Session 3B – UAV Navigation
Date: Wednesday, 18th November 2026
Time: 16:00 – 17:30
Room: Aula Capella
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
16:00 – Sim-to-Real Reinforcement Learning for Indoor UAV Navigation and Obstacle Avoidance Using Lightweight Sensing
Authors: João J. Martins¹,², Ricardo Morais¹, André Dias¹ and José Almeida¹
Affiliations: ¹ INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² FEUP – Faculty of Engineering of the University of Porto, Portugal.
Abstract
This paper presents a sim-to-real reinforcement learning (RL) framework for autonomous navigation of small unmanned aerial vehicles (UAVs) operating in cluttered indoor intralogistics environments under strict sensing and computational constraints. The approach targets lightweight aerial platforms and adopts a minimalist perception configuration consisting of a single forward-facing time-of-flight (ToF) depth camera and a 2D LiDAR. This sensing setup provides complementary short- and mid-range obstacle perception while remaining compatible with embedded low-power onboard computing.
A navigation policy is learned using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm and trained in a ROS-Gazebo simulation environment through a velocity and heading control interface. The agent observes goal-relative features, body-frame velocity estimates, and local obstacle information derived from fused ToF and 2D LiDAR measurements. The reward formulation promotes safe, smooth, and goal-directed navigation in cluttered indoor spaces.
The trained policy is evaluated on a training map and an unseen map with extended goal distances and directly deployed on a real intralogistics UAV without modifying the perception or control pipeline. Experimental results demonstrate effective sim-to-real transfer and show that lightweight sensing combined with RL enables effective indoor UAV navigation on resource-constrained platforms.
16:15 – Temporal Range Augmentation for Dynamic Intralogistics UAV Navigation with Reinforcement Learning
Authors: João J. Martins¹,², André Dias¹ and José Almeida¹
Affiliations: ¹ INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² FEUP – Faculty of Engineering of the University of Porto, Portugal.
Abstract
Indoor UAV operation in intralogistics environments such as warehouses requires reliable navigation in narrow aisles and intersections while safely interacting with moving obstacles under strict onboard compute and sensing constraints. This work investigates reinforcement learning for indoor UAV navigation and dynamic obstacle avoidance using lightweight ToF and 2D LiDAR fusion and a PX4-compatible velocity control interface. We formulate the task as a partially observable Markov decision process and train a Twin Delayed Deep Deterministic Policy Gradient policy that outputs linear velocity and yaw-rate commands. To improve safety in dynamic scenes without introducing explicit tracking or prediction, we propose a minimal temporal augmentation of the observation by appending the previous distance sector to the current range measurements, enabling the policy to infer short-horizon motion cues from successive scans. The reward balances goal progress, directional motion, heading alignment, obstacle clearance, and smooth control. We evaluate the approach in PX4-in-the-loop simulation across dynamic intralogistics scenarios with randomized obstacle behaviors, including moving geometric obstacles with 3D motion and person-shaped agents following trajectories, under sensing noise and varied interaction patterns. Across two dynamic intralogistics environments and 100 episodes per environment, the learned policy achieves success rates of 86-89% at a contact-level collision threshold (0.36 m), with mean minimum clearances of 0.47-0.58 m. Removing the temporal augmentation reduces success from 89% to 53% with moving obstacles and from 86% to 21% with person-shaped agents, showing that the previous range vector is critical for dynamic obstacle avoidance. The residual 9-11% collision rate indicates that a certified safety layer remains necessary for operation near personnel. An ablation with a reactive TTC-based safety filter shows that, for the evaluated filter, velocity-level intervention significantly increases out-of-bounds terminations without significantly reducing collisions, motivating minimally invasive safety layers.
16:30 – TIM-MARS: Selected-Target Memory for Safer UAV Person-Following Perception
Authors: Francisco Tavares, Rodrigo Ventura and Meysam Basiri
Affiliation: Institute for Systems and Robotics, Instituto Superior Técnico, Lisbon, Portugal.
Abstract
Reliable UAV person following requires preserving the identity of the operator-selected person, not only detecting and tracking people in each frame. A tracker may continue to output plausible person tracks while the selected identity is fragmented, lost, or assigned to another person. This paper presents TIM-MARS, a modular selected-target memory layer placed after detection, multi-object tracking, and raw target selection. TIM-MARS uses temporal memory, geometric consistency, tracker identity, conservative state transitions, hard-negative memory, and MARS appearance evidence to decide whether the selected target should be output, suppressed, or reacquired. The method is evaluated with selected-target correctness metrics that separate correct, wrong, and lost output duration. Across challenging outdoor ambiguity sequences, TIM-MARS reduces wrong-target output from 7.68 s to 0.10 s in a two-person crossing sequence and from 12.60 s to 0.00 s in a dense-group ambiguity sequence. In a distractor re-entry sequence, TIM-MARS exposes a conservative trade-off: wrong-target duration remains low at 0.60 s, but lost-target duration increases. These results show that TIM-MARS is useful as a safety-oriented selected-target filter for UAV person following, while ambiguous re-entry still requires stronger reacquisition confidence.
16:45 – A fleet management and global planning tool for heterogeneous unmanned operated flights
Authors: Alejandra González Valido¹, Jorge Rodríguez Meijide¹, Juan Manuel Marredo Píriz², Alexandre Rodríguez Rendo¹ and Antidio Viguria Jiménez²
Affiliations: ¹ CATEC, Santiago de Compostela, Spain; ² CATEC, Sevilla, Spain.
Abstract
The concept presented in this paper consists in the first version of a tool which addresses the fleet management problem faced by Unmanned Aerial Systems (UAS) operators who must schedule multiple flights to accomplish a variety of tasks. Its primary objective is to facilitate access to unmanned airspace by simplifying flight and mission planning. Referred to throughout this paper as the global planner -or simply planner-, the tool integrates airspace information and potential conflicts with other operations as input data. By combining the capabilities of various established tools in the field, it leverages their respective strengths to provide a useful, simple and practical solution for fleet and task management.
17:00 – Unity-Based Simulation Framework for UAV Mission Execution
Authors: Francisco Javier Roman-Escorza, Alvaro Caballero and Anibal Ollero
Affiliation: GRVC Robotics Lab, University of Seville, Seville, Spain.
Abstract
Inspection and Maintenance (I&M) of critical infrastructures has increasingly involved the use of aerial robotics in recent years since multi-rotor UAVs (Unmanned Aerial Vehicles) are a suitable solution that enables the automation of such missions while reducing costs and risks. In this context, simulation frameworks have become essential tools as they help to prevent dangerous behaviours before real missions. This paper presents a framework powered by Unity to simulate UAV missions in scenarios under realistic conditions. The simulation framework includes the implementation of the multi-rotor UAV mechanical model and its controller, sensor models, and the integration with external modules such as a GCS (Ground Control Station) and a vision-based module for autonomous landing. The resulting dynamic library is capable of simulating the behaviour of one or several UAVs simultaneously under different wind conditions and can be operated like real UAVs through the adopted GCS. In this way, the library is beneficial for mission planning and testing before real executions. Simulation results demonstrate the benefits of the proposed approach.
17:15 – Staravia: Traversability from RGB Drone Imagery: Depth and Obstacle Prediction Using Multimodal Models
Authors: Cesar Ortego¹, Javier Bermejo² and Rafael Barea¹
Affiliations: ¹ Polytechnic University School, Nebrija University, Madrid, Spain; ² Faculty of Engineering and Technology, La Rioja University (UNIR), Madrid, Spain.
Abstract
Staravia (*A-VIA) is a method presented for estimating the traversability of ground vehicles in unstructured environments using exclusively aerial optical imagery acquired by drones, without relying on expensive sensors such as LiDAR or sensitive systems such as stereo cameras.
The approach is composed of three independent modules. The first predicts scene geometry via a neural network and subsequently derives terrain properties such as slope and curvature. The second performs terrain type classification through supersegmentation, enabling the determination of the full terrain composition. The third module detects obstacles using a generic object prediction model filtered through a multimodal embedding. The final per-pixel traversability probability is obtained by statistically combining the outputs of the three modules under an independence assumption. Based on the resulting traversability map, the safest route is computed using a weighted A* algorithm.
The method is evaluated on orthophotos of La Marañosa, a national park in Madrid, at a spatial resolution of 0.5m and 2 m per pixel and compared against state-of-the-art approaches. The proposed model demonstrates a substantial improvement across evaluation metrics relative to the compared methods. Furthermore, it provides interpretability by separating geometric and semantic computations.
This work represents a significant step towards autonomous navigation in unknown and unstructured environments, with potential applications in interplanetary exploration, operations in conflict zones, disaster response and search-and-rescue among others.