KIO-planner: Attention-Guided Single-Stage Motion Planning with Dual Mapping for UAV Navigation

1 st Dexing Yao†, 1 st Haochen Li†, 2 nd Junhao Wei, 3 rd Yifu Zhao, 4 th Yanxiao Li, 5 th Jiahui Xu, 6 th Jinxuan Hu, 7 th Lele Tian, 8 th Baili Lu, 9 th Zikun Li, 10 th Xu Yang, 11 th Sio-Kei Im, 12 th Dingcheng Yang, 13 th Yapeng Wang* · 2026

Read KIO-planner: Attention-Guided Single-Stage Motion Planning with Dual Mapping for UAV Navigation by 1 st Dexing Yao†, 1 st Haochen Li†, 2 nd Junhao Wei, 3 rd Yifu Zhao, 4 th Yanxiao Li, 5 th Jiahui Xu, 6 th Jinxuan Hu, 7 th Lele Tian, 8 th Baili Lu, 9 th Zikun Li, 10 th Xu Yang, 11 th Sio-Kei Im, 12 th Dingcheng Yang, 13 th Yapeng Wang* (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-19703.txt.

Introduction

Yao and team build KIO-planner, an attention-guided planner for drones in tight spaces. Standard planners lag and stall among dense obstacles, while learning methods miss depth detail and lack motion limits. KIO-planner adds attention to its backbone for edges and open space, plus Dual Mapping with a Safety Shield in depth-pixel space for collision-free flight. Simulated tests show flight at up to 3 point 0 metres per second with 24 milliseconds latency. Paths cut cost by 28 point 4 percent, and safety margins rise from 0 point 48 to 0 point 76 metres.

Catalog record

CREATOR1 st Dexing Yao†, 1 st Haochen Li†, 2 nd Junhao Wei, 3 rd Yifu Zhao, 4 th Yanxiao Li, 5 th Jiahui Xu, 6 th Jinxuan Hu, 7 th Lele Tian, 8 th Baili Lu, 9 th Zikun Li, 10 th Xu Yang, 11 th Sio-Kei Im, 12 th Dingcheng Yang, 13 th Yapeng Wang*
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25629
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.19703

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

KIO-planner: Attention-Guided Single-Stage Motion Planning with Dual Mapping for UAV Navigation Thanks: This work has been accepted for publication in an IEEE Vehicular Technology Conference (VTC). © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. 1 st Dexing Yao†, 1 st Haochen Li†, 2 nd Junhao Wei, 3 rd Yifu Zhao, 4 th Yanxiao Li, 5 th Jiahui Xu, 6 th Jinxuan Hu, 7 th Lele Tian, 8 th Baili Lu, 9 th Zikun Li, 10 th Xu Yang, 11 th Sio-Kei Im, 12 th Dingcheng Yang, 13 th Yapeng Wang* Abstract Autonomous UAV flight in confined, wall-dense environments requires low-latency and reliable motion planning under strict safety constraints. Traditional optimization-based planners suffer from mapping latency and easily fall into local minima when navigating through dense structural obstacles. Meanwhile, existing end-to-end learning methods struggle to extract fine-grained geometric features from raw depth images and lack hard kinodynamic constraints, leading to unpredictable collisions near walls. To address these issues, we propose KIO-planner, an attention-guided single-stage trajectory planning framework. First, we integrate a Convolutional Block Attention Module (CBAM) into the perception backbone to adaptively focus on critical structural edges and traversable space. Second, we introduce a novel Dual Mapping mechanism—comprising physical bounds activation and a deterministic Geometric Safety Shield in the depth-pixel space—to enforce kinodynamic feasibility and collision-free flight without global map fusion. Extensive high-fidelity simulated experiments demonstrate that KIO-planner enables highly agile navigation at speeds up to 3.0 m/s. Compared to the state-of-the-art baseline, KIO-planner achieves lower inference latency ( ∼ \sim 24 ms) and generates significantly smoother trajectories (reducing control cost by 28.4%). Most notably, our Dual Mapping substantially increases the worst-case safety margin (minimum distance to obstacles) from 0.48 m to 0.76 m, ensuring fast, smooth, and safer navigation in highly constrained environments. *Corresponding author I INTRODUCTION In recent years, the deployment of Unmanned Aerial Vehicles (UAVs) in confined and structured environments, such as indoor navigation, tight corridors, and industrial facility inspections, has significantly increased. High-speed autonomous navigation in these wall-dense scenarios requires extreme agility, low latency, and reliable safety. Traditional optimization-based planners, such as Fast-Planner [1] and EGO-planner [2], have achieved strong performance by utilizing Euclidean Signed Distance Fields (ESDFs) or collision-free guiding paths to optimize trajectories. Furthermore, advanced trajectory optimization frameworks like MINCO [3] leverage smooth maps to handle geometric constraints efficiently. However, in highly constrained environments with dense uncooperative walls, these decoupled architectures still suffer from significant computational mapping latency and are highly prone to getting trapped in local minima during gradient optimization. To bypass the latency of explicit mapping and complex optimization, learning-based end-to-end navigation approaches have emerged as a promising alternative. Recent pioneering works have demonstrated the promising potential of convolutional neural networks (CNNs) in mapping sensory observations directly to collision-free trajectories, enabling high-speed flight in the wild [4] and agile dynamic obstacle avoidance [5]. Despite their fast inference speeds, these ’black-box’ models suffer from a major limitation: the lack of explicit kinodynamic and geometric hard constraints. When operating in extremely close proximity to walls, even a minor probabilistic prediction error can lead to collision risks. To overcome these challenges, we propose KIO-planner, an attention-guided single-stage trajectory generation framework equipped with improved safety guarantees. To significantly enhance the network’s spatial perception, we seamlessly integrate the Convolutional Block Attention Module (CBAM) [ 6 ] into our feature extraction backbone. By adaptively recalibrating both channel and spatial feature responses, KIO-planner accurately focuses on critical navigable gaps while suppressing irrelevant background noise. More importantly, to bridge the gap between neural network unpredictability and hard safety constraints, we introduce a novel Dual Mapping mechanism. Inspired by the concepts of lazy search over local 3D data [ 7 ] and safe learning via shielding [ 8 ] , our Geometric Safety Shield deterministically evaluates and crops unsafe trajectory candidates directly in the depth-pixel space, eliminating the need for global map fusion. Finally, to ensure kinodynamic feasibility, the framework generates closed-form polynomial motion primitives [ 9 ] that are strictly mapped to the UAV’s physical actuation limits. The main contributions of this paper are as follows: • We propose KIO-planner, an efficient single-stage planning framework that achieves millisecond-level inference for high-speed UAV navigation in confined wall environments. • We introduce a CBAM-enhanced perception backbone that significantly improves the network’s ability to extract fine-grained structural edges and narrow gaps. • We design a Dual Mapping mechanism, comprising physical bounds activation and a Geometric Safety Shield, to provide a deterministic safety margin and guarantee collision avoidance. • Extensive high-fidelity simulated experiments demonstrate that KIO-planner achieves superior agility, smoothness, and safety compared to state-of-the-art baselines. Fig. 1: The proposed KIO-planner framework: The depth image and odometry are processed by a CBAM-enhanced backbone to predict raw primitives, which are then strictly constrained by the Dual Mapping mechanism to ensure kinodynamic and geometric safety. II RELATED WORK II-A UAV Local Motion Planning Traditional UAV local planning typically follows a hierarchical ’perception-mapping-planning’ pipeline. Gradient-based trajectory optimizers, such as Fast-Planner [ 1 ] and EGO-planner [ 2 ] , have demonstrated impressive agility by utilizing Euclidean Signed Distance Fields (ESDFs) or collision-free guiding paths to push trajectories away from obstacles. To handle complex geometric and kinodynamic constraints, advanced frameworks like MINCO [ 3 ] parameterize trajectories with spatial-temporal deformations. However, in highly confined environments with dense, non-convex walls, these optimization-based methods often suffer from severe mapping latency and easily get trapped in local minima, leading to oscillatory behaviors. To mitigate explicit mapping overhead, learning-based end-to-end approaches have gained significant traction. Kaufmann et al. [ 4 ] demonstrated that mapping raw sensory observations directly to actions can enable high-speed flight in the wild. Similarly, Deep-PANTHER [ 5 ] leverages imitation learning for rapid perception-aware replanning. Despite the ultra-low latency of end-to-end neural planners, these ’black-box’ models generally lack explicit kinodynamic boundaries. Consequently, their probabilistic predictions often lead to collision risks when navigating near walls where the margin for error is close to zero. II-B Attention-Guided Representation Learning Extracting robust geometric features from high-dimensional depth images is crucial for reliable navigation. While standard Convolutional Neural Networks (CNNs) fuse spatial and channel-wise information within local receptive fields, their representational power can be further enhanced by attention mechanisms. SENet [ 10 ] introduced the Squeeze-and-Excitation block to adaptively recalibrate channel-wise feature responses, proving highly effective in classification tasks. However, in UAV navigation, the spatial location of structural edges and narrow gaps is just as critical as the channel features. Therefore, our KIO-planner framework incorporates the Convolutional Block Attention Module (CBAM) [ 6 ] , which sequentially infers attention maps along both channel and spatial dimensions. This dual attention allows KIO-planner to explicitly focus on traversable spaces while ignoring irrelevant background noise, a capability missing in standard learning-based planners. II-C Safety-Critical and Kinodynamic Navigation Ensuring absolute safety during high-speed flight requires bridging the gap between motion planning and control limits. Mueller et al. [ 9 ] pioneered the rapid generation of closed-form

Read the complete text