Learning to Fly: Stable Vision-Guided UAV Servoing with Compact Target-Centric Cues and Reinforcement Learning

Saurbh Singh Jamwal; Nived Chebrolu

cs.ROSep 7, 2026

Learning to Fly: Stable Vision-Guided UAV Servoing with Compact Target-Centric Cues and Reinforcement Learning

Authors: Saurbh Singh Jamwal, Nived Chebrolu

Organizations: Department of Computer Science and Engineering Indian Institute of Technology Bombay

Abstract

Vision-guided reinforcement learning for Unmanned Aerial Vehicles (UAVs) remains challenging due to unstable policy optimisation, aggressive exploration, and the cost of high-dimensional visual perception. In this work, we investigate long-horizon UAV visual servoing using compact target-centric cues combined with low-dimensional sensor measurements. Rather than learning directly from RGB images, lightweight target segmentation provides image-space offsets and relative depth, which are combined with quadrotor velocity and projected-gravity measurements into a compact 12D policy observation. We compare Direct PPO with three matched-budget curriculum strategies: a Visual curriculum that progressively expands target placement difficulty, a Dynamics curriculum that gradually relaxes action constraints and smoothing, and a Joint curriculum that combines both progressions. All strategies reach comparable nominal performance, with complementary advantages across tracking metrics. Observation ablations show that proprioceptive measurements are critical for stable flight and image-space cues for target alignment, while explicit depth is not necessary for strong performance in the evaluated setting. Against tuned classical visual-servo controllers, learned policies show greater robustness to strong control and visual perturbations, while the Visual curriculum exhibits the smallest degradation under unseen target motion. Overall, the results demonstrate that compact target-centric representations can support robust long-horizon aerial visual servoing and that visual curriculum training can improve robustness to dynamic distribution shifts despite limited gains in nominal performance.

Figures & tables

Figure 1: Overview of the proposed vision-guided UAV visual servoing framework. The RGB-D observation is processed through target segmentation to extract compact image-offset and depth cues, which are combined with sensor measurements into a 12D observation state. A PPO policy maps this state to collective thrust and body-moment commands, while the curriculum scheduler adjusts the IsaacLab training difficulty.

Figure 2: Equal-budget training comparison averaged across five random seeds. The panels show ( a ) episodic reward, ( b ) episode length, ( c ) successful-step rate, and ( d ) center error. Shaded regions indicate one standard deviation.

Method	Reward $\uparrow$	Ep. Length $\uparrow$	Center Error $\downarrow$	Depth Error $\downarrow$	Success $\uparrow$
Direct PPO	96.01 $\pm$ 1.89	495.9 $\pm$ 4.7	0.0233 $\pm$ 0.0077	1.3925 $\pm$ 0.6671	0.2825 $\pm$ 0.2046
Visual Curriculum PPO	95.94 $\pm$ 1.41	496.4 $\pm$ 4.9	0.0388 $\pm$ 0.0329	1.1669 $\pm$ 0.2578	0.3835 $\pm$ 0.1306
Dynamics Curriculum PPO	96.56 $\pm$ 2.07	498.2 $\pm$ 1.6	0.0257 $\pm$ 0.0113	1.4297 $\pm$ 0.2884	0.3257 $\pm$ 0.1054
Joint Curriculum PPO	96.35 $\pm$ 2.78	496.9 $\pm$ 2.8	0.0515 $\pm$ 0.0539	1.7296 $\pm$ 0.5923	0.2877 $\pm$ 0.2242

Table 1: Equal-budget comparison under the hardest static evaluation configuration, averaged across five random seeds. Bold indicates the best mean value.

Observation	Dim.	Reward $\uparrow$	Center Error $\downarrow$	Depth Error $\downarrow$	Success $\uparrow$	Survival $\uparrow$
Full	12	89.09 $\pm$ 3.22	0.0230 $\pm$ 0.0032	0.9779 $\pm$ 0.1060	0.4719 $\pm$ 0.0371	97.1%
No Depth	11	94.17 $\pm$ 3.72	0.0182 $\pm$ 0.0052	0.7986 $\pm$ 0.1100	0.5318 $\pm$ 0.0605	98.7%
No Proprioception	3	2.87 $\pm$ 1.09	0.2649 $\pm$ 0.0234	0.8022 $\pm$ 0.0846	0.3674 $\pm$ 0.0516	4.4%
State Only	9	58.01 $\pm$ 4.30	0.2649 $\pm$ 0.0389	0.8494 $\pm$ 0.3045	0.3131 $\pm$ 0.0937	100.0%

Table 2: Observation ablation under the hard evaluation setting. Results are mean

\pm

standard deviation across five seeds. The variants of PPO isolate the roles of target-relative depth, proprioception, and target-centric visual cues.

Method	Static $R$	Lateral $R$	Helix $R$	$\Delta R_{\mathrm{H}}$	Helix Success	Helix Survival
Direct PPO	89.37 $\pm$ 3.26	87.86 $\pm$ 3.71	82.60 $\pm$ 3.91	$-7.58\%$	0.4050 $\pm$ 0.0691	97.3%
Visual Curriculum PPO	87.41 $\pm$ 1.44	88.60 $\pm$ 2.12	83.97 $\pm$ 1.51	$-$ 3.94%	0.4216 $\pm$ 0.0247	97.7%
Classical Servo 2D	99.37 $\pm$ 1.28	97.29 $\pm$ 1.05	87.74 $\pm$ 0.91	$-11.71\%$	0.5248 $\pm$ 0.0182	94.6%
Classical Servo 3D	102.12 $\pm$ 1.16	99.77 $\pm$ 1.13	85.97 $\pm$ 3.35	$-15.82\%$	0.6818 $\pm$ 0.0195	84.9%

Table 3: Zero-shot target-motion generalisation.

\Delta R_{\mathrm{H}}

is the reward change from static to hard-helix evaluation. Results are mean

\pm

standard deviation over five runs.

Method	Action	Visual	Dropout	Vis. Surv.
Direct PPO	$+2.08\%$	$-27.09\%$	$+1.75\%$	$97.3\%$
Visual Curr. PPO	$+0.01\%$	$-25.43\%$	$+0.40\%$	$97.2\%$
Classical Servo 2D	$-9.08\%$	$-36.62\%$	$+0.46\%$	$86.7\%$
Classical Servo 3D	$-6.16\%$	$-39.05\%$	$-0.16\%$	$85.1\%$

Table 4: Evaluation-time robustness. Action, Visual, and Dropout denote percentage reward change relative to clean evaluation; Vis. Surv. denotes survival rate under visual noise.

Method	$\Delta R_{0.8}$	$\Delta R_{1.2}$	$\Delta S_{0.8}$	$\Delta S_{1.2}$
Direct PPO (Full)	$+0.4\%$	$-0.7\%$	$+0.018$	$-0.002$
Direct PPO (No Depth)	$-1.2\%$	$+0.3\%$	$+0.009$	$-0.005$
Classical Servo 2D	$-0.4\%$	$+0.6\%$	$+0.000$	$+0.001$
Classical Servo 3D	$-0.8\%$	$+0.5\%$	$-0.006$	$-0.001$

Table 5: Performance change under

\pm 20\%

focal-length variation relative to nominal camera intrinsics.

\Delta R

and

\Delta S

denote changes in reward and successful-step rate, respectively.

Vision-Based Robot Control

Vision-based robot control uses camera images to guide robots’ navigation and manipulation through closed-loop perception and action.

See all 221 papers

Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach
Jul 5, 2026Yaxuan Li, Jiarui Zeng, Shaofei Huang +1Diffusion-Based Planning UAV Navigation
FST and ICI, University of Macau · Macau SAR, China
CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UAV Tracking
Jul 16, 2026Ruilong Ren, Songsheng Cheng, Yunpeng Zhou +9Aerial Robotics Visual Object Tracking
1Autel Robotics · 2Northeast Normal University · 3Southern University of Science and Technology +2
RefineFly: Failure-Aware Post-Training for Aerial Vision-Language Navigation
Aug 10, 2026Boxiong Wang, Hui Kang, Geng Sun +3UAV Navigation Vision-Language Navigation
1Jilin University · 4Zhongguancun Academy · 2Tsinghua University +1

Jul 5, 2026J/K move · Enter open · S save

Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach

Yaxuan Li, Jiarui Zeng, Shaofei Huang, Zhedong Zheng

FST and ICI, University of Macau · Macau SAR, China

Diffusion-Based Planning UAV Navigation Vision-Based Robot Control

Open paper AskPDF arXiv

In this work, we study the last-meter precision navigation for UAVs, e.g., autonomously reaching a target within the final 10 meters using monocular vision. This task is challenging due to scale ambiguity, rotation discontinuities, and the need for fine-grained spatial reasoning. Existing methods often fail under large viewpoint changes or lack generalization to unseen environments. To this end, we propose DreamNav, a coarse-to-fine diffusion-refined aerial visual servoing framework. In the first coarse-estimation stage, a robust regression policy employs a trigonometric parameterization to predict rotation by jointly modeling sine and cosine components, effectively mitigating optimization instabilities caused by angular periodicity. Given this coarse estimate, the second diffusion-refined stage utilizes a pre-trained world model to simulate future visual observations for candidate actions, selecting the trajectory that minimizes visual discrepancy with the target through a process of visual imagination. To support rigorous evaluation, we contribute PairUAV, a large-scale benchmark comprising 4.8 million image pairs across 72 scenes, curated from the University-1652 dataset. Extensive experiments show DreamNav outperforms strong visual servoing and foundation model baselines in accuracy and generalization, with zero-shot transfer to unseen scenes.

Abstract

Vision-Based Robot Control

Explore similar work

Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach

CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UAV Tracking

RefineFly: Failure-Aware Post-Training for Aerial Vision-Language Navigation