” documents the inaugural AERPAW Challenge, “Find a Rover.” Teams had to localize a narrowband radio signal and estimate the hidden rover’s location after 3 minutes and again after 10 minutes, within a twenty-acre area. The flight region was geofenced, the rover region partially inaccessible, the drone altitude was constrained between 20 m and 110 m, and speed was constrained to under 10 m/s. The top teams—NYU, UNT, and UGA—used markedly different strategies, including Gaussian process regression + Bayesian optimization and recursive perimeter sweep. A key result was methodological rather than purely algorithmic: NYU dominated the emulator, while UNT achieved the best real-world performance. The paper attributes much of this difference to digital-twin mismatch, since the channel emulator used free-space propagation model with 10 dB added white noise, whereas the real testbed experienced 30–40 dB of noise at times.

“The 2nd Workshop on Maritime Computer Vision (MaCVi) 2024” shows a different benchmark logic: one UAV task plus several USV tasks under realistic maritime conditions. Its five tracks are UAV-based Maritime Object Tracking with Re-identification, USV-based Obstacle Segmentation, USV-based Embedded Obstacle Segmentation, USV-based Obstacle Detection, and USV-based Maritime Boat Tracking. The workshop reports over 195 submissions, public datasets, evaluation code, and a public leaderboard. This provides a mature example of challenge governance, hidden-test evaluation, and embedded-efficiency constraints, all of which are directly relevant to later UAV benchmark design.

“Multi-Modal UAV Detection, Classification and Tracking Algorithm — Technical Report for CVPR 2024 UG2 Challenge” describes the first-place system for UG2+, a task in the CVPR 2024 UAV Tracking and Pose-Estimation Challenge. The problem includes drone detection, UAV-type classification, and 2D/3D trajectory estimation under extreme weather conditions with stereo vision, various Lidars, Radars, and, in the abstract, audio arrays. The implemented winning system, however, is explicitly decomposed: camera ROIs drive four-class UAV type classification—Phantom 4, M300, M30T, Mavic 3—while lidar/radar point clouds drive 3D pose tracking. The paper is especially informative because it argues for late fusion under unsynchronized measurements and because it reports strong improvements in center-regression MSE from 0.27 to 0.05.

Taken together, these benchmarks suggest that the practical meaning of a 2025 UAV challenge is plural rather than singular. Some benchmark families center on perception and tracking, some on wireless localization, some on maritime deployment, and some on multi-modal anti-UAV sensing.

4. System-level challenge dimensions beyond pure perception

A broader systems interpretation of a UAVM-like challenge emerges from papers that define autonomy as a stack rather than a single leaderboard score. “Autonomous Advanced Aerial Mobility — An End-to-end Autonomy Framework for UAVs and Beyond” organizes aerial autonomy into sensing, perception, planning, and controls, and situates those blocks within simulation, cloud/onboard compute, V2X, fleet operations, and certification. The paper distinguishes automation from autonomy, arguing that autonomy implies satisfactory performance under uncertainty and the ability to compensate for failures without external intervention. It also places strong emphasis on application-conditioned planning for package delivery, inspection, survey and mapping, and ISR, and on validation regimes influenced by DO-178C, EASA work on learning assurance, and the W-shaped development process for learning-enabled systems (Mishra et al., 2023).

A communications-centered extension is provided by “Machine Learning-Assisted UAV Operations with UTM: Requirements, Challenges, and Solutions,” which identifies four UTM pillars—operation planning, situational awareness, status and advisors, and security—for low-altitude, often BVLOS, operations below 400 feet. The paper treats trajectory prediction, connectivity prediction, obstacle awareness, failure prediction, recovery, and remote identification as distinct machine-learning problems. “Waveform and Spectrum Management for Unmanned Aerial Systems Beyond 2025” adds a separate but complementary communications view: future dense civilian UAS operation must handle both low-throughput, ultra-reliable, low-latency control signaling and higher-rate payload traffic, and the paper argues that dynamic spectrum access will be needed because static allocation will be insufficient (Abdalla et al., 2020, Kakar et al., 2017).

Other system papers translate these concerns into challenge-worthy operational tasks. “Challenges in Close-Proximity Safe and Seamless Operation of Manned and Unmanned Aircraft in Shared Airspace” proposes an integrated autonomy stack for terminal-area mixed traffic with vision-only aircraft detection, intent understanding, natural-language interaction, GAIL + MCTS + STL planning, and runtime safety filtering. “Validation of Collision Detection and Avoidance Methods for Urban Air Mobility through Simulation” adds a benchmark-like detect-and-avoid framework using decision trees and three safety envelopes—collision, warning, and caution—evaluated with Closest Point of Approach and mission delay, including the explicit delay decomposition $D_T(s)=D_G+D_A$. “Practical Challenges in Landing a UAV on a Dynamic Target” decomposes precision landing into approach, tracking, and descent, and contrasts GPS, RF, optical flow, marker-based landing, IR, IMU fusion, and control families such as PID, fuzzy PID, and MPC (Patrikar et al., 2022, Panchal et al., 2023, Salagame et al., 2022).

This systems literature suggests that any comprehensive 2025 UAV challenge would need to measure more than detection or classification. It would need to incorporate at least some combination of communication reliability, airspace integration, contingency handling, disturbance rejection, mission legality, and safety assurance.

5. Multimodal and embodied-autonomy interpretations of “UAVM”

If the term is interpreted through multimodal embodied intelligence rather than through a specific named competition, three papers become especially important. First, “UAVM: Towards Unifying Audio and Visual Models” defines UAVM as a Unified Audio-Visual Model with modality-specific front ends followed by a shared modality-agnostic Transformer and shared classifier. Its base setting is $N=N_s=3$ with shared dimension $S_{dim}=1024$; it achieves 65.8% top-1 accuracy on VGGSound fusion and uses 76% of the parameters of the modality-independent counterpart. This usage is technically precise but is a separate acronymic lineage from UAV benchmark naming (Gong et al., 2022).

Second, “UAV-VLPA*: A Vision-Language-Path-Action System for Optimal Route Generation on a Large Scales” shows what a multimodal 2025 UAV task can look like when natural language, satellite imagery, global route ordering, and local obstacle avoidance are combined. The system interprets commands such as visiting all buildings while avoiding lakes, extracts waypoints from overhead imagery, solves global ordering with 2-opt for TSP, refines route segments with A* on a 5 × 5 pixel occupancy grid, and reports a final trajectory length of 51.27 km versus an updated human route of 62.95 km, corresponding to the paper’s headline 18.5% reduction. Runtime is reported as no more than 3 minutes on an RTX 4090 plus Intel Core i9-13900K (Sautenkov et al., 4 Mar 2025).

Third, “Vision-and-Language Navigation for UAVs: Progress, Challenges, and a Research Roadmap” provides a formal challenge grammar for UAV-VLN. It defines UAV-VLN as a POMDP with state, action, transition, reward, observation, and observation function, and it emphasizes continuous 3D control, long-horizon reasoning, sim-to-real gaps, and safety. The survey identifies benchmark families such as AerialVLN, AVDN, CityNav, OpenFly, UAV-Flow, and SAGE-Bench, and reports two quantitative gaps that have become reference points: AerialVLN baselines at 5.1% SR versus 80.8% for human pilots, and an OpenFly seen/unseen drop from 33.2% SR to 10.7% SR. The paper also argues that UAV evaluation should combine holistic metrics such as SR, OSR, NE, SPL, nDTW, and SDTW with robustness metrics such as Collision Rate, Sim-to-Real SR Drop, and Inference Frequency (Chen et al., 15 Apr 2026).

This suggests a second, multimodal reading of a “UAVM 2025 Challenge”: not a single benchmark already fixed in the literature, but a family of tasks centered on language grounding, multi-sensor perception, long-horizon navigation, and constrained onboard deployment.

6. Unresolved issues, controversies, and prospective meaning

The first unresolved issue is simply nominal. The cited literature does not define a single canonical “UAVM 2025 Challenge.” The most explicit 2025 challenge paper nearby, AG-VPReID 2025, states that it does not identify itself that way; the UAV-VLN survey likewise denies the existence of a challenge under that title (Nguyen et al., 28 Jun 2025, Chen et al., 15 Apr 2026).

The second unresolved issue is challenge governance and reproducibility. AG-VPReID 2025 contains an unresolved contradiction on external data usage and does not state the final ranking rule, even though it repeatedly identifies X-TFCLIP as winner. AERPAW demonstrates a different difficulty: digital twin fidelity. The channel emulator used 10 dB white noise, but real flights saw 30–40 dB fluctuations. UAV3D is large and carefully structured, but it is explicitly simulated, not real, and its experiments collapse 17 vehicle categories into the single benchmark class “car” (Nguyen et al., 28 Jun 2025, Kudyba et al., 2024, Ye et al., 2024).

The third unresolved issue is synchronization and modality realism. The UG2+/MMUAD winning system is instructive precisely because it rejects end-to-end early fusion under unsynchronized measurements and instead relies on modality-specialized branches with late temporal fusion. This suggests that many future UAV challenge datasets will continue to reward robust engineering of cross-modal asynchrony rather than idealized fusion architectures (Deng et al., 2024).

The fourth unresolved issue is certification-grade safety. The end-to-end AAM perspective, the UTM paper, the shared-airspace teaming paper, the dynamic-target landing review, and the UAM collision-avoidance study all indicate that performance metrics alone are insufficient. Certification, operational envelopes, fault handling, safety cases for learned components, and compliance with shared-airspace procedures remain open research and engineering problems (Mishra et al., 2023, Abdalla et al., 2020, Patrikar et al., 2022, Salagame et al., 2022, Panchal et al., 2023).

In the available literature, the most defensible interpretation is therefore composite. “UAVM 2025 Challenge” denotes no single arXiv-standardized competition title; rather, it points to a 2025-era UAV benchmark constellation that includes high-altitude aerial-ground ReID, collaborative 3D perception, wireless localization, maritime UAV evaluation, multi-modal anti-UAV tracking, and emerging multimodal navigation and planning. The term’s ambiguity is itself revealing: by 2025, UAV challenge design had already expanded from narrow perception tasks toward multi-sensor, multi-agent, safety-aware, and mission-level evaluation.