Low-Cost Rapid-Development Air-Ground Robotic Solution for Nuclear Power Plant Inspection

Low-Cost Rapid-Development Air-Ground Robotic Solution for Nuclear Power Plant Inspection* *Lessons learnt from participating Enrich 2025 Robot Hackthon Changda Tian1,†, Sasanka Kuruppu Arachchige2, Haichuan Li3, Juan Jos´ e Garc´ ıa C´ ardenas4, Hamidreza Raei5, Enes Dincer6, Alperen Kenan7, Paul Bremner7, Manuel Giuliani8, Gerhard Neumann6, Arash Ajoudani5, Adriana Tapus4, Tomi Westerlund3, Joni–Kristian K¨ am¨ ar¨ ainen2, Luis Figueredo9, Simon Watson10, and Panos Trahanias1 1Institute of Computer Science, Foundation for Research and Technology - Hellas (FORTH), Heraklion, Greece 2Tampere University, Finland 3University of Turku, Finland 4ENSTA Paris, France 5Istituto Italiano di Tecnologia (IIT), Italy 6Karlsruhe Institute of Technology (KIT), Germany 7University of the West of England, UK 8Kempten University of Applied Sciences, Kempten, Germany. 9School of Computer Science, University of Nottingham, UK 10 The University of Manchester, UK †Corresponding author: [email protected] Abstract—This paper presents a practical, low-cost, and swiftly developed heterogeneous robotic system-comprising an aerial drone and a ground vehicle-designed for inspection and situational awareness in nuclear power plant (NPP) environments. Developed during the EnRicH 2025 competition in AKW Zwentendorf, our solution prioritized rapid prototyping using ROS2, off-the-shelf components, and open-source frameworks, achieving deployment within 3 weeks. We detail the design methodology: a 3D-printed airframe for the drone, an Ackermann-steering UGV, and a combined sensor suite composed of LiDAR, IMU, and radiation detection modules. Control and navigation algorithms were implemented using ROS2, leveraging SLAM and a hybrid autonomy paradigm to address GNSS denial inside containment structures. We also tackle communication challenges within shielded environments by integrating a modular mesh network and signal repeaters. Field trials in a real NPP demonstrated robust localization, reliable communication despite severe attenuation, and real-time radiation monitoring. Our contribution is a fully reproducible, low-cost air-ground inspection solution validated in a real NPP, together with a complete open-source release of mechanical designs, BOMs, firmware, and ROS2 software stacks.1 Index Terms—Robots for nuclear power plant, heterogeneous robots inspection. I. INTRODUCTION Nuclear power plants are inherently dangerous environments, where radiation, confined spaces, and complex infrastructure make both routine inspections and emergency responses risky and difficult to manage. Autonomous robots can help mitigate these risks by allowing human operators to remain at a safe distance while performing vital tasks 1All hardware designs and software have been released at our open-source project site: UAV platform: https://sasakuruppuarachchi.github.io/agipix/; UGV platform: https://github.com/RAICAM-EU-Project/Enrich UGV; Onboard control mode: https://github.com/RAICAM-EU-Project/px4 onboard control; Radiation sensor module: https://github.com/RAICAM-EU-Project/geiger monitor. Fig. 1: Our team in AKW Zwentendorf with our robots such as mapping, sensing, and manipulation [1]. Due to this reliance on remote control, operators require reliable and intuitive teleoperation systems [2]. Despite this potential, current robotic systems are often proprietary, expensive, and lack the flexibility needed to adapt efficiently to the specific conditions of each site [3]. Moreover, many advanced inspection robots – for example, legged quadrupeds such as ANYmal or collision-tolerant drones like the Flyability Elios – are proprietary and cost-prohibitive, which hinders their rapid deployment in emergency scenarios. In contrast, our approach emphasizes low-cost hardware and open-source design for quick adaptation to specific site requirements. The 5th European Robotics Hackathon – EnRicH 2025 [4] offered us an unprecedented opportunity to test real-world robotic solutions inside the decommissioned Zwentendorf Nuclear Power Plant under authentic radiological conditions. Held from June 30 to July 4, 2025, the EnRicH event brought together robotics teams and CBRNE practitioners [5] to assess the performance of unmanned aerial and ground vehicles on tasks such as 3D environment mapping, radiation detection, valve manipulation, and the search and rescue of dummy victims. Designed as a research-focused trial rather than a competitive race, EnRicH challenges participants to conduct real radiation measurements, coordinate UAV and UGV operations, and navigate the communication difficulties caused by dense concrete and steel structures. Motivated by these challenges and by the competition’s emphasis on open science and practitioner-driven scenarios, we developed a low-cost, rapid-development air–ground robotic solution for NPP inspection. Within only 3 weeks we assembled a quadrotor and an Ackermann-steering UGV from off-the-shelf parts, integrated LiDAR, IMU, camera, and Geiger-counter sensors, and implemented control and navigation algorithms based on PX4 [6] entirely in ROS2 [7]. To maintain reliable connectivity within the heavily shielded interiors, we leveraged the plant’s existing Wi-Fi network with multiple repeater nodes. In this paper we describe our system design, implementation, and deployment in the competition, report on our results during EnRicH 2025 trials, and reflect on lessons learned for rapid prototyping of field-ready robots in challenging environments. Our experience demonstrates that resource-constrained teams can deliver operationally viable solutions aligned with end-user needs in nuclear inspection and emergency response. All of our mechanical designs, electronic schematics, and software algorithms are already open-sourced to facilitate secondary development and rapid adaptation by the research community and industry partners. II. SYSTEM ARCHITECTURE A. Hardware Our heterogeneous platform consists of an Ackermannsteering UGV and a quadrotor UAV, both designed for rapid deployment, low cost, and full open-source reproducibility. 1) Unified highand low-level control framework: To accelerate development, both robots share the same highand low-level control stack. The high-level platform is based on an NVIDIA Orin NX 8GB module mounted on a ConnectTech Boson carrier board running Ubuntu 22.04 and ROS2 Humble. Perception and autonomy, including SLAM (via DLIO [8] and LIO-SAM [9]), path planning, and radiation mapping, are implemented as ROS2 nodes fusing LiDAR, IMU, and camera data. After sensor processing, optimized motion commands are issued to the PX4-based low-level motion controller through Fast DDS. Both vehicles share a unified power architecture driven by a 14.8 V LiPo battery, providing regulated 12 V rails for computation and sensors. 2) UGV framework: The UGV employs a modified 1:10scale Reely Stagger RC chassis, providing a robust yet lightweight base at low cost. With four-wheel drive and 58 mm of clearance, it can traverse uneven floors up to 45 km/h. Propulsion and steering are managed by a PX4 Mini autopilot interfacing with the stock ESC and servo, while a 2.4 GHz (a) UGV chassis (b) High-level board UGV (c) UAV framework (d) Radiation sensor Fig. 2: Hardware overview: (a) UGV chassis. (b) high-level computation board. (c) UAV framework. (d) Radiation sensor component radio link offers manual override for safety. Power is supplied by a compact 7.4 V LiPo pack. 3) UAV framework: The Agipix V2 [10] UAV features a carbon-fiber frame with integrated propeller guards, designed to carry LiDAR–SLAM sensors while maintaining agility and low weight. It achieves a 2.8:1 thrust-to-weight ratio using a 6S power system with high-thrust brushless motors. The compact 35 ×30 cm frame allows safe indoor flight in constrained industrial spaces. 4) Radiation sensor integration: A DFRobot Gravity Geiger Counter serves as the radiation sensing module, combining a GM tube, high-voltage supply, and TTL pulse output. An Arduino-compatible MCU counts incoming pulses via hardware interrupts and transmits the counts-per-minute rate through serial at 115200 baud. A lightweight ROS2 node running on the Orin NX parses and republishes this data in real time. This open-source design ensures modularity, low-latency reporting, and straightforward replication for future inspection platforms. B. Algorithms 1) Perception: Our system builds on the Direct LiDARInertial Odometry (DLIO) framework of Chen et al. [8], which tightly couples raw LiDAR scans with high-rate IMU data via a nonlinear geometric observer and continuous-time motion model. Initially, the observer integrates IMU measurements to estimate pose, velocity, and biases, providing strong priors for motion correction. Assuming constant jerk and constant angular acceleration over each LiDAR sweep, we then perform closed-form per-point de-skewing in parallel, removing motion distortion efficiently. Rather than matching scans pairwise, DLIO registers each deskewed sweep directly against a global map, simplifying the back-end and reducing latency. We deploy DLIO on both platforms: the UGV carries a Livox Mid360 LiDAR rigidly mounted alongside an Xsens MTI-300 IMU, while the UAV uses the same LiDAR (pitched downward by 45◦) with the Cube Orange’s onboard IMU. Prior to trials, we perform full extrinsic and temporal calibration, and maintain timestamp alignment in the field via Precision Time Protocol (PTP) [11] over the Ethernet backbone. To adapt DLIO to the Zwentendorf NPP - characterized by large, cluttered halls filled with metal structures – we introduce several preprocessing and weighting steps. First, each LiDAR sweep is voxel downsampled at 0.05 m and cropped to a region of interest to limit map size while retaining structural elements such as walls and floors. Next, small-scale clutter (e.g. cables, handrails) is down-weighted using a statistical outlier filter, improving scan-to-map registration stability in densely occupied areas. During optimization, we assign higher weight to planar inliers (floors, walls) detected via RANSAC plane fitting [12], ensuring that large, stable surfaces anchor the pose estimate. Loop closures are triggered when a histogram comparison between recent scans exceeds a threshold, and are enforced via a coarse-to-fine ICP alignment [13] to correct long-term drift. These modifications preserve DLIO’s direct scan-to-map efficiency while delivering low-drift SLAM over long trajectories in the GNSS-denied, metal-cluttered NPP interior. 2) UGV planning and control: We partition the plant environment into fixed-size sectors and maintain a 0.25 m resolution OctoMap updated by incoming LiDAR scans. Within each sector, frontiers are detected and clustered, and each cluster C is assigned a utility: U(C) = α I(C)−β T(C) + γ R(C) that balances expected information gain I(C)(unknown map area reduced) against travel cost T(C), plus a reward R(C) if the cluster is in a region of elevated radiation. We tune the weights α,β,γto favor efficient exploration while biasing toward areas with higher radiation readings. The UGV uses D∗Lite [14] to plan to the highest-utility frontier cluster, then A∗on the voxel grid to refine a path. A proportional heading controller drives the UGV along the path (ω=ksteer[θgoal−θ], v=vmaxe−|ω|) and a dynamic window approach [15] provides reactive obstacle avoidance. When a sector is fully explored, the robot selects the next unexplored sector via a Travelling Salesman heuristic [16]. Once all sectors are covered, D* Lite plans a path for the UGV to return to the start. This hierarchical strategy yields efficient, complete coverage and reliable homing even in a large, cluttered NPP layout. 3) UAV Planning and Control: Teleoperation of the UAV during the competition was infeasible due to severe network limitations, extremely thick walls, and metallic structures within the nuclear facility. Furthermore, fully autonomous exploration was deemed inefficient because of the strict time constraints. To address these challenges, we designed a semiautonomous control framework where the high-level controller sends relative waypoints via UDP, and the UAV locally executes these waypoints while actively avoiding obstacles. An occupancy grid map is built online using LiDAR point clouds fused with odometry from DLIO. Unlike the UGV, the UAV employs a custom map manager based on voxel grids to efficiently store and query large-scale 3D environments. This map serves as the foundation for a global planner based on the RRT* algorithm [17], which incrementally re-plans a path toward a goal point at the beginning of each trajectory segment. The planned path is partitioned into manageable segments, and each is further optimized using a variant of the Minimum Snap Trajectory method [18], [19], tailored to account for the UAV’s dynamic constraints. To enhance safety and efficiency in cluttered environments, the trajectory is further refined with time-optimal adjustments near obstacles [20]. Finally, the optimized trajectory is tracked in real time using a cascaded PID attitude controller, with adaptive commands sent in either velocity or attitude mode depending on the UAV’s flight regime and proximity to obstacles. Fig. 3: The overview of the Agipix Control pipeline 4) Simulation: We developed a photorealistic simulation platform containing UAV and UGV in IsaacSIM [21]. One shot Sim to Real deployment is enabled using Docker to simulate High level controller of the hardware platform. The low level control in PX4 is simulated using a customized version of the Pegasus Simulator [22]. The simulation models of UAV and UGV are ensured to have close characteristics physically and visually by doing unit testing against the respective hardware platforms. III. ENRICH-2025 COMPETITION ENVIRONMENT AND SETUP A. Environment and Tasks The Zwentendorf NPP presents large, multi-level halls lined with steel piping, control consoles, and machinery, often with low ambient light and narrow passageways (Fig. 5). Participating teams faced a 30-minute time limit to perform tasks including 3D mapping of the interior and radiation hotspot localization. In particular, a detailed point cloud map Fig. 4: The overview of the Agipix Simulation environment Fig. 5: The Enrich-2025 Competition Environment. of the engine hall had to be produced, and the UAV was tasked with radiation mapping in an upper-level turbine hall to locate and quantify hidden radiation sources. Each run was strictly limited to the 30 min window, with no mid-run adjustments or system reboots permitted. B. Communication Strategy Thick concrete walls and metal structures severely attenuate wireless signals, making reliable control and telemetry a major challenge. Some teams addressed this challenge with expensive custom solutions such as dedicated high-power base stations or tethered fiber-optic links. Instead of deploying such specialized communication infrastructure, we leveraged the NPP’s existing Wi-Fi mesh by assigning static IP addresses to the UGV, UAV, and operator consoles, eliminating DHCP negotiation delays. Each robot runs a background process that periodically scans for available SSIDs and, upon detecting a stronger access point, gracefully disconnects and re-associates to maintain the best link quality. IV. EXPERIMENTS AND RESULTS A. Autonomy and Teleoperation Strategy Given the three-week development timeframe, we adopted a parallel design-and-test workflow and a hybrid autonomy approach to maximize reliability. The UGV’s navigation stack Fig. 6: 3D Maps of the NPP built by our robots. ran onboard and published velocity commands to the PX4 lowlevel controller via UDP, allowing real-time autonomous driving. However, to handle the large environment efficiently, we enabled seamless human intervention through teleoperation: a forward-facing camera feed was streamed live to the operator station, and a human pilot could send drive commands when needed. If such teleop commands were received, the UGV would temporarily cede control to the operator; if the link was lost or no commands arrived, the robot would automatically fall back to its autonomous exploration mode. This ensured continuous operation despite intermittent connectivity. Fully manual control of the UAV was infeasible given the network limitations, so we adopted a semi-autonomous scheme. An operator could designate relative waypoints for the drone, which were transmitted over UDP; the UAV’s onboard computer then planned and executed trajectories to those waypoints while actively avoiding obstacles. In essence, the operator provided high-level guidance (e.g., selecting exploration targets), and the UAV handled low-level flight control autonomously, guaranteeing safe navigation even when realtime piloting was impossible. To validate our system before the competition runs, we built a digital twin of the NPP environment and conducted extensive hardware-in-the-loop simulations. Using preliminary scans of the site, we recreated the floorplan in Isaac Sim (see Section V) and ran our full software stack with a PX4 SITL model of the drone. This allowed us to tune control gains and test fail-safes in a realistic scenario, ensuring a smooth transfer to the real robots. All code, launch scripts, and configurations were managed in a single open-source repository with continuous integration, which facilitated rapid iteration and testing throughout development. This hybrid autonomy strategy proved essential during the trials: the UGV could continue exploring under autonomous control whenever network issues interrupted teleoperation, and the UAV could carry out its mission with minimal operator input, significantly mitigating the impact of communication blackouts. B. Field Trial Results In the EnRicH 2025 competition trials, our robotic team achieved most of the mission objectives. The UGV successfully performed the 3D mapping task on the ground floor, covering the entire designated area in under 20 minutes and autonomously returning to its starting point. The generated point-cloud map captured all major structural elements of the environment. Qualitatively, the map shows good alignment with the known floor plan, indicating minimal SLAM drift; Fig. 7: Left: 3D Maps the UGV is building. Right: the 3D map and the current view of the camera on the UGV. when the UGV completed its loop, the odometry had only a negligible pose error (on the order of a few tens of centimeters over 200 m traveled). The final map spans several hundred square meters of floor area at approximately 5 cm resolution. During its run, the UGV also demonstrated the ability to bypass obstacles and debris - for example, it navigated around a partially blocked corridor and over small floor protrusions (Fig. 9). The UAV likewise took off and mapped a substantial portion of the facility’s interior. It created a detailed 3D point cloud of the first-level turbine hall while simultaneously measuring radiation levels. Throughout the first half of its flight, the UAV maintained stable communication and responded to the operator’s high-level commands. However, approximately halfway through the mission, the wireless link deteriorated abruptly due to external interference (another team’s radio system). Consequently, the UAV’s incoming commands were lost, and it hovered in place awaiting further instructions. Because of this connectivity loss, the drone could not proceed to the second level of the plant, leaving that part of the radiation mapping task incomplete. Despite this setback, our radiation sensing module functioned as intended. On the ground floor, the Geiger counter registered only normal background radiation (roughly 10-20 counts per minute) since no artificial sources were present there. The intended radiation source on the upper level could not be reached in time. Nevertheless, the sensor system is capable of detecting such sources within a few meters, and it would likely have identified any significant radioactive hotspot had the UAV been able to complete its route. During missions, network performance proved to be a critical factor. We observed Wi-Fi latency fluctuating from a low 30 ms up to over 700 ms in certain areas. At latencies above roughly 300 ms, manual control became impractical. In these instances, the importance of our autonomy fallback was clearly demonstrated: Upon losing contact, the UAV remained stable, and the UGV continued exploring through periods of control lag until connectivity was restored. Overall, our low-cost system performed comparably to more expensive, specialized solutions at the event, validating the effectiveness of our design decisions and rapid-prototyping approach. (a) UGV Simulation (b) UAV Simulation Fig. 8: Simulating two robots at the simulation environment. V. DIGITAL TWIN OF THE NPP Using data obtained from the NPP, a digital twin of the environment was created to serve as a testing platform for researchers. This simulation, developed using real-world NPP data, enables the evaluation of a diverse range of robots in challenging environments. Multiple types of data were collected during and after the trials. Prior to the experiments, an approximate floor plan of the engine area and turbine hall was provided to the participants. Based on this floor plan, point cloud data were collected using the LiDAR sensors mounted on both the ground robot and the UAV. By combining the approximate floor plan with the 3D point cloud data, a detailed 3D model of the building was generated. Video footage was recorded within the testing area with permission from the competition organisers. This footage was used both to identify and locate objects within the building and to provide insights into the dimensions and materials of these objects. Using the collected data, a simulation environment was created with the Unity game engine. 2Fig. 8 shows two instances captured during tests of the simulation environment. Subfig. 8a presents an example of remotely operating the UGV, while subfig. 8b illustrates a UAV being controlled within the simulation environment. VI. LESSONS WE LEARNED AND FUTURE WORK Our experience in EnRicH 2025 highlighted several challenges. First, communication proved to be the most critical factor. By relying on the NPP’s Wi-Fi mesh, we discovered that unexpected interference (e.g., another team activating a high-power base station) can abruptly sever control. The strict 30-minute time limit and the plant’s sprawling, cluttered interiors forced us to juggle autonomy and teleoperation, adding complexity to our control architecture. Although our simulations in IsaacSim faithfully recreated the facility’s layout, they did not model wireless interference or moving obstacles, revealing a gap between virtual testing and reality. In addition, we underestimated the size and complexity of the NPP environment relative to the UAV’s 15-minute flight limit, which made it difficult to complete the entire planned flight path within a single battery cycle. Therefore, a lighter, longerendurance drone would be necessary for covering a larger area. 2The GitHub repository for the simulation is publicly available as an openaccess resource. https://github.com/kenanalperen/NPP-Digital-Twin Fig. 9: CCTV footage of our UGV and UAV operating in the NPP, navigating around obstacles. From these experiences, we recommend reserving dedicated network channels or incorporating redundant radio links to safeguard control traffic, and embedding automatic handoff routines that detect link loss and switch to safe, preplanned autonomy modes. Enhancing simulation environments with RF propagation models and dynamic agents will better prepare systems for field deployment. Finally, integrating remaining battery life into planning heuristics and establishing on-site charging checkpoints can help ensure mission completion in large, high-risk environments. ACKNOWLEDGMENT This research work has received full funding from the European Commission’s HORIZON.1.2 - Marie Skłodowska-Curie Actions (MSCA) under Grant agreement No. 101072634, project RAICAM. REFERENCES [1] Ioannis Tsitsimpelis, C James Taylor, Barry Lennox, and Malcolm J Joyce. A review of ground-based robotic systems for the characterization of nuclear environments. Progress in nuclear energy, 111:109–124, 2019. [2] Alperen Kenan, Paul Bremner, and Manuel Giuliani. Robot teleoperation design requirements from end users in nuclear facilities. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025). IEEE, 2025. [3] Robert Bogue. Robots in the nuclear industry: a review of technologies and applications. Industrial Robot: An International Journal, 38(2):113– 118, 2011. [4] European Robotics Group. Enrich – a new robotics competition. https: //enrich.european-robotics.eu/, 2025. 5th European Robotics Hackathon “EnRicH 2025” held at Zwentendorf Nuclear Power Plant, Austria (June 30–July 4, 2025). [5] C Norman Coleman, Judith L Bader, John F Koerner, Chad Hrdina, Kenneth D Cliffer, John L Hick, James J James, Monique K Mansoura, Alicia A Livinski, Scott V Nystrom, et al. Chemical, biological, radiological, nuclear, and explosive (cbrne) science and the cbrne science medical operations science support expert (cmosse). Disaster medicine and public health preparedness, 13(5-6):995–1010, 2019. [6] Lorenz Meier, Dominik Honegger, and Marc Pollefeys. Px4: A nodebased multithreaded open source robotics framework for deeply embedded platforms. In 2015 IEEE international conference on robotics and automation (ICRA), pages 6235–6240. IEEE, 2015. [7] Steven Macenski, Tully Foote, Brian Gerkey, Chris Lalancette, and William Woodall. Robot operating system 2: Design, architecture, and uses in the wild. Science robotics, 7(66):eabm6074, 2022. [8] Kenny Chen, Ryan Nemiroff, and Brett T Lopez. Direct lidar-inertial odometry: Lightweight lio with continuous-time motion correction. arXiv preprint arXiv:2203.03749, 2022. [9] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 5135–5142. IEEE, 2020. [10] Sasa Kuruppuarachchi. Agipix: Agile quadrotor platform based on px4 low-level control to aid in aerial robotic research, 2025. Accessed: 202502-11. [11] Steve T Watt, Shankar Achanta, Hamza Abubakari, Eric Sagen, Zafer Korkmaz, and Husam Ahmed. Understanding and applying precision time protocol. In 2015 Saudi Arabia Smart Grid (SASG), pages 1–7. IEEE, 2015. [12] Konstantinos G Derpanis. Overview of the ransac algorithm. Image Rochester NY, 4(1):2–3, 2010. [13] Xin Wang, Yun Li, Yaxin Peng, and Shihui Ying. A coarse-to-fine generalized-icp algorithm with trimmed strategy. Ieee Access, 8:40692– 40703, 2020. [14] Sven Koenig and Maxim Likhachev. D* lite. In Eighteenth national conference on Artificial intelligence, pages 476–483, 2002. [15] Dieter Fox, Wolfram Burgard, and Sebastian Thrun. The dynamic window approach to collision avoidance. IEEE robotics & automation magazine, 4(1):23–33, 2002. [16] Rajesh Matai, Surya Prakash Singh, and Murari Lal Mittal. Traveling salesman problem: an overview of applications, formulations, and solution approaches. Traveling salesman problem, theory and applications, 1(1):1–25, 2010. [17] Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846–894, 2011. [18] Daniel Mellinger and Vijay Kumar. Minimum snap trajectory generation and control for quadrotors. In 2011 IEEE International Conference on Robotics and Automation, pages 2520–2525, 2011. [19] Zhefan Xu, Yumeng Xiu, Xiaoyang Zhan, Baihan Chen, and Kenji Shimada. Vision-aided uav navigation and dynamic obstacle avoidance using gradient-based b-spline trajectory optimization. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1214–1220, 2023. [20] Zhefan Xu and Kenji Shimada. Quadcopter trajectory time minimization and robust collision avoidance via optimal time allocation. arXiv preprint arXiv:2309.08544, 2023. [21] NVIDIA. Nvidia isaac sim. https://developer.nvidia.com/isaac-sim, 2023. [22] Marcelo Jacinto, Jo˜ ao Pinto, Jay Patrikar, John Keller, Rita Cunha, Sebastian Scherer, and Ant´ onio Pascoal. Pegasus simulator: An isaac sim framework for multiple aerial vehicles simulation. In 2024 International Conference on Unmanned Aircraft Systems (ICUAS), pages 917–922, 2024.