UAV-VL-R1: Lightweight Aerial Reasoning Model

• UAV-VL-R1 is a lightweight vision-language model designed for structured reasoning on UAV imagery, achieving significantly higher accuracy than general-purpose VLMs.
• It leverages a hybrid training pipeline—combining supervised fine-tuning and multi-stage GRPO—to enhance semantic alignment and cross-task transfer on eight UAV-specific reasoning tasks.
• The accompanying HRVQA-VL dataset of over 50,000 samples and low-memory inference (2.5–3.9 GB) supports practical, real-time deployment on resource-constrained UAV platforms.

Searching arXiv for the target paper and closely related UAV vision-language systems to ground the article. arXiv search query: (Guan et al., 15 Aug 2025) UAV-VL-R1 UAV-VL-R1 is a lightweight vision-LLM for reasoning over UAV aerial imagery, introduced as a domain-specialized alternative to general-purpose VLMs whose performance degrades on drone-view scenes with high resolution, complex spatial semantics, and strict real-time constraints. Built on Qwen2-VL-2B-Instruct and trained with a hybrid pipeline of supervised fine-tuning and multi-stage Group Relative Policy Optimization, it is designed for structured and interpretable inference on eight UAV-relevant reasoning tasks. Its accompanying dataset, HRVQA-VL, contains 50,019 annotated samples, and the model is reported to require 3.9 GB of memory under FP16 inference and 2.5 GB with GPTQ INT8 quantization, with the stated goal of supporting real-time deployment on resource-constrained UAV platforms (Guan et al., 15 Aug 2025).

1. Identity, scope, and problem setting

UAV-VL-R1 addresses a specific gap in multimodal aerial intelligence: strong performance on natural-image tasks does not transfer cleanly to UAV imagery. The paper attributes this degradation to the visual distribution of aerial scenes, including top-down or oblique viewpoints, severe scale variation, sparse semantic cues, cluttered layouts, occlusion, and strong spatial regularity, as well as to the fact that many UAV tasks require explicit spatial and compositional reasoning rather than object-centric captioning (Guan et al., 15 Aug 2025). The model is therefore framed not simply as an aerial VQA system, but as a compact aerial reasoner optimized for structured output, cross-task transfer, and deployability.

The work positions UAV-VL-R1 as the first lightweight VLM tailored to eight UAV aerial reasoning tasks and trained with a hybrid supervised fine-tuning plus multi-stage GRPO pipeline (Guan et al., 15 Aug 2025). The stated design goals are lightweight deployment, structured and interpretable inference, cross-task aerial reasoning, and training stability under reinforcement learning. In this formulation, the model is meant to support downstream use in navigation, monitoring, and decision support, but the paper itself evaluates image-based reasoning rather than flight control or mission execution.

The name is easily confused with other UAV vision-language systems published in the same period. UAV-VLRR is a search-and-rescue stack that combines ChatGPT-4o, a quantized Molmo model, and point-to-point NMPC for rapid response, and the paper explicitly notes that “UAV-VL-R1” does not appear there (Yaqoot et al., 4 Mar 2025). UAV-VLA and UAV-VLPA* are map-grounded mission-generation systems that translate natural-language requests and satellite imagery into waypoints or optimized routes, rather than image-question-answer reasoning over UAV scenes (Sautenkov et al., 9 Jan 2025); (Sautenkov et al., 4 Mar 2025). UAV-VL-R1 is therefore best understood as a reasoning model, not a planner, controller, or mission compiler.

2. Model architecture and deployment profile

Architecturally, UAV-VL-R1 is built on Qwen2-VL-2B-Instruct. The paper does not introduce a new backbone; instead, it describes the model as a post-trained specialization of the 2B base model for UAV-domain reasoning (Guan et al., 15 Aug 2025). The visual encoder, LLM, and multimodal alignment machinery are inherited from Qwen2-VL. During supervised fine-tuning, the backbone model is frozen, the visual encoder remains trainable, and LoRA is used for lightweight adaptation.

The SFT configuration is explicit. LoRA uses rank 32 and scaling factor $\text{lora-alpha}=48$, and the model is trained for 4 epochs with batch size 1, gradient accumulation 4, bfloat16 mixed precision, and early stopping (Guan et al., 15 Aug 2025). The implementation uses PyTorch, DeepSpeed, and FlashAttention-2, and experiments run on 2× NVIDIA A100 40GB PCIe and 2× RTX A6000 48GB GPUs.

The paper does not report a new high-resolution visual module or a modified projector. It repeatedly emphasizes high-resolution aerial imagery as a challenge, but does not provide a new explicit image-resolution handling mechanism beyond using Qwen2-VL-2B-Instruct as the multimodal base (Guan et al., 15 Aug 2025). The architectural novelty is therefore primarily in post-training and reasoning format rather than in backbone redesign.

Deployment characteristics are central to the paper’s claims. UAV-VL-R1 remains at roughly 2B parameters, while the 72B comparison model is described as $36\times$ larger (Guan et al., 15 Aug 2025). The reported inference memory is 3.9 GB under FP16 and 2.5 GB with GPTQ INT8 quantization. The paper presents these numbers as evidence that the model is suitable for onboard or edge deployment on UAV platforms, although it does not provide a detailed latency or throughput table.

3. Training methodology: SFT and multi-stage GRPO

The training pipeline has two stages: supervised fine-tuning followed by multi-stage reinforcement learning. In the SFT phase, each sample is represented as $(i,q,r,a)$, where $i$ is the input image, $q$ the question, $r$ the reasoning path, and $a$ the answer (Guan et al., 15 Aug 2025). The model is trained to maximize the conditional likelihood of the reasoning-plus-answer sequence given image and question:

$\mathcal{L}_{\text{SFT} = -\mathbb{E}_{(i,q,r,a)\sim\mathcal{D} \left[ \sum_{t=1}^{T}\log \pi_\theta(y_t \mid i,q,y_{<t}) \right]$

Operationally, SFT establishes semantic alignment and trains the output schema in which reasoning is enclosed in > ...</think> and the prediction in <answer>...</answer> (Guan et al., 15 Aug 2025).

The RL stage uses Group Relative Policy Optimization. Rather than employing a learned value function or preference model, GRPO samples a group of candidate outputs from the old policy, scores them with deterministic rule-guided rewards, and computes a relative advantage within the group (Guan et al., 15 Aug 2025). The intended objective is given in PPO-like clipped form with KL regularization:

$\mathcal{J}_{\text{GRPO}(\theta) = \mathbb{E}\Bigl[ q \sim P(\mathcal{Q}), \{o_i\}_{i=1}^{G}\sim \pi_{\theta_{\text{old}(\cdot \mid q) \Bigr] \left[ \frac{1}{G}\sum_{i=1}^{G}\frac{1}{|o_i|}\sum_{t=1}^{|o_i|} \min\left( \frac{\pi_\theta(o_i\mid q)}{\pi_{\theta_{\text{old}(o_i\mid q)}A_i,\; \operatorname{clip}\!\left( \frac{\pi_\theta(o_i\mid q)}{\pi_{\theta_{\text{old}(o_i\mid q)}, 1-\varepsilon,1+\varepsilon \right)A_i \right) \right] -\beta D_{\mathrm{KL}[\pi_\theta\|\pi_{\mathrm{ref}]$

The printed advantage formula is:

$$A_i = \frac{r_i - \max(\{r_1,r_2,\ldots,r_G\})}{\operatorname{std}(\{r_1,r_2,\ldots,r_G\})}$$

The surrounding explanation, however, describes reward standardization within the group. This discrepancy is noted in the paper’s detailed discussion and indicates a notation or editing inconsistency rather than a different optimization principle (Guan et al., 15 Aug 2025).

The reward function is deliberately simple. A format reward checks whether both <think>... and <answer>...</answer> are present; if so, the reward is $36\times$0, otherwise $36\times$1. An accuracy reward checks whether the content inside <answer>...</answer> exactly matches the ground-truth answer; if correct, the reward is $36\times$2, otherwise $36\times$3. The total reward is:

$36\times$4

with

$36\times$5

The RL curriculum is divided into three stages. RL1 uses Stage-A tasks, RL2 uses Stage-B tasks, and RL3 uses Stage-C tasks (Guan et al., 15 Aug 2025). This curriculum progresses from low-level attribute reasoning to counting and category recognition, and then to location reasoning and scene understanding. The paper argues that this staged schedule improves stability and supports transfer across task groups.

4. HRVQA-VL dataset and task organization

HRVQA-VL is the benchmark introduced for training and evaluation. It is described as a cleaned, structured subset of the prior HRVQA aerial VQA dataset and contains 50,019 carefully filtered samples (Guan et al., 15 Aug 2025). The paper motivates this refinement by noting that HRVQA, though very large at approximately 1.07M samples, is semi-artificially synthesized, contains annotation errors, and uses verbose language not ideal for high-quality VLM training.

The dataset contains eight UAV-relevant reasoning tasks: color, size, yes/no, shape, number, transportation, location, and scene (Guan et al., 15 Aug 2025). These are grouped into three curriculum stages.

Stage	Tasks
Stage A	color, size, yes/no
Stage B	number, transportation, shape
Stage C	location, scene

The split reported by the paper is 42,465 training samples and 7,554 test samples, corresponding to an 85:15 train/test partition (Guan et al., 15 Aug 2025). Within training, 19,187 samples are used for SFT and 23,278 for RL. The RL portion is further divided into 5,434 samples for Stage A, 9,257 for Stage B, and 8,587 for Stage C.

The dataset is explicitly UAV-oriented rather than satellite-oriented. The paper contrasts it with datasets such as VisDrone, HazyDet, UAV3D, and UAVid, which target detection, tracking, 3D perception, or segmentation rather than VQA, and with FloodNet and RSVQA, which do support VQA but are domain-limited or satellite-oriented (Guan et al., 15 Aug 2025). HRVQA-VL is presented as a VLM-compatible dataset organized for reasoning supervision and RL, rather than as a general perception benchmark.

A plausible implication is that HRVQA-VL fills a methodological gap between UAV perception datasets and UAV control systems. It provides a structured image-question-answer substrate for aerial reasoning, but it does not itself encode control, action, or temporal interaction. That distinction is important when comparing UAV-VL-R1 with planning-centric systems.

5. Empirical performance and ablation findings

The paper evaluates UAV-VL-R1 under cross-task generalization and full-test multi-task generalization. On the full HRVQA-VL test set, the reported “Zero-shot Plain” and “Zero-shot Prompting” results are as follows (Guan et al., 15 Aug 2025).

Model	Zero-shot Plain	Zero-shot Prompting
Qwen2-VL-2B-Instruct	36.60%	22.20%
Qwen2-VL-2B-Instruct (SFT)	52.60%	26.93%
InstructBLIP-7B	37.67%	32.91%
Qwen2-VL-7B-Instruct	43.83%	38.40%
Qwen2-VL-72B-Instruct	51.46%	46.67%
UAV-VL-R1	68.94%	72.13%

These figures support two central claims. First, UAV-VL-R1 outperforms all reported baselines, including Qwen2-VL-72B-Instruct, despite being based on a 2B model (Guan et al., 15 Aug 2025). Second, structured prompting with <think> and <answer> degrades most general-purpose baselines but improves UAV-VL-R1, suggesting that the model has internalized the required output format rather than merely tolerating it.

The cross-task generalization results are similarly strong. Qwen2-VL-72B-Instruct records Stage-A 56.82, Stage-B 51.39, Stage-C 40.05, and Overall 49.42, while UAV-VL-R1 trained on all stages records Stage-A 86.27, Stage-B 63.40, Stage-C 64.22, and Overall 71.30 (Guan et al., 15 Aug 2025). The paper also highlights the transfer behavior of the Stage-A-only model, which achieves 48.37 on unseen Stage B and 46.59 on unseen Stage C, with 60.44 overall. This is interpreted as evidence that low-level visual-language alignment learned from attribute tasks transfers unusually well across task groups.

The abstract states that UAV-VL-R1 achieves a 48.17% higher zero-shot accuracy than the Qwen2-VL-2B-Instruct baseline (Guan et al., 15 Aug 2025). The tabulated values, however, make this wording ambiguous. On the full test set under Zero-shot Prompting, the scores are 72.13% and 22.20%; on the cross-task table, the corresponding Overall figures are 71.30% and 23.96%. The number 48.17 matches the difference between 72.13 and 23.96, which come from different evaluation tables. The paper’s concrete table values are therefore clearer than the headline phrasing.

The ablation study compares GRPO-only training with SFT-initialized GRPO. GRPO-ABC achieves Stage-A 70.17, Stage-B 64.94, Stage-C 43.32, and Overall 59.48, whereas SFT-GRPO-ABC achieves 86.27, 63.40, 64.22, and 71.30 (Guan et al., 15 Aug 2025). This supports the paper’s more nuanced conclusion: SFT improves semantic alignment and especially helps Stage-C spatial and scene tasks, but may reduce reasoning diversity in Stage-B numerical or compositional tasks, where GRPO-only training is slightly stronger. GRPO is presented as compensating for this rigidity by increasing logical flexibility and inference robustness.

Training dynamics reinforce that interpretation. The format reward rises rapidly to its maximum of 0.5 within roughly 30 steps, while the accuracy reward remains low initially and then begins rising around step 200 before stabilizing around step 1400 (Guan et al., 15 Aug 2025). The paper describes this as a two-phase pattern in which structure is learned first and reasoning quality improves later.

6. Relation to adjacent UAV vision-language systems and principal limitations

Within the UAV vision-language literature, UAV-VL-R1 occupies a reasoning-centric position. UAV-VLA and UAV-VLPA* focus on converting natural-language requests and satellite imagery into path-action plans, using Molmo-based grounding and classical route generation such as 2-opt TSP and A* (Sautenkov et al., 9 Jan 2025); (Sautenkov et al., 4 Mar 2025). UAV-VLRR moves further toward execution by turning image-text mission descriptions into target and obstacle coordinates that drive a point-to-point NMPC controller with built-in obstacle avoidance (Yaqoot et al., 4 Mar 2025). VLM-RRT uses a VLM as a directional prior for RRT sampling in autonomous navigation, but keeps geometric feasibility in a classical planner (Ye et al., 29 May 2025). By contrast, UAV-VL-R1 does not claim mission generation, replanning, obstacle avoidance, or control; its contribution is aerial visual reasoning over structured tasks.

This distinction matters because it clarifies what UAV-VL-R1 is not. It is not a low-level controller, not a flight stack, not a map-grounded route optimizer, and not a perception-only detection dataset. The role it can plausibly play in a larger UAV autonomy architecture is as a reasoning module that interprets aerial scenes in a structured way and emits parseable outputs for downstream systems.

The limitations acknowledged or implied by the paper are correspondingly specific. The authors state that future work will expand dataset scope, include more task types and modalities, and refine reward design to improve RL stability (Guan et al., 15 Aug 2025). The paper does not provide measured latency, FPS, or tokens-per-second on edge hardware, so its real-time deployment claim rests primarily on memory feasibility rather than quantified onboard throughput. It also does not deeply address weather shifts, altitude shifts, temporal reasoning, video understanding, hallucination control, uncertainty estimation, or open-world robustness. Evaluation is based on exact-match extraction from <answer>...</answer>, which gives a precise benchmark but leaves broader answer-space generalization open.

A broader limitation emerges by comparison with perception-side resources such as LRDDv3, which supplies 4K RGB imagery, paired 640x512 thermal images, and range metadata for long-range drone detection but no language annotations (Peterson et al., 25 May 2026). This suggests that UAV-VL-R1 addresses the reasoning layer more directly than the sensor-grounding layer. A plausible implication is that a more complete UAV multimodal stack would require tighter coupling between aerial reasoning benchmarks such as HRVQA-VL and perception datasets designed for tiny-object detection, thermal fusion, or range-aware sensing.

In summary, UAV-VL-R1 is best characterized as a post-trained, lightweight aerial VLM specialized for structured reasoning over UAV imagery. Its significance lies less in architectural novelty than in the combination of a refined aerial reasoning dataset, explicit reasoning-format supervision, and multi-stage GRPO that enables a 2B model to outperform much larger general-purpose VLMs on the authors’ benchmark while remaining compatible with constrained deployment settings (Guan et al., 15 Aug 2025).